# Jensen Huang – Will Nvidia’s moat persist?

## Metadata
- Channel: Dwarkesh Patel
- Duration: 103 min
- YouTube: https://www.youtube.com/watch?v=Hrbq66XqtCo

## Transcript

**[00:00] Speaker A:** We've seen the valuations of a bunch of software companies crash because people are expecting AI to commoditize software. There's a potentially naive way of thinking about things, which is: look, Nvidia sends a GDS2 file to TSMC.  
我们看到一大批软件公司的估值暴跌,因为人们预期 AI 会让软件变成大路货。有一种可能过于简单的思路是:你看,Nvidia 把 GDS2 文件发给 TSMC。  
**[00:14] Speaker A:** TSMC builds the logic dies, it builds the switches, then it packages them with the HBM that SK Hynix, Micron, and Samsung make. Then it sends it to an ODM in Taiwan where they assemble the racks.  
TSMC 制造逻辑芯片和交换机,然后把它们和 SK Hynix、Micron、Samsung 生产的 HBM 封装在一起。接着送到台湾的 ODM 厂商那里组装成机架。  
**[00:23] Speaker A:** Nvidia is fundamentally making software that other people are manufacturing, and if software gets commoditized, does Nvidia get commoditized?  
Nvidia 本质上是在做软件,只不过由别人来制造。如果软件被商品化了,Nvidia 会不会也被商品化?  
**[00:32] Speaker B:** In the end, something has to transform electrons to tokens.  
归根结底,总得有东西把电子转换成 token。  
**[00:42] Speaker B:** The transformation of electrons to tokens and making those tokens more valuable over time is hard to completely commoditize.  
把电子转换成 token,并且让这些 token 随着时间推移变得更有价值,这件事很难被完全商品化。  
**[00:59] Speaker B:** The transformation from electrons to tokens is such an incredible journey.  
从电子到 token 的转换是一段不可思议的旅程。  
**[01:05] Speaker B:** Making that token is like making one molecule more valuable than another molecule, making one token more valuable than another. The amount of artistry, engineering, science, and invention that goes into making that token valuable, obviously we're watching it happen in real time.  
生成 token 就像让一个分子比另一个分子更有价值,让一个 token 比另一个更有价值。要让 token 有价值,需要投入的艺术、工程、科学和发明的量级,我们显然正在实时见证这一切。  
**[01:21] Speaker B:** The transformation, the manufacturing, all of the science that goes in there is far from deeply understood and the journey is far from over.  
这个转换过程、制造过程,以及其中涉及的所有科学,远远谈不上被深入理解,这段旅程也远未结束。  
**[01:38] Speaker B:** I doubt that it will happen. We're going to make it more efficient, of course.  
我不认为会发生商品化。当然,我们会让它更高效。  
**[01:46] Speaker B:** The way that you framed the question is my mental model of our company.  
你提问的方式正是我对我们公司的心智模型。  
**[01:50] Speaker B:** The input is electrons, the output is tokens. In the middle is Nvidia. Our job is to do as much as necessary and as little as possible to enable that transformation to be done at incredible capabilities.  
输入是电子,输出是 token。中间是 Nvidia。我们的工作是做必须做的事,同时尽可能少做,以便让这个转换能够以惊人的能力完成。  
**[02:05] Speaker B:** What I mean by "as little as possible," whatever I don't need to do, I partner with somebody and make it part of my ecosystem.  
我说的「尽可能少做」是指,凡是我不需要做的,我就和别人合作,让它成为我生态系统的一部分。  
**[02:16] Speaker B:** If you look at Nvidia today, we probably have the largest ecosystem of partners, both in the supply chain upstream and downstream, all of the computer companies, application developers, and model makers.  
如果你看今天的 Nvidia,我们可能拥有最大的合作伙伴生态系统,既包括供应链上下游,也包括所有的计算机公司、应用开发者和模型制造商。  
**[02:26] Speaker B:** AI is a five-layer cake, if you will.  
AI 可以说是一个五层蛋糕。  
**[02:34] Speaker A:** We have ecosystems across the entire five layers. We try to do as little as possible, but the part that we have to do, as it turns out, is insanely hard.  
我们在整个五层都有生态系统。我们尽量少做,但我们必须做的那部分,事实证明难得离谱。  
**[02:46] Speaker A:** I don't think that gets commoditized. In fact, I also don't think the enterprise software companies, the tool makers... Most software companies today are tool makers.  
我不认为这部分会被商品化。事实上,我也不认为企业软件公司、工具制造商会被商品化……今天大多数软件公司都是工具制造商。  
**[03:00] Speaker A:** Some of them are not. Some of them are workflow codification systems.  
有些不是。有些是工作流固化系统。  
**[03:09] Speaker A:** But for a lot of companies, they're tool makers. For example, Excel is a tool, PowerPoint is a tool, Cadence makes tools, Synopsys makes tools. I actually see the opposite of what people see.  
但对很多公司来说,它们是工具制造商。比如 Excel 是工具,PowerPoint 是工具,Cadence 做工具,Synopsys 做工具。我看到的其实和大家看到的相反。  
**[03:22] Speaker A:** I think the number of agents is going to grow exponentially, and the number of tool users is going to grow exponentially.  
我认为 agent 的数量会指数级增长,工具使用者的数量也会指数级增长。  
**[03:34] Speaker A:** It's very likely that the number of instances of all these tools is going to skyrocket. It's very likely that the number of instances of Synopsys Design Compiler is going to skyrocket, along with the number of agents using the floor planners, our layout tools, and our design rule checkers.  
很可能所有这些工具的实例数量都会暴增。Synopsys Design Compiler 的实例数量很可能会暴增,使用布图规划器、我们的布局工具和设计规则检查器的 agent 数量也会暴增。  
**[03:58] Speaker A:** Today we're limited by the number of engineers. Tomorrow, those engineers are going to be supported by a bunch of agents.  
今天我们受限于工程师的数量。明天,这些工程师会得到一群 agent 的支持。  
**[04:04] Speaker A:** We're going to be exploring the design space like you've never seen before, and we're going to use the tools that we use today.  
我们将以前所未有的方式探索设计空间,而且会使用我们今天用的工具。  
**[04:10] Speaker A:** I think tool use is going to cause the software companies to skyrocket.  
我认为工具使用会让软件公司暴涨。  
**[04:14] Speaker A:** The reason why it hasn't happened yet is because the agents aren't good enough at using their tools yet.  
还没发生的原因是 agent 还不够擅长使用它们的工具。  
**[04:17] Speaker A:** Either these companies are going to build the agents themselves, or agents are going to get good enough to be able to use those tools.  
要么这些公司自己构建 agent,要么 agent 会变得足够好,能够使用这些工具。  
**[04:26] Speaker A:** I think it's going to be a combination of both.  
我记得在你们最新的财报文件中,你们在晶圆厂、内存和封装方面有将近1000亿美元的采购承诺。  
**[04:26] Speaker B:** I think in your latest filings, you had almost $100 billion in purchase commitments with foundries, memory, and packaging.  
我记得在你们最新的财报文件中,你们在晶圆厂、内存和封装方面有将近1000亿美元的采购承诺。  
**[04:38] Speaker B:** SemiAnalysis has reported that you will have $250 billion of these kinds of purchase commitments.  
SemiAnalysis 报道说你们将会有2500亿美元的这类采购承诺。  
**[04:44] Speaker B:** One interpretation is that Nvidia's moat is really that you've locked up many years of these scarce components.  
一种解读是,Nvidia 的护城河其实就在于你们锁定了这些稀缺组件未来多年的供应。  
**[04:48] Speaker B:** Somebody else might have an accelerator, but can they actually get the memory to build it? Can they actually get the logic to build it?  
其他公司可能有加速器,但他们真的能拿到内存来制造吗?他们真的能拿到逻辑芯片来制造吗?  
**[04:58] Speaker A:** Is this really Nvidia's big moat for the next few years?  
这真的是 Nvidia 未来几年最大的护城河吗?  
**[05:01] Speaker B:** It's one of the things that we can do that is hard for someone else to do. We've made enormous commitments upstream.  
这是我们能做到而别人很难做到的事情之一。我们在上游做了巨大的承诺。  
**[05:14] Speaker B:** Some of it is implicit. For example, a lot of the investments that are upstream are made by our supply chain because I said to the CEOs, "Let me tell you how big this industry is going to be, let me explain to you why, let me reason through it with you, and let me show you what I see."  
有些承诺是隐性的。比如,上游的很多投资是由我们的供应链做出的,因为我对那些 CEO 们说:「让我告诉你这个行业会有多大,让我解释原因,让我和你一起推理,让我展示我看到的东西。」  
**[05:33] Speaker B:** As a result of that process of informing, inspiring, and aligning with CEOs of all different industries upstream, they're willing to make the investments.  
通过这个告知、激励和与上游各行业 CEO 达成共识的过程,他们愿意做出投资。  
**[05:48] Speaker B:** Why are they willing to make the investments for me and not someone else?  
他们为什么愿意为我投资而不是为别人投资?  
**[05:51] Speaker B:** The reason for that is because they know that I have the capacity to buy their supply and sell it through my downstream. The fact is that Nvidia's downstream supply chain and our downstream demand is so large, they're willing to make the investment upstream.  
原因是他们知道我有能力购买他们的供应并通过我的下游渠道销售出去。事实是 Nvidia 的下游供应链和下游需求规模如此之大,他们才愿意在上游做投资。  
**[06:11] Speaker B:** If you look at GTC, people are marveled by the scale of it and the people that go.  
如果你看 GTC 大会,人们会惊叹于它的规模和参会者。  
**[06:18] Speaker B:** It's a full 360 degrees, the entire universe of AI all in one place.  
这是一个360度全方位的盛会,整个 AI 宇宙都汇聚在一个地方。  
**[06:25] Speaker B:** They're all in one place because they need to see each other.  
他们都聚在一个地方,因为他们需要见到彼此。  
**[06:27] Speaker B:** I bring them together so that the downstream can see the upstream, the upstream can see the downstream, and all of them can see the advances in AI.  
我把他们聚在一起,让下游能看到上游,上游能看到下游,所有人都能看到 AI 的进展。  
**[06:36] Speaker B:** Very importantly, they can all meet the AI natives, all the AI startups being built, and all the amazing things happening so they can see firsthand all the things that I tell them.  
非常重要的是,他们都能见到 AI 原生企业、所有正在创建的 AI 初创公司,以及所有正在发生的惊人事情,这样他们就能亲眼看到我告诉他们的一切。  
**[06:48] Speaker B:** I spend a lot of my time informing, directly or indirectly, our supply chain, partners, and ecosystem about the opportunity in front of us. Some people always say, "Jensen, in most keynotes, it's one announcement after another."  
我花很多时间直接或间接地向我们的供应链、合作伙伴和生态系统介绍我们面前的机会。有些人总是说:「Jensen,大多数主题演讲都是一个接一个的发布。」  
**[07:06] Speaker B:** With our keynotes, there's always a part of it that's a little torturous in the sense that it almost comes across like education.  
而我们的主题演讲,总有一部分有点折磨人,因为它几乎像是在上课。  
**[07:22] Speaker B:** In fact, that's exactly on my mind. I need to make sure the entire supply chain, upstream and downstream, the ecosystem, understands what is coming at us, why it's coming,  
事实上,这正是我心里想的。我需要确保整个供应链,上游和下游,整个生态系统,都理解即将到来的是什么,为什么会到来,  
**[07:35] Speaker A:** When it's coming, how big it's going to be, and is able to reason about it systematically, just like I reason about it. Regarding the moat as you describe it, we're able to build for a future.  
什么时候到来,规模会有多大,并且能够像我一样系统地推理这件事。关于你所说的护城河,我们能够为未来而建。  
**[07:56] Speaker A:** If our next several years are a trillion dollars in scale, we have the supply chain to do it. Without our reach, the velocity of our business...  
如果我们未来几年的规模是万亿美元级别,我们有供应链来实现它。没有我们的影响力,没有我们业务的速度...  
**[08:05] Speaker A:** Just as there's cash flow, there's supply chain flow, there's churn.  
就像有现金流一样,也有供应链流,有周转。  
**[08:10] Speaker A:** Nobody is going to build a supply chain for an architecture if the business churn is low.  
如果业务周转率低,没有人会为一个架构建立供应链。  
**[08:17] Speaker A:** Our ability to sustain the scale is only because our downstream demand is so great.  
我们能够维持这个规模,只是因为我们的下游需求如此巨大。  
**[08:23] Speaker A:** And they see it, they hear about it, they see it all coming.  
他们看到了,他们听说了,他们看到这一切正在到来。  
**[08:27] Speaker A:** That allows us to do the things we're able to do at the scale we do them.  
正是这一点让我们能够以现在的规模做我们正在做的事情。  
**[08:32] Speaker B:** I do want to understand more concretely whether the upstream can keep up.  
我确实想更具体地了解上游供应能否跟得上。  
**[08:37] Speaker B:** For many years now, you guys have been 2x-ing revenue year over year.  
这么多年来,你们的营收一直在逐年翻倍。  
**[08:41] Speaker B:** You've been more than tripling the amount of flops you're providing to the world year over year.  
你们提供给世界的算力(flops)每年增长超过三倍。  
**[08:44] Speaker B:** And 2x-ing at this scale now is really incredible.  
确实如此。  
**[08:44] Speaker A:** Exactly.  
确实如此。  
**[08:49] Speaker B:** But then you look at logic. You're the biggest customer on TSMC's N3 node, and you're one of the biggest on N2.  
但再看看逻辑芯片这块。你们是 TSMC N3 制程节点的最大客户,也是 N2 制程最大客户之一。  
**[08:57] Speaker B:** AI as a whole this year is going to be 60% of N3.  
今年整个 AI 领域将占据 N3 产能的 60%。  
**[09:00] Speaker B:** It's going to be 86% next year, according to SemiAnalysis.  
根据 SemiAnalysis 的数据,明年这个比例会达到 86%。  
**[09:03] Speaker B:** How do you double if you're the majority? And how do you do that year over year?  
如果你们已经占了大部分产能,怎么还能翻倍?而且怎么年复一年地做到?  
**[09:09] Speaker B:** Are we in a regime now where the growth rate in AI compute has to slow because of upstream?  
我们现在是不是进入了一个阶段,AI 算力的增长速度必须因为上游供应而放缓?  
**[09:14] Speaker B:** Do you see a way to get around this? How do we build 2x more fabs year over year, ultimately?  
你看到绕过这个问题的办法了吗?我们最终要怎么做到每年建造两倍数量的晶圆厂?  
**[09:25] Speaker A:** At some level, the instantaneous demand is greater than the supply upstream and downstream in the world.  
在某种程度上,瞬时需求总是大于全球上下游的供应能力。  
**[09:38] Speaker A:** At any instant, we could be limited by the number of plumbers, which actually happens.  
在任何时刻,我们都可能受限于水管工的数量,这种情况确实发生过。  
**[09:46] Speaker A:** The plumbers are invited to next year's GTC. By the way, great idea.  
水管工们会被邀请参加明年的 GTC 大会。顺便说一句,这主意不错。  
**[09:52] Speaker A:** But that's a good condition. You want an industry where the instantaneous demand is greater than the total supply of the industry. The opposite is obviously less good.  
但这其实是个好现象。你会希望一个行业的瞬时需求大于整个行业的总供应能力。反过来显然就不太好了。  
**[10:05] Speaker A:** If we're too far apart, if one particular component is too far away, the industry swarms it.  
如果供需差距太大,如果某个特定组件供应严重不足,整个行业就会蜂拥而上解决它。  
**[10:15] Speaker A:** For example, notice people aren't talking very much about CoWoS anymore.  
比如你注意到,人们现在不怎么谈论 CoWoS 了。  
**[10:20] Speaker A:** The reason for that is because for two years we swarmed the living daylights out of it.  
原因是我们花了两年时间全力攻克它。  
**[10:25] Speaker A:** We doubled, doubled, doubled on several doubles. Now I think we're in fairly good shape.  
我们翻倍、翻倍、再翻倍,连续翻了好几轮。现在我觉得我们的状况相当不错了。  
**[10:30] Speaker A:** TSMC now knows that CoWoS supply has to keep up with the rest of the logic demand and the memory demand. They're scaling CoWoS and future packaging technologies at the same level as they scale logic.  
TSMC 现在明白了,CoWoS 的供应必须跟上逻辑芯片需求和内存需求的步伐。他们正在以与逻辑芯片相同的规模扩展 CoWoS 和未来的封装技术。  
**[10:46] Speaker A:** This is terrific, because for a long time, CoWoS and HBM memory were rather specialty. But they're not specialties anymore. People now realize they're mainstream computing technology.  
这太好了,因为很长一段时间里,CoWoS 和 HBM 内存都算是比较小众的技术。但它们现在不再小众了。人们现在意识到它们是主流计算技术。  
**[11:01] Speaker A:** Of course, we're now much more able to influence a larger scope of our supply chain.  
当然,我们现在能够影响供应链的范围也大得多了。  
**[11:09] Speaker A:** At the beginning of the AI revolution, all the things that I say now, I was saying five years ago. Some people believed in it and invested in it, for example, Sanjay and the Micron team. I still remember the meeting really well where I was clear about exactly what was going to happen, why it was going to happen, and the predictions of today. They really doubled down on it.  
在 AI 革命刚开始的时候,我现在说的这些话,五年前我就在说了。有些人相信并投资了,比如 Sanjay 和 Micron 团队。我还清楚记得那次会议,我明确说明了将会发生什么、为什么会发生,以及对今天的预测。他们真的全力投入了。  
**[11:38] Speaker A:** We partnered with them across LPDDR and HBM memories, and they really invested in it.  
我们在 LPDDR 和 HBM 内存方面与他们合作,他们真的大力投资了。  
**[11:46] Speaker A:** It obviously has been tremendous for the company.  
这对公司来说显然是巨大的成功。  
**[11:49] Speaker A:** Some people came a little bit later, but now they're all here.  
有些人来得稍晚一些,但现在他们都到齐了。  
**[11:56] Speaker A:** Each one of these bottlenecks gets a great deal of attention.  
每一个瓶颈都得到了极大的关注。  
**[12:02] Speaker A:** Now we're prefetching the bottlenecks years in advance.  
现在我们提前好几年就在预判和解决这些瓶颈。  
**[12:06] Speaker A:** For example, the investments that we've done with Lumentum, Coherent, and the silicon photonics ecosystem over the last several years really reshaped the supply chain.  
比如说,过去几年我们对 Lumentum、Coherent 以及硅光子生态系统的投资,真正重塑了整个供应链。  
**[12:23] Speaker A:** We built up an entire supply chain around TSMC. We partnered with them on COUPE, invented a whole bunch of technology, and licensed those patents to the supply chain to keep it nice and open.  
我们围绕 TSMC 建立了整个供应链。我们和他们在 COUPE 上合作,发明了一大堆技术,然后把这些专利授权给供应链,保持它的开放性。  
**[12:36] Speaker A:** We're preparing the supply chain through the invention of new technologies, new workflows,  
我们通过发明新技术、新工作流程来为供应链做准备,  
**[12:42] Speaker A:** New testing equipment like double-sided probing, investing in companies, and helping them scale up their capacity. You can see that we're trying to shape the ecosystem so that the supply chain is ready to support the scale.  
比如双面探针这样的新测试设备,投资相关公司,帮助它们扩大产能。你可以看到我们在努力塑造整个生态系统,让供应链能够支撑这样的规模。  
**[12:57] Speaker B:** It seems like some bottlenecks are easier than others.  
看起来有些瓶颈比其他的更容易解决。  
**[13:00] Speaker A:** Scaling up CoWoS versus scaling up—I went to the hardest one, by the way.  
扩大 CoWoS 产能和扩大——顺便说一句,我说的是最难的那个。  
**[13:04] Speaker B:** Which is?  
水管工。水管工和电工。这也是我对那些末日论者描述工作终结、工作岗位消失的担忧之一。  
**[13:04] Speaker A:** Plumbers. Plumbers and electricians. This is one of the concerns that I have about the doomers describing the end of work and killing of jobs.  
水管工。水管工和电工。这也是我对那些末日论者描述工作终结、工作岗位消失的担忧之一。  
**[13:26] Speaker A:** If we discourage people from being software engineers, we're going to run out of software engineers. The same prediction happened ten years ago.  
如果我们劝阻人们成为软件工程师,我们就会面临软件工程师短缺。同样的预测十年前就出现过。  
**[13:35] Speaker A:** Some of the doomers were telling people, "Whatever you do, don't be a radiologist."  
一些末日论者当时告诉人们:「无论如何,千万别当放射科医生。」  
**[13:43] Speaker A:** You might hear some of those videos still on the web saying radiology is going to be the first career to go and the world is not going to need any more radiologists.  
你现在可能还能在网上听到一些这样的视频,说放射科会是第一个消失的职业,世界不再需要更多放射科医生。  
**[13:51] Speaker A:** Guess what we're short of? Radiologists.  
你猜我们现在缺什么?放射科医生。  
**[13:58] Speaker B:** Going back to this point about how some things you can scale, and other things... How do you actually manufacture 2x the amount of logic a year?  
回到刚才那个话题,有些东西你可以扩大规模,有些则不行……你怎么能在一年内把逻辑芯片的产量翻倍?  
**[14:03] Speaker B:** Ultimately, memory and logic are bottlenecked by EUV.  
归根结底,内存和逻辑芯片都受制于 EUV 光刻机。  
**[14:07] Speaker B:** How do you get to 2x as many EUV machines year over year?  
你怎么做到每年 EUV 机器的数量翻倍?  
**[14:10] Speaker A:** None of that is impossible to scale quickly. All of that is easy to do within two or three years. You just need a demand signal.  
这些都不是不可能快速扩大规模的。这些在两三年内都很容易做到。你只需要一个需求信号。  
**[14:23] Speaker A:** Once you can build one, you can build ten, and once you can build ten, you can build a million.  
一旦你能造一台,你就能造十台,一旦你能造十台,你就能造一百万台。  
**[14:28] Speaker A:** These things are not hard to replicate.  
这些东西复制起来并不难。  
**[14:32] Speaker B:** How far down the supply chain do you go?  
你会去找 ASML 说:「嘿,如果我展望三年后,要让 Nvidia 实现每年两万亿的收入,我们需要多得多的 EUV 机器」吗?  
**[14:32] Speaker B:** Do you go to ASML and say, "Hey, if I look out three years from now, for Nvidia to be generating two trillion a year in revenue, we need way more EUV machines"?  
你会去找 ASML 说:「嘿,如果我展望三年后,要让 Nvidia 实现每年两万亿的收入,我们需要多得多的 EUV 机器」吗?  
**[14:42] Speaker A:** Some of them I have to directly, some of them indirectly, and some of them... If I can convince TSMC, ASML will be convinced. We have to think about the critical pinch points.  
有些我必须直接去谈,有些是间接的,还有些……如果我能说服 TSMC,ASML 自然就会被说服。我们必须考虑关键的卡点在哪里。  
**[14:55] Speaker A:** But if TSMC is convinced, you'll have plenty of EUV machines in a few years.  
但如果 TSMC 被说服了，几年内你就会有大量的 EUV 光刻机。  
**[15:04] Speaker A:** My point is that none of the bottlenecks last longer than a couple of years, two, three years, none of them. Meanwhile, we're improving computing efficiency by 10x, 20x, and in the case of Hopper to Blackwell, 30x to 50x.  
我想说的是，这些瓶颈没有一个会持续超过几年，两三年而已，一个都没有。与此同时，我们正在将计算效率提升 10 倍、20 倍，而从 Hopper 到 Blackwell 这一代，提升了 30 到 50 倍。  
**[15:19] Speaker A:** We're coming up with new algorithms because CUDA is so flexible.  
我们正在开发新的算法，因为 CUDA 非常灵活。  
**[15:24] Speaker A:** We're developing all kinds of new techniques so that we drive efficiency in addition to increasing capacity. None of those things worry me.  
我们正在开发各种新技术，在提升产能的同时也提高效率。这些事情都不让我担心。  
**[15:36] Speaker A:** It's the stuff that's downstream from us. Energy policies that prevent energy from... You can't create an industry without energy. You can't create a whole new manufacturing industry without energy.  
真正让我担心的是我们下游的那些事情。那些阻碍能源供应的能源政策……没有能源就无法创建一个产业。没有能源就无法创建一个全新的制造业。  
**[15:52] Speaker A:** We want to reindustrialize the United States. We want to bring back chip manufacturing, computer manufacturing, and packaging.  
我们想让美国重新工业化。我们想把芯片制造、计算机制造和封装带回来。  
**[15:58] Speaker A:** We want to build new things like EVs and robots. We want to build AI factories.  
我们想制造新的东西，比如电动汽车和机器人。我们想建造 AI 工厂。  
**[16:02] Speaker A:** You can't build any of these things without energy, and those things take a long time.  
没有能源，这些东西一个都造不出来，而且这些事情需要很长时间。  
**[16:08] Speaker A:** More chip capacity, that's a 2-3 year problem. More CoWoS capacity, 2-3 year problem.  
更多的芯片产能，那是个 2 到 3 年的问题。更多的 CoWoS 产能，也是 2 到 3 年的问题。  
**[16:13] Speaker B:** Interesting. I feel like I have guests tell me the exact opposite thing sometimes.  
有意思。我感觉有时候我的嘉宾会告诉我完全相反的观点。  
**[16:17] Speaker B:** In this case, I just don't have the technical knowledge to adjudicate.  
在这种情况下，我只是没有足够的技术知识来做判断。  
**[16:20] Speaker A:** The beautiful thing is you're talking to the expert.  
好在你现在正在和专家对话。  
**[16:23] Speaker B:** True. I want to ask about your competitors. If you look at the TPU, arguably two out of the top three models in the world, Claude and Gemini, were trained on TPU.  
确实。我想问问你的竞争对手。如果看 TPU 的话，可以说世界上排名前三的模型中有两个，Claude 和 Gemini，都是在 TPU 上训练的。  
**[16:39] Speaker B:** What does that mean for Nvidia going forward?  
这对 Nvidia 未来意味着什么？  
**[16:47] Speaker A:** We build a very different thing. What Nvidia built is accelerated computing, not a tensor processing unit.  
我们构建的是非常不同的东西。Nvidia 构建的是加速计算，而不是张量处理单元。  
**[16:56] Speaker A:** Accelerated computing is used for all kinds of things: molecular dynamics, quantum chromodynamics, data processing, data frames, structured data, and unstructured data.  
加速计算可以用于各种各样的事情：分子动力学、量子色动力学、数据处理、数据框架、结构化数据和非结构化数据。  
**[17:09] Speaker A:** It's also used for fluid dynamics and particle physics.  
它还用于流体动力学和粒子物理学。  
**[17:14] Speaker A:** In addition, we use it for AI. Accelerated computing is much more diverse.  
此外，我们还用它来做 AI。加速计算的应用范围要广泛得多。  
**[17:22] Speaker A:** Although AI is the conversation today and is obviously very important and impactful,  
虽然 AI 是今天的话题，而且显然非常重要和有影响力，  
**[17:28] Speaker A:** Computing is much broader than that. Nvidia has reinvented the way computing is done, moving from general-purpose computing to accelerated computing.  
但计算的范围要广泛得多。Nvidia 重新定义了计算的方式，从通用计算转向了加速计算。  
**[17:38] Speaker A:** Our market reach is far greater than any TPU or ASIC can possibly have.  
我们的市场覆盖范围远远超过任何 TPU 或 ASIC 所能达到的。  
**[17:47] Speaker A:** If you look at our position, we're the only company that accelerates applications of all kinds. We have a gigantic ecosystem. So all kinds of frameworks and algorithms run on Nvidia.  
如果你看我们的定位，我们是唯一一家能加速各种应用的公司。我们有一个庞大的生态系统。所以各种框架和算法都能在 Nvidia 上运行。  
**[18:02] Speaker A:** Because our computers are designed to be operated by other people, anyone who's an operator can buy our systems.  
因为我们的计算机是设计给其他人操作的，任何运营商都可以购买我们的系统。  
**[18:08] Speaker A:** With most of these home-built systems, you have to be your own operator because they were never designed to be flexible enough for others to operate.  
而对于大多数这些自研系统，你必须自己做运营商，因为它们从来没有被设计得足够灵活以供他人操作。  
**[18:19] Speaker A:** Because anybody can operate our systems, we're in every cloud, including Google, Amazon, Azure, and OCI.  
因为任何人都可以操作我们的系统，所以我们在每个云平台上都有，包括 Google、Amazon、Azure 和 OCI。  
**[18:31] Speaker A:** If you want to operate it to rent, you better have a large ecosystem of customers in many industries to be the offtakers.  
如果你想通过租赁来运营,那你最好拥有一个庞大的客户生态系统,覆盖多个行业,这些客户可以成为承租方。  
**[18:40] Speaker A:** If you want to operate it for yourself, we obviously have the ability to help you operate it yourself, like we did for Elon with xAI.  
如果你想自己运营,我们显然有能力帮助你自主运营,就像我们为 Elon 的 xAI 所做的那样。  
**[18:55] Speaker A:** And because we can enable operators in any company and any industry, you could use it to build a supercomputer for scientific research and drug discovery at Lilly.  
因为我们能够为任何公司、任何行业的运营者提供支持,你可以用它来构建超级计算机,用于科学研究和药物发现,比如在 Lilly 公司。  
**[19:10] Speaker A:** We can help them operate their own supercomputer and use it for the entire diversity of drug discovery and biological sciences that we accelerate.  
我们可以帮助他们运营自己的超级计算机,并将其用于药物发现和生物科学的各个领域,这些都是我们能够加速的方向。  
**[19:21] Speaker A:** There are just a whole bunch of applications that we can address that you can't do with TPUs.  
有大量的应用场景是我们能够支持的,而这些是 TPU 无法做到的。  
**[19:28] Speaker A:** Nvidia built CUDA to be a fantastic tensor processing unit as well, but it also handles every life cycle of data processing, computing, AI, and so on.  
Nvidia 构建 CUDA 时,不仅让它成为一个出色的张量处理单元,还让它能够处理数据处理、计算、AI 等各个生命周期的任务。  
**[19:41] Speaker A:** Our market opportunity is just a lot larger, and our reach is a lot greater.  
我们的市场机会要大得多,覆盖范围也广得多。  
**[19:48] Speaker A:** Because we support every application in the world now, you can build Nvidia systems anywhere and know that there will be customers for it. It's a very different thing.  
因为我们现在支持全球所有的应用,你可以在任何地方构建 Nvidia 系统,并且知道一定会有客户需要它。这是完全不同的。  
**[20:00] Speaker B:** This is going to be a long question. You have spectacular revenue, and you're not making $60 billion a quarter from pharma and quantum.  
这个问题会比较长。你们的营收非常惊人,但你们每季度 600 亿美元的收入并不是来自制药和量子计算。  
**[20:10] Speaker A:** You're making it because AI is an unprecedented technology that is growing unprecedentedly fast. The question then is what is best for AI specifically.  
你们能赚到这些钱,是因为 AI 是一项前所未有的技术,而且正在以前所未有的速度增长。那么问题就是,什么对 AI 本身最有利。  
**[20:19] Speaker A:** I'm not in the details, but I talk to my AI researcher friends and they say, "Look, when I use a TPU, it's this big systolic array that's perfect for doing matrix multiplies, whereas a GPU is very flexible. It's great when you have lots of branching or irregular memory access."  
我不了解细节,但我和做 AI 研究的朋友聊过,他们说:「你看,当我使用 TPU 时,它是一个大型脉动阵列,非常适合做矩阵乘法,而 GPU 则非常灵活。当你有大量分支或不规则的内存访问时,GPU 表现很好。」  
**[20:30] Speaker A:** But what is AI? It's just these very predictable matrix multiplies again and again and again. You don't have to give up any die area for warp schedulers or switches between threads and memory banks.  
但 AI 是什么呢?就是这些非常可预测的矩阵乘法,一遍又一遍地重复。你不需要为 warp 调度器或线程与内存库之间的切换牺牲任何芯片面积。  
**[20:47] Speaker A:** And the TPU is really optimized for the bulk of this growth in revenue and use case for compute that is coming online right now. I wonder how you react to that.  
而 TPU 真的是针对当前正在上线的这波计算增长和用例的大部分需求进行了优化。我想知道你对此有何反应。  
**[21:01] Speaker B:** Matrix multiplies are an important part of AI, but they're not the only part.  
矩阵乘法是 AI 的重要组成部分,但不是全部。  
**[21:07] Speaker B:** If you want to come up with a new attention mechanism, disaggregate in a different way, or invent a whole new type of architecture altogether—like a hybrid SSM—you want an architecture that's generally programmable.  
如果你想提出一种新的注意力机制,以不同的方式进行解耦,或者发明一种全新的架构——比如混合 SSM——你需要一个通用可编程的架构。  
**[21:23] Speaker B:** If you want to create a model that fuses diffusion and autoregressive techniques, you want an architecture that's just generally programmable.  
如果你想创建一个融合扩散和自回归技术的模型,你需要的就是一个通用可编程的架构。  
**[21:38] Speaker B:** We run everything you can imagine. That's the advantage. It allows for the invention of new algorithms a lot more easily, because it's a programmable system.  
我们可以运行你能想到的一切。这就是优势所在。它让新算法的发明变得容易得多,因为这是一个可编程的系统。  
**[21:52] Speaker B:** The ability to invent new algorithms is really what makes AI advance so quickly.  
发明新算法的能力,正是让 AI 进步如此之快的原因。  
**[22:00] Speaker B:** TPUs, like anything else, are impacted by Moore's Law, which we know is increasing by about 25% per year. The only way to really get 10x or 100x leaps is to fundamentally change the algorithm and how it's computed every single year.  
TPU 和其他任何东西一样,都受到摩尔定律的影响,我们知道摩尔定律每年大约提升 25%。要真正实现 10 倍或 100 倍的飞跃,唯一的方法就是每年从根本上改变算法及其计算方式。  
**[22:15] Speaker B:** That's Nvidia's fundamental advantage. The only reason we were able to make Blackwell to Hopper 50x… When I first announced Blackwell was going to be 35x more energy efficient than Hopper, nobody believed it.  
这就是 Nvidia 的根本优势。我们之所以能让 Blackwell 相比 Hopper 提升 50 倍……当我第一次宣布 Blackwell 的能效将比 Hopper 高 35 倍时,没人相信。  
**[22:42] Speaker B:** Then Dylan wrote an article saying I sandbagged, and it's actually fifty times.  
然后 Dylan 写了一篇文章说我保守了,实际上是 50 倍。  
**[22:49] Speaker B:** You can't reasonably do that with just Moore's Law.  
仅靠摩尔定律是不可能做到这一点的。  
**[22:53] Speaker A:** The way we solve that problem is with new models, like MoEs, that are parallelized, disaggregated, and distributed across a computing system. Without the ability to really get down and come up with new kernels with CUDA, it's really hard to do.  
我们解决这个问题的方法是使用新模型,比如 MoE,它们是并行化的、解耦的,并分布在整个计算系统中。如果没有能力真正深入并用 CUDA 开发新的内核,这真的很难做到。  
**[23:15] Speaker A:** It's the combination of the programmability of our architecture and the fact that Nvidia is an extreme co-design company. We can even offload some of the computation into the fabric itself, like NVLink, or into the network with Spectrum-X.  
这是我们架构的可编程性与 Nvidia 作为一家极致协同设计公司这一事实的结合。我们甚至可以将一些计算卸载到结构本身,比如 NVLink,或者卸载到网络中,比如 Spectrum-X。  
**[23:36] Speaker A:** We could affect change across the processors, the system, the fabric, the libraries, and the algorithm simultaneously. Without CUDA to do that, I wouldn't even know where to start.  
我们可以同时在处理器、系统、结构、库和算法层面实现变革。如果没有 CUDA 来做这件事,我甚至不知道从哪里开始。  
**[23:51] Speaker B:** My sponsor Crusoe was among the first clouds to offer NVIDIA's Blackwell and Blackwell Ultra platforms.  
我的赞助商 Crusoe 是最早提供 NVIDIA Blackwell 和 Blackwell Ultra 平台的云服务商之一。  
**[23:58] Speaker B:** And they just announced their NVIDIA Vera Rubin deployment scheduled for later this year.  
他们刚刚宣布将在今年晚些时候部署 NVIDIA Vera Rubin。  
**[24:02] Speaker B:** But access to state-of-the-art hardware is only part of the story.  
但能用上最先进的硬件只是故事的一部分。  
**[24:05] Speaker B:** For example, most inference engines already do KV caching for a single user's forward passes.  
比如说,大多数推理引擎已经会为单个用户的前向传播做 KV 缓存。  
**[24:09] Speaker B:** But Crusoe does it across users and GPUs. So if a thousand agents are running on the same system prompt, Crusoe only has to compute the KV cache once for it to become available to every single GPU in the cluster.  
但 Crusoe 是跨用户和跨 GPU 做缓存的。所以如果有一千个 agent 在用同一个系统提示词,Crusoe 只需要计算一次 KV 缓存,就能让集群里的每个 GPU 都用上。  
**[24:17] Speaker B:** This is especially important as systems get more agentic and require much longer prefixes in order to use tools and access files.  
这在系统变得更智能化、需要更长的前缀来使用工具和访问文件时尤其重要。  
**[24:27] Speaker B:** In a recent benchmark, Crusoe was able to deliver up to 10x faster time-to-first-token and up to 5x better throughput than vLLM.  
在最近的基准测试中,Crusoe 的首 token 时间比 vLLM 快了最多 10 倍,吞吐量提升了最多 5 倍。  
**[24:32] Speaker B:** This is just one among many reasons that you should run your inference workload with Crusoe. And if you need GPUs for training, you don't need to switch clouds. Crusoe's got you covered there too.  
这只是你应该在 Crusoe 上运行推理工作负载的众多理由之一。而且如果你需要 GPU 来做训练,也不用换云服务商,Crusoe 也能满足你。  
**[24:43] Speaker B:** Go to crusoe.ai/dwarkesh to learn more. This gets at an interesting question about Nvidia's clientele.  
访问 crusoe.ai/dwarkesh 了解更多。这引出了一个关于 Nvidia 客户群的有趣问题。  
**[24:50] Speaker B:** 60% of your revenue is coming from these big five hyperscalers.  
你们 60% 的收入来自这五大超大规模云服务商。  
**[25:00] Speaker B:** In a different era with different customers—let's say professors running experiments—they need CUDA. They can't use another accelerator. They just needed to run PyTorch with CUDA and...  
在另一个时代,面对不同的客户——比如说做实验的教授们——他们需要 CUDA,不能用别的加速器。他们只需要用 CUDA 跑 PyTorch,然后……  
**[25:12] Speaker A:** Have everything optimized. But these hyperscalers have the resources to write their own kernels. In fact, they have to in order to get that last 5% of performance they need for their specific architecture.  
把一切都优化好。但这些超大规模云服务商有资源自己写 kernel。事实上,为了在他们特定的架构上榨出最后 5% 的性能,他们必须这么做。  
**[25:23] Speaker A:** Anthropic and Google are mostly running their own accelerators or running TPUs and Trainium.  
Anthropic 和 Google 主要在用自己的加速器,或者在用 TPU 和 Trainium。  
**[25:30] Speaker A:** But even OpenAI, using GPUs, has Triton because they need their own kernels.  
但即使是用 GPU 的 OpenAI,也有 Triton,因为他们需要自己的 kernel。  
**[25:38] Speaker A:** Down to CUDA C++, instead of using cuBLAS and NCCL, they've got their own stack which compiles to other accelerators as well. If most of your customers can and do make replacements for CUDA, to what extent is CUDA really the thing that is going to make frontier AI happen on Nvidia?  
深入到 CUDA C++ 层面,他们不用 cuBLAS 和 NCCL,而是有自己的技术栈,还能编译到其他加速器上。如果你的大部分客户都能够、而且确实在替换 CUDA,那 CUDA 到底在多大程度上真的是让前沿 AI 在 Nvidia 上实现的关键?  
**[25:55] Speaker B:** CUDA is a rich ecosystem. If you want to build on any computer first, building on CUDA first is incredibly smart.  
CUDA 是一个丰富的生态系统。如果你想先在某个计算平台上构建,先在 CUDA 上构建是非常明智的。  
**[26:11] Speaker B:** Because the ecosystem is so rich, we support every framework.  
因为这个生态系统非常丰富,我们支持所有框架。  
**[26:16] Speaker B:** If you want to create custom kernels... For example, we contribute enormously to Triton.  
如果你想创建自定义 kernel……比如说,我们对 Triton 贡献巨大。  
**[26:23] Speaker B:** So the back end of Triton has huge amounts of Nvidia technology.  
所以 Triton 的后端有大量 Nvidia 的技术。  
**[26:28] Speaker B:** We're delighted to help every framework become as great as it can be.  
我们很乐意帮助每个框架变得尽可能优秀。  
**[26:33] Speaker B:** There are lots and lots of frameworks. There's Triton, vLLM, SGLang, and more.  
框架有很多很多,有 Triton、vLLM、SGLang 等等。  
**[26:38] Speaker B:** Now there's a whole bunch of new reinforcement learning frameworks coming out, like verl and NeMo RL. With post-training and reinforcement learning, that entire area is just exploding.  
现在还有一大批新的强化学习框架出现,比如 verl 和 NeMo RL。在后训练和强化学习方面,整个领域正在爆发式增长。  
**[26:50] Speaker B:** So if you want to build on an architecture, building on CUDA makes the most sense because you know the ecosystem is great.  
所以如果你想在某个架构上构建,在 CUDA 上构建最合理,因为你知道生态系统很棒。  
**[27:00] Speaker B:** You know that if something happens, it's more likely in your code and not in the mountain of code underneath. Don't forget the amount of code you're dealing with when building these systems.  
你知道如果出了问题,更可能是你的代码有问题,而不是底层那一大堆代码。别忘了构建这些系统时你要处理的代码量有多大。  
**[27:08] Speaker B:** When something doesn't work, was it you or was it the computer? You would like it to always be you and to be able to trust the computer.  
当某个东西不工作时,是你的问题还是计算机的问题?你希望问题总是出在你这边,能够信任计算机。  
**[27:17] Speaker B:** Obviously, we still have lots of bugs ourselves, but our system is so well wrung out that you can at least build on top of the foundation.  
显然,我们自己也还有很多 bug,但我们的系统已经经过了充分的锤炼,至少你可以在这个基础上进行开发。  
**[27:31] Speaker A:** That's number one: the richness, programmability, and capability of the ecosystem.  
第一点是:生态系统的丰富性、可编程性和能力。  
**[27:34] Speaker A:** The second thing is, if you're a developer building anything at all, the single most important thing you want is an install base. You want the software you write to run on a whole bunch of other computers. You're not building software just for yourself.  
第二点是,如果你是开发者,无论在做什么,你最需要的就是安装基数。你希望自己写的软件能在大量其他计算机上运行。你不是只为自己开发软件。  
**[27:52] Speaker A:** You're building it for your fleet or everybody else's fleet because you're a framework builder.  
你是在为自己的机群或其他所有人的机群开发,因为你是框架构建者。  
**[27:57] Speaker A:** Nvidia's CUDA ecosystem is ultimately its great treasure.  
Nvidia 的 CUDA 生态系统最终是它最宝贵的财富。  
**[28:02] Speaker A:** We have several hundred million GPUs out there now. Every cloud has it. It goes back to the A10, A100, H100, H200, the L series, the P series. There's a whole bunch of them.  
我们现在有数亿块 GPU 在外面运行。每个云平台都有。从 A10、A100、H100、H200,到 L 系列、P 系列,有一大堆型号。  
**[28:21] Speaker A:** They're in all kinds of sizes and shapes. If you're a robotics company, you want that CUDA stack to actually run in the robot itself. We're literally everywhere. The install base means that once you develop the software or the model, it's going to be useful everywhere. That is just incredibly valuable. Lastly, the fact that we're in every single cloud makes us genuinely unique.  
它们有各种尺寸和形态。如果你是机器人公司,你会希望 CUDA 技术栈能在机器人本身上运行。我们真的无处不在。这个安装基数意味着,一旦你开发了软件或模型,它就能在任何地方使用。这非常有价值。最后,我们存在于每一个云平台这个事实,让我们真正独一无二。  
**[28:46] Speaker A:** If you're an AI company or developer, you're not exactly sure which cloud service provider you're going to partner with or where you'd like to run it.  
如果你是 AI 公司或开发者,你不一定确定要和哪个云服务提供商合作,或者想在哪里运行。  
**[28:55] Speaker A:** We run everywhere, including on-prem for you if you like.  
我们在任何地方都能运行,如果你愿意,也包括本地部署。  
**[29:01] Speaker A:** The combination of the richness of the ecosystem, the expansiveness of the install base, and the versatility of where we are makes CUDA invaluable.  
生态系统的丰富性、安装基数的广泛性,以及我们所在位置的多样性,这些结合起来让 CUDA 变得不可或缺。  
**[29:10] Speaker B:** That makes a lot of sense.  
这很有道理。  
**[29:17] Speaker B:** I guess the thing I'm curious about is whether those advantages matter a lot to your main customers. There's many people for whom they might matter.  
我好奇的是,这些优势对你们的主要客户来说是否真的很重要。对很多人来说可能确实重要。  
**[29:29] Speaker B:** The kind of person who can actually build their own software stack makes up most of your revenue.  
但那种能够自己构建软件栈的人,才是你们大部分收入的来源。  
**[29:34] Speaker B:** Especially if you go to a world where AI is getting especially good at the things which have tight verification loops where you can RL on them.  
特别是如果进入一个 AI 在那些有严密验证循环、可以用强化学习的事情上变得特别擅长的世界……  
**[29:42] Speaker B:** This question of how do you write a kernel that does attention or MLP the most efficiently across a scale-up? It's a very verifiable sort of feedback loop.  
比如如何写一个 kernel 来在扩展规模上最高效地执行 attention 或 MLP?这是一个非常可验证的反馈循环。  
**[29:54] Speaker B:** Can all the hyperscalers write these custom kernels for themselves?  
所有超大规模云厂商能不能自己写这些定制 kernel?  
**[29:59] Speaker A:** Nvidia still has great price performance, so they might still prefer to use Nvidia. But then the question is, does it just become a question of who is offering the best specs, the best flops and memory bandwidth for a given dollar.  
Nvidia 仍然有很好的性价比,所以他们可能还是更愿意用 Nvidia。但问题就变成了,这是否只是一个谁能在给定价格下提供最好规格、最好的浮点运算和内存带宽的问题。  
**[30:13] Speaker A:** Whereas historically Nvidia has just had, and still has, the best margins in all of AI across hardware and software, over 70%, because of this CUDA moat.  
而历史上 Nvidia 一直拥有,现在仍然拥有 AI 领域所有硬件和软件中最高的利润率,超过 70%,就是因为这个 CUDA 护城河。  
**[30:21] Speaker A:** And the question is, can you sustain those margins if for most of your customers, they can actually afford to build instead of the CUDA moat?  
问题是,如果你的大多数客户实际上有能力自己构建,而不是依赖 CUDA 护城河,你还能维持这些利润率吗?  
**[30:28] Speaker B:** The number of engineers we have assigned to these AI labs is insane, working with them, optimizing their stack.  
我们派驻到这些 AI 实验室的工程师数量是惊人的,和他们一起工作,优化他们的技术栈。  
**[30:41] Speaker B:** The reason for that is because nobody knows our architecture better than we do.  
原因是没有人比我们更了解我们的架构。  
**[30:46] Speaker B:** These architectures are not as general purpose as a CPU.  
这些架构不像 CPU 那样通用。  
**[30:54] Speaker B:** A CPU is kind of like a Cadillac. It's a nice cruiser. It never goes too fast. Everybody drives it pretty well. It's got cruise control, and everything's easy.  
CPU 有点像 Cadillac。它是一辆不错的巡航车。从不开得太快。每个人都能开得很好。它有定速巡航,一切都很简单。  
**[31:10] Speaker B:** But in a lot of ways, Nvidia's GPUs, accelerators, are like F1 racers.  
但在很多方面,Nvidia 的 GPU、加速器,就像 F1 赛车。  
**[31:18] Speaker B:** I could imagine everybody's able to drive it at a hundred miles an hour, but it takes quite a bit of expertise to be able to push it to the limit.  
我能想象每个人都能开到一百英里每小时,但要把它推到极限,需要相当多的专业知识。  
**[31:24] Speaker B:** We use a ton of AI to create the kernels that we have.  
我们用了大量 AI 来创建现有的这些 kernel。  
**[31:30] Speaker B:** I'm pretty sure we're going to still be needed for quite some time.  
我很确定,在相当长的一段时间内,我们还是会被需要的。  
**[31:34] Speaker B:** Our expertise helps our AI lab partners to get another 2x out of their stack easily oftentimes.  
我们的专业能力经常能帮助 AI 实验室合作伙伴轻松地从他们的技术栈中再榨出 2 倍性能。  
**[31:44] Speaker B:** It's not unusual that by the time we're done optimizing their stack or optimizing a particular kernel, their model sped up by 3x, 2x, 50%.  
等我们优化完他们的技术栈或某个特定 kernel 后,模型速度提升 3 倍、2 倍或 50% 都不罕见。  
**[32:01] Speaker B:** That's a huge number, especially when you're talking about the install base of the fleet that they have, of all the Hoppers and Blackwells that they have.  
这个数字非常可观,尤其是考虑到他们拥有的那些 Hopper 和 Blackwell 芯片的装机规模。  
**[32:09] Speaker B:** When you increase it by a factor of two, that doubles the revenues. That directly translates to revenues.  
当你把性能提升 2 倍时,收入就翻倍了。这直接转化为收入。  
**[32:16] Speaker B:** Nvidia's computing stack is the best performance per TCO in the world, bar none.  
Nvidia 的计算栈在全世界范围内拥有最佳的性能 TCO 比,没有之一。  
**[32:24] Speaker B:** Nobody can demonstrate to me that any single platform in the world today has a better performance-TCO ratio. Not one company. In fact, the benchmarks that are out there.  
没有人能向我证明当今世界上有任何一个平台拥有更好的性能-TCO 比率。一家公司都没有。事实上,现有的那些基准测试就摆在那里。  
**[32:38] Speaker A:** Dylan's InferenceMAX is sitting out there for everybody to use, and not one... TPU won't come, Trainium won't come.  
Dylan 的 InferenceMAX 就公开放在那里供所有人使用,但没有一个……TPU 不会来,Trainium 也不会来。  
**[32:46] Speaker A:** I encourage them to use InferenceMAX and demonstrate their incredible inference cost. It's really hard. Nobody wants to show up.  
我鼓励他们使用 InferenceMAX 来展示他们那惊人的推理成本。这真的很难。没人愿意露面。  
**[32:55] Speaker A:** MLPerf. I would welcome Trainium to demonstrate their 40% that they claim all the time. I would love to hear them demonstrate the cost advantage of TPUs.  
MLPerf。我欢迎 Trainium 来证明他们一直声称的那 40% 优势。我很想听他们展示 TPU 的成本优势。  
**[33:10] Speaker A:** It makes no sense in my mind. It makes absolutely zero sense. On first principles, it makes no sense.  
在我看来这完全说不通。绝对说不通。从第一性原理来看,就说不通。  
**[33:18] Speaker A:** So I think the reason why we're so successful is simply because our TCO is so great.  
所以我认为我们如此成功的原因,就是因为我们的 TCO 太出色了。  
**[33:27] Speaker A:** Secondly, you say 60% of our customers are the top five, but most of that business is external.  
其次,你说我们 60% 的客户是前五大云厂商,但这些业务大部分是面向外部的。  
**[33:36] Speaker A:** For example, most of Nvidia in AWS is for external customers, not internal use.  
比如,AWS 上的 Nvidia 大部分是给外部客户用的,不是内部使用。  
**[33:42] Speaker A:** Most of our customers at Azure, obviously all of our customers are external.  
Azure 上我们的大部分客户,显然全都是外部客户。  
**[33:46] Speaker A:** All of our customers at OCI are external, not internal use.  
OCI 上我们所有的客户都是外部的,不是内部使用。  
**[33:49] Speaker A:** The reason why they favor us is because our reach is so great.  
他们青睐我们的原因是我们的覆盖面太广了。  
**[33:54] Speaker A:** We can bring them all of the great customers in the world. They're all built on Nvidia. And the reason why all these companies are built on Nvidia is because our reach and our versatility is so great.  
我们能为他们带来全世界所有优质客户。这些客户都构建在 Nvidia 上。而所有这些公司之所以构建在 Nvidia 上,是因为我们的覆盖面和通用性太强了。  
**[34:01] Speaker A:** So I think the flywheel is really install base, the programmability of our architecture, the richness of our ecosystem, and the fact that there's so many AI companies in the world. There's tens of thousands of them now.  
所以我认为这个飞轮效应实际上就是:装机量、我们架构的可编程性、生态系统的丰富度,以及世界上有如此多的 AI 公司这个事实。现在已经有数万家了。  
**[34:22] Speaker A:** If you were one of those AI startups, what architecture would you choose?  
如果你是那些 AI 初创公司之一,你会选择什么架构?  
**[34:26] Speaker A:** You would choose an architecture that's most abundant.  
你会选择最普及的架构。  
**[34:29] Speaker A:** We're the most abundant in the world. You'd choose the one that has the largest installed base. We're the largest install base. And you'd choose the one that has a rich ecosystem.  
我们是世界上最普及的。你会选择装机量最大的。我们就是装机量最大的。你还会选择生态系统丰富的。  
**[34:36] Speaker A:** So that's the flywheel. That's the reason why, between the combination of: one, our perf per dollar is so great that they have the lowest cost tokens.  
这就是飞轮效应。这就是原因,综合来看:第一,我们的性价比太出色了,所以他们的 token 成本最低。  
**[34:49] Speaker A:** Second, our perf per watt is the highest in the world.  
第二,我们的能效比是全世界最高的。  
**[34:53] Speaker A:** So if one of these companies, if our partners, built a one gigawatt data center, that one gigawatt data center better deliver the maximum amount of revenues and number of tokens, which directly translates to revenues.  
如果这些公司中的某一家,也就是我们的合作伙伴,建造了一个 1 吉瓦的数据中心,那这个数据中心最好能产生最大的收入和 token 数量,而 token 数量直接对应收入。  
**[35:07] Speaker A:** You want it to generate as many tokens as possible, maximize the revenues for that data center.  
你希望它生成尽可能多的 token,让数据中心的收入最大化。  
**[35:13] Speaker A:** We are the highest tokens per watt architecture in the world.  
我们是全球每瓦特产出 token 数最高的架构。  
**[35:17] Speaker A:** Lastly, if your goal is to rent the infrastructure, we have the most customers in the world. So that's the reason why the flywheel works.  
最后,如果你的目标是出租基础设施,我们拥有全球最多的客户。这就是飞轮效应能够运转的原因。  
**[35:25] Speaker B:** Interesting. I guess the question comes down to, what is the actual market structure here?  
有意思。我想问题归结为,这里的实际市场结构是什么?  
**[35:30] Speaker B:** Because even if there's other companies... There could have been a world where there's tens of thousands of AI companies that have roughly equal share of compute.  
因为即使有其他公司存在……本来可能会有一个世界,那里有成千上万家 AI 公司,它们的算力份额大致相当。  
**[35:38] Speaker B:** But even through these five hyperscalers, really the people on Amazon using the compute are Anthropic, OpenAI, and these big foundation labs who can themselves afford and have the ability to make different accelerators work.  
但即使通过这五大云服务商,真正在 Amazon 上使用算力的其实是 Anthropic、OpenAI 这些大型基础模型实验室,它们自己有能力也负担得起让不同的加速器运行起来。  
**[35:51] Speaker A:** No, I think your premise is wrong.  
不,我认为你的前提是错的。  
**[35:58] Speaker B:** Maybe. But let me ask you a slightly different question.  
也许吧。但让我换个问题问你。  
**[36:01] Speaker A:** Come back and make me correct your premise.  
好的。让我换个问题问你。  
**[36:01] Speaker B:** Okay. Let me just ask you a different question.  
好的。让我换个问题问你。  
**[36:08] Speaker A:** But still make sure to make me come back and fix because it's just too important to AI. It's too important to the future of science. It's too important to the future of the industry. That premise... Look —  
但还是要确保让我回来纠正,因为这对 AI 太重要了。对科学的未来太重要了。对整个行业的未来太重要了。那个前提……你看——  
**[36:16] Speaker B:** Let me just finish the question and then we can address it together.  
让我先把问题问完,然后我们一起讨论。  
**[36:23] Speaker A:** Yeah.  
好。  
**[36:29] Speaker B:** If all these things are true about price, performance, and performance per watt, et cetera, are true, why do you think it is the case that, say, Anthropic for example, just announced a couple days ago they have a multi-gigawatt deal with Broadcom and Google for TPUs and majority of their compute?  
如果关于价格、性能、每瓦特性能等等这些都是真的,那为什么比如说 Anthropic,几天前刚宣布他们与 Broadcom 和 Google 达成了多吉瓦的 TPU 协议,而且大部分算力都用 TPU?  
**[36:42] Speaker B:** Obviously for Google, TPU is a majority of compute.  
显然对 Google 来说,TPU 占了算力的大部分。  
**[36:47] Speaker B:** So if I look at these big AI companies, it seems like a lot of their compute... There was some point where it's all Nvidia and now it's not.  
所以如果我看这些大型 AI 公司,似乎它们的很多算力……曾经有个时期全是 Nvidia,但现在不是了。  
**[36:57] Speaker A:** So I'm curious how to square, if these things are true on paper, why are they going with other accelerators?  
所以我很好奇,如果这些在纸面上都是真的,为什么它们要选择其他加速器?  
**[37:01] Speaker B:** Anthropic is a unique instance, not a trend. Without Anthropic, why would there be any TPU growth at all? It's 100% Anthropic. Without Anthropic, why would there be Trainium growth at all? It's 100% Anthropic. I think that's fairly well known and well understood.  
Anthropic 是个特例,不是趋势。如果没有 Anthropic,为什么会有任何 TPU 增长?100% 是因为 Anthropic。如果没有 Anthropic,为什么会有 Trainium 增长?100% 是因为 Anthropic。我认为这是众所周知和被充分理解的。  
**[37:21] Speaker B:** It's not that there's an abundance of ASIC opportunities. There's only one Anthropic.  
并不是有大量的 ASIC 机会。只有一个 Anthropic。  
**[37:27] Speaker A:** But OpenAI's deals with AMD... They're building their own Titan accelerator.  
但 OpenAI 与 AMD 的协议……他们在构建自己的 Titan 加速器。  
**[37:33] Speaker B:** Yeah, but I think we could all acknowledge they're vastly Nvidia.  
是的,但我想我们都承认他们绝大部分还是用 Nvidia。  
**[37:36] Speaker B:** We're going to still do a lot of work together. I'm not offended by other people using something else and trying things.  
我们还会继续做很多合作。我不会因为别人用其他东西、尝试其他方案而感到被冒犯。  
**[37:50] Speaker B:** If they don't try these other things, how would they know how good ours is?  
如果他们不尝试这些其他东西,怎么知道我们的有多好呢?  
**[37:55] Speaker B:** Sometimes you've got to be reminded of it.  
有时候需要被提醒一下。  
**[37:55] Speaker B:** We have to continuously earn the position that we're in. There are always big claims. Look at the number of ASICs that have been canceled.  
我们必须持续证明自己配得上现在的位置。市场上总有很多豪言壮语,但你看看有多少 ASIC 项目最终都取消了。  
**[38:09] Speaker B:** Just because you're going to build an ASIC... You still have to build something better than Nvidia.  
不是说你要做 ASIC 就能成功,你还得做出比 Nvidia 更好的产品才行。  
**[38:15] Speaker B:** It's not that easy building something better than Nvidia. It's not sensible, actually. Nvidia's got to be missing something, seriously.  
做出比 Nvidia 更好的产品可不容易。实际上这想法本身就不太合理——Nvidia 肯定是在某些方面有严重缺陷,否则怎么可能被超越呢?  
**[38:20] Speaker B:** Because of our scale, our velocity, we're the only company in the world that's cranking it out every single year. Big leaps, every single year.  
凭借我们的规模和速度,我们是世界上唯一一家每年都能推出新品的公司,而且每年都有重大突破。  
**[38:32] Speaker A:** I guess their logic is, "Hey, it doesn't need to be better. It just needs to be not more than 70% worse," because they're paying you 70% margins.  
我猜他们的逻辑是:「嘿,不需要做得更好,只要性能不比 Nvidia 差超过 70% 就行」,因为他们付给你 70% 的利润率嘛。  
**[38:39] Speaker B:** No, don't forget, even in ASICs margins are really quite high. Nvidia's margin is 70%, let's say. But ASIC margins are 65%. What are you really saving?  
不对,别忘了即使是 ASIC 的利润率也非常高。假设 Nvidia 的利润率是 70%,但 ASIC 的利润率也有 65%,你到底能省多少钱?  
**[38:51] Speaker A:** Oh, you mean from Broadcom or something like that?  
对,没错。你总得付钱给某个人。据我所知,ASIC 的利润率高得惊人。他们自己也相信这一点,还挺为自己超高的 ASIC 利润率感到自豪呢。  
**[38:51] Speaker B:** Yeah, sure. You've got to pay somebody. I think the ASIC margins are incredibly good, from what I can tell. They believe it too. They're quite proud of their incredible ASIC margins.  
对,没错。你总得付钱给某个人。据我所知,ASIC 的利润率高得惊人。他们自己也相信这一点,还挺为自己超高的 ASIC 利润率感到自豪呢。  
**[39:06] Speaker A:** So, you asked the question why.  
所以,你问了为什么这个问题。  
**[39:12] Speaker B:** A long time ago, we just didn't have the ability to do it. At the time, I didn't deeply internalize how difficult it would be to build a foundation  
很久以前,我们确实没有能力做这件事。当时我没有深刻意识到,建立一个像 OpenAI 和 Anthropic 这样的基础  
**[39:29] Speaker A:** AI labs like OpenAI and Anthropic, and the fact that they needed huge investments from the suppliers themselves.  
AI 实验室会有多困难,以及它们需要供应商自己投入巨额资金这个事实。  
**[39:37] Speaker A:** We just weren't in a position to make the multi-billion dollar investment into Anthropic so that they could use our compute.  
我们当时根本没有能力向 Anthropic 投资数十亿美元,让他们使用我们的算力。  
**[39:47] Speaker A:** But Google and AWS were. They put in huge investments in the beginning so that Anthropic, in return, used their compute.  
但 Google 和 AWS 有这个能力。他们一开始就投入了巨额资金,作为回报,Anthropic 使用他们的算力。  
**[39:56] Speaker A:** We just weren't in a position to do that at the time.  
我们当时就是没有这个条件。  
**[40:00] Speaker A:** I would say my mistake is I didn't deeply internalize that they really had no other options, that a VC would never put in $5-10 billion of investment into an AI lab with the hopes of it turning out to be Anthropic.  
我觉得我的失误在于,我没有深刻认识到他们真的别无选择——风投永远不会向一个 AI 实验室投入 50 到 100 亿美元,指望它能成为下一个 Anthropic。  
**[40:20] Speaker A:** So that was my miss. But even if I understood it, I don't think we would've been in a position to do that at the time.  
这就是我的失误。但即使我当时理解了这一点,我们也不太可能有条件去做。  
**[40:27] Speaker A:** But I'm not going to make that same mistake again. I'm delighted to invest in OpenAI, and I'm delighted to help them scale, and I believe it's essential to do so.  
但我不会再犯同样的错误了。我很高兴能投资 OpenAI,很高兴能帮助他们扩大规模,我认为这是必须做的事。  
**[40:40] Speaker A:** And then, when I was able to, when Anthropic came to us, I'm delighted to be an investor, delighted to help them scale.  
后来当我有能力的时候,Anthropic 来找我们,我也很高兴成为投资者,很高兴帮助他们扩大规模。  
**[40:54] Speaker A:** We just weren't, at the time, able to do it.  
如果能重来一次——假如 Nvidia 当时就有现在这么大的规模——我会非常乐意去做这件事。  
**[40:54] Speaker A:** If I could rewind everything—and Nvidia could have been as big back then as we are now—I would've been more than happy to do it.  
如果能重来一次——假如 Nvidia 当时就有现在这么大的规模——我会非常乐意去做这件事。  
**[41:06] Speaker B:** This is actually quite interesting. For many years Nvidia has been the company in AI making money, making lots of money. Now you're investing it. It's been reported that you've done up to $30 billion in OpenAI and $10 billion in Anthropic.  
这其实挺有意思的。多年来 Nvidia 一直是 AI 领域赚钱的公司,而且赚了很多钱。现在你们在投资这些钱。据报道你们向 OpenAI 投了最多 300 亿美元,向 Anthropic 投了 100 亿美元。  
**[41:22] Speaker B:** But now their valuations have increased, and I'm sure they'll continue to increase.  
但现在他们的估值已经上涨了,而且我相信还会继续上涨。  
**[41:28] Speaker B:** So if over these many years you were giving them the compute, you saw where it was headed, and they were worth like one tenth what they're worth now a couple years ago—or even a year ago in some cases and you had all this cash—there's a world where either Nvidia themselves becomes a foundation lab, does a huge investment to make that possible, or has made the deals you've made now at current valuations much earlier on.  
所以如果这么多年来你们一直在给他们提供算力,你们看到了发展方向,而他们几年前——甚至在某些情况下一年前——的估值只有现在的十分之一,而你们又有这么多现金的话,那么就存在这样一种可能:要么 Nvidia 自己成为一个基础模型实验室,投入巨资让这成为可能,要么以当前估值更早地达成你们现在达成的交易。  
**[41:57] Speaker B:** And you had the cash to do it. So I am curious, actually, why not have done it earlier?  
而且你们有现金去做这件事。所以我确实很好奇,为什么不早点做呢?  
**[42:00] Speaker A:** We did it as soon as we could have.  
我们是在能做的时候就尽快做了。  
**[42:05] Speaker A:** We did it as soon as we could have, and if I could have, I would've done it even earlier.  
我们在能做的时候就尽快做了,如果可以的话,我其实想更早就做。  
**[42:12] Speaker A:** At the time that Anthropic needed us to do it, we just weren't in a position to do it.  
当时 Anthropic 需要我们投资的时候,我们确实还没有条件去做这件事。  
**[42:17] Speaker A:** It wasn't in our sensibility to do so.  
怎么说?是资金的问题吗?  
**[42:17] Speaker B:** How so? Was it like a cash thing?  
怎么说?是资金的问题吗?  
**[42:23] Speaker A:** Yeah, the level of investment. We had never invested outside the company at the time, and not that much. We didn't realize we needed to.  
对,投资规模的问题。我们当时从来没有对外投资过,而且投资额也没那么大。我们没意识到需要这么做。  
**[42:36] Speaker A:** I always thought that they could just go raise from VCs, for God's sakes, like all companies do.  
我一直觉得他们可以像所有公司一样去找风投融资就行了。  
**[42:42] Speaker A:** But what they were trying to do couldn't have been done through VCs.  
但他们想做的事情,靠风投是做不成的。  
**[42:51] Speaker A:** What OpenAI wanted to do couldn't have been done through VCs. I recognize that now. I didn't know it then. But that's their genius. That's why they're smart. They realized then that they had to do something like that. And I'm delighted that they did.  
OpenAI 想做的事情,靠风投是做不成的。我现在明白了,但当时不知道。这就是他们的天才之处,这就是他们聪明的地方。他们当时就意识到必须采取那样的方式。我很高兴他们做到了。  
**[43:07] Speaker A:** Even though we caused Anthropic to have to go to somebody else, I'm still happy that it happened.  
虽然我们导致 Anthropic 不得不去找别人,但我仍然很高兴这件事发生了。  
**[43:17] Speaker A:** Anthropic's existence is great for the world. I'm delighted for it.  
Anthropic 的存在对世界来说是件好事,我为此感到高兴。  
**[43:21] Speaker B:** I guess you still are making a ton of money, and you're making way more money quarter after quarter. It's still okay to have regrets.  
我想你们现在还是在赚很多钱,而且一个季度比一个季度赚得更多。有遗憾也是正常的。  
**[43:29] Speaker B:** So the question still arises. Okay, now that we're here and you have all this money that you keep making, what should Nvidia be doing with it?  
所以问题还是存在。好吧,既然现在你们有了这么多钱,而且还在不断赚钱,Nvidia 应该用这些钱做什么?  
**[43:39] Speaker B:** There's one answer which is that there's this whole middleman ecosystem that has popped up for converting CapEx into OpEx for these labs so that they can rent compute.  
有一个答案是,现在出现了一整个中间商生态系统,专门把资本支出转换成运营支出,让这些实验室可以租用算力。  
**[43:45] Speaker B:** Because the chips are really expensive, they make a lot of money over their lifetime because the AI models are getting better.  
因为芯片真的很贵,但它们在整个生命周期内能赚很多钱,因为 AI 模型在不断变好。  
**[43:53] Speaker B:** So the value that they generate, their tokens, is increasing, but they're expensive to set up.  
所以它们生成的价值,也就是 token 的价值在增加,但前期建设成本很高。  
**[43:57] Speaker B:** Nvidia has the money to do the CapEx. In fact, it's been reported, you are backstopping CoreWeave up to $6.3 billion and have invested $2 billion.  
Nvidia 有钱做资本支出。事实上,据报道,你们为 CoreWeave 提供了高达 63 亿美元的担保,还投资了 20 亿美元。  
**[44:08] Speaker B:** Why doesn't Nvidia become a cloud themselves? Why doesn't it become a hyperscaler themselves and rent this compute out? You have all this cash to do it.  
为什么 Nvidia 不自己做云服务?为什么不自己成为超大规模云服务商,把算力租出去?你们有足够的现金来做这件事。  
**[44:15] Speaker A:** This is a philosophy of the company, and I think it's wise.  
这是公司的理念,我认为这很明智。  
**[44:18] Speaker A:** We should do as much as needed, as little as possible.  
我们应该做必须做的事,但尽可能少做。  
**[44:24] Speaker A:** What that means is, the work that we do with building our computing platform, if we don't do it, I genuinely believe it doesn't get done.  
这意味着,我们在构建计算平台方面做的工作,如果我们不做,我真心相信就不会有人做。  
**[44:31] Speaker A:** If we didn't take the risk that we take—if we didn't build NVLink the way we built it, if we didn't build the whole stack, if we didn't create the ecosystem the way we did, if we didn't dedicate ourselves to 20 years of CUDA while losing money most of that time—if we didn't do it, nobody else would have done it.  
如果我们不承担我们承担的风险——如果我们不按照我们的方式构建 NVLink,如果我们不构建整个技术栈,如果我们不按照我们的方式创建生态系统,如果我们不在 CUDA 上投入 20 年时间,而且大部分时间都在亏钱——如果我们不做这些,就不会有其他人做。  
**[44:52] Speaker A:** If we didn't create all the CUDA-X libraries so that they're all domain-specific… A decade and a half ago, we pushed into domain-specific libraries because we realized that if we didn't create these domain-specific libraries, whether it's for ray tracing or image generation or even the early works of AI, these models, if we didn't create them, for data processing, structured data processing, or vector data processing, if we didn't create them, nobody would.  
如果我们不创建所有这些针对特定领域的 CUDA-X 库……十五年前,我们就开始推进领域专用库,因为我们意识到,如果我们不创建这些领域专用库,无论是用于光线追踪、图像生成,还是 AI 的早期工作、这些模型,如果我们不创建它们,用于数据处理、结构化数据处理或向量数据处理的库,如果我们不创建它们,就不会有人做。  
**[45:19] Speaker A:** I am completely certain of that.  
我们创建了一个用于计算光刻的库,叫 cuLitho。如果我们不创建它,就不会有人做。  
**[45:19] Speaker A:** We created a library for computational lithography called cuLitho. If we didn't create it, nobody would have.  
我们创建了一个用于计算光刻的库,叫 cuLitho。如果我们不创建它,就不会有人做。  
**[45:29] Speaker A:** So accelerated computing wouldn't advance the way it has if we didn't do what we did. So we should do that. We should dedicate our company, all of our might, wholeheartedly to go do that.  
所以如果我们不做我们做的事情,加速计算就不会像现在这样发展。所以我们应该做这件事。我们应该全力以赴,全心全意地投入公司所有力量去做这件事。  
**[45:41] Speaker A:** However, the world has lots of clouds. If I didn't do it, somebody would show up.  
不过,世界上有很多云服务商。如果我不做,也会有别人出现。  
**[45:46] Speaker A:** So following the recipe, the philosophy, of doing as much as needed but as little as possible—as little as possible—that philosophy exists in our company today.  
所以遵循这个原则,这个哲学——做必要的事,但尽可能少做——尽可能少做——这个哲学今天仍然存在于我们公司。  
**[45:58] Speaker A:** Everything I do, I do it with that lens.  
就云服务而言,如果我们不支持 CoreWeave 的存在,这些新兴云、这些 AI 云就不会存在。  
**[45:58] Speaker A:** In the case of clouds, if we didn't support CoreWeave to exist, these neoclouds, these AI clouds, wouldn't exist.  
就云服务而言,如果我们不支持 CoreWeave 的存在,这些新兴云、这些 AI 云就不会存在。  
**[46:11] Speaker A:** If we didn't help CoreWeave exist, they would not exist.  
如果我们不帮助 CoreWeave 存在,他们就不会存在。  
**[46:15] Speaker A:** If we didn't support Nscale, they wouldn't be where they are today.  
如果我们不支持 Nscale,他们不会有今天的成就。  
**[46:19] Speaker A:** If we didn't support Nebius, they wouldn't be what they are today. Now they're doing fantastically.  
如果我们不支持 Nebius,他们不会是今天的样子。现在他们发展得非常好。  
**[46:25] Speaker B:** Is that a business model?  
我们应该做必要的事,但尽可能少做。所以我们投资我们的生态系统。  
**[46:25] Speaker A:** We should do as much as needed, as little as possible. So we invest in our ecosystem.  
我们应该做必要的事,但尽可能少做。所以我们投资我们的生态系统。  
**[46:34] Speaker A:** Because I want our ecosystem to thrive. I want the architecture and AI to be able to connect with as many industries as possible, as many countries as possible, and make it possible for the planet to be built on AI and to be built on the American tech stack.  
因为我希望我们的生态系统蓬勃发展。我希望这个架构和 AI 能够连接尽可能多的行业、尽可能多的国家,让整个地球能够建立在 AI 之上,建立在美国的技术栈之上。  
**[46:56] Speaker A:** That vision is exactly what we're pursuing. Now, one of the things that you mentioned... There are so many great, amazing foundation model companies, and we try to invest in all of them.  
这个愿景正是我们在追求的。现在,你提到的一件事……有很多很棒的、了不起的基础模型公司,我们试图投资所有这些公司。  
**[47:08] Speaker A:** This is another thing that we do. We don't pick winners. We need to support everyone. It's part of our joy of doing so. It's imperative to our business. But we also go out of our way not to pick winners. So when I invest in one of them, I invest in all of them.  
这是我们做的另一件事。我们不挑选赢家。我们需要支持所有人。这是我们乐于做的事情的一部分。这对我们的业务至关重要。但我们也特意不去挑选赢家。所以当我投资其中一家时,我会投资所有公司。  
**[47:25] Speaker B:** Why do you go out of your way not to pick winners?  
你为什么特意不挑选赢家?  
**[47:29] Speaker A:** Because it's not our job to, number one. Number two, when Nvidia first started, there were 60 3D graphics companies. We are the only one that survived.  
因为第一,这不是我们的工作。第二,当 Nvidia 刚开始的时候,有 60 家 3D 图形公司。我们是唯一幸存下来的。  
**[47:42] Speaker A:** If you would have taken those 60 graphics companies and asked yourself which one was going to make it, Nvidia would be at the top of that list not to make it.  
如果你拿那 60 家图形公司,问自己哪一家会成功,Nvidia 会排在最不可能成功的名单的首位。  
**[47:53] Speaker A:** This is long before you, but Nvidia's graphics architecture was precisely wrong. It's not a little bit wrong. We created an architecture that was precisely wrong, and it was an impossible thing for developers to support.  
这是在你之前很久的事了,但 Nvidia 的图形架构完全是错的。不是有点错,而是完全错了。我们创造了一个完全错误的架构,对开发者来说是不可能支持的。  
**[48:07] Speaker A:** It was never going to make it. We reasoned about it from good first principles, but we ended up with the wrong solution. Everybody would have counted us out. And here we are. So I have enough humility to recognize that. Don't pick winners. Either let them all take care of themselves, or take care of all of them.  
它永远不可能成功。我们从良好的第一性原理出发进行推理,但最终得到了错误的解决方案。所有人都会认为我们出局了。但我们还在这里。所以我有足够的谦逊来认识到这一点。不要挑选赢家。要么让他们都自己照顾自己,要么照顾所有人。  
**[48:31] Speaker B:** One thing I didn't understand is you said, "Look, we're not prioritizing these neoclouds just because they are neoclouds and we want to prop them up." But you also listed a bunch of neoclouds and said they wouldn't exist if it wasn't for NVIDIA. How are those two things compatible?  
有一件事我不理解,你说「我们不会仅仅因为这些新兴云是新兴云就优先考虑它们,我们不想扶持它们」。但你也列举了一堆新兴云,说如果没有 NVIDIA 它们就不会存在。这两件事怎么能兼容?  
**[48:51] Speaker A:** First of all, they need to want to exist, and they come to ask us for help. When they want to exist and they have a business plan, expertise,  
首先,他们需要想要存在,他们来向我们寻求帮助。当他们想要存在,并且有商业计划、专业知识,  
**[49:01] Speaker A:** And the passion for it... They obviously have to have some capabilities themselves. But if, at the end of the day, they need some investment in order to get it off the ground, we would be there for them.  
以及对此的热情……他们显然必须自己具备一些能力。但如果到最后,他们需要一些投资才能起步,我们会支持他们。  
**[49:11] Speaker A:** But the sooner they get their flywheel going... Your question was, "Do we want to be in the financing business?" The answer is no.  
但他们越早让自己的飞轮转起来……你的问题是「我们想做融资业务吗?」答案是不想。  
**[49:26] Speaker A:** There are people in the financing business, and we'd rather work with all the people in the financing business than be a financier ourselves.  
有专门做融资业务的人,我们宁愿与所有做融资业务的人合作,也不愿自己成为融资方。  
**[49:30] Speaker A:** Our goal is to focus on what we do, keep our business model as simple as possible, and support our ecosystem.  
我们的目标是专注于我们所做的事,保持我们的商业模式尽可能简单,并支持我们的生态系统。  
**[49:41] Speaker A:** When someone like OpenAI needs an investment of a $30 billion scale because it's still before their IPO, and we deeply believe in them and I deeply believe that they're going to be an... Well, they're an extraordinary company already today.  
当像 OpenAI 这样的公司需要 300 亿美元规模的投资,因为他们还没有 IPO,而我们深信他们,我深信他们将会是一家……嗯,他们今天已经是一家非凡的公司了。  
**[49:58] Speaker A:** They're going to be an incredible company. The world needs them to exist.  
他们将会是一家令人难以置信的公司。世界需要他们存在。  
**[50:02] Speaker A:** The world wants them to exist. I want them to exist. They have the wind at their back.  
世界需要它们存在,我也希望它们存在。它们现在正处于顺风顺水的时期。  
**[50:06] Speaker A:** Let's support them and let them scale. Those investments we'll do because they need us to do it.  
让我们支持它们,帮助它们扩大规模。这些投资我们会做,因为它们需要我们这样做。  
**[50:20] Speaker A:** But we're not trying to do as much as possible. We're trying to do as little as possible.  
但我们的目标不是尽可能多做,而是尽可能少做。  
**[51:04] Speaker A:** I've linked the GitHub repo in the description. And if you have a tool that you've been wanting to build, you should make it happen. Go to cursor.com/dwarkesh to get started.  
我已经在描述中放了 GitHub 仓库的链接。如果你一直想开发某个工具,现在就该动手了。访问 cursor.com/dwarkesh 开始吧。  
**[51:13] Speaker A:** This may be an obvious question, but we've lived many years in this situation where there's a shortage of GPUs, and it's grown now because models are getting better.  
这个问题可能很明显,但我们已经在 GPU 短缺的情况下生活了很多年,而且现在因为模型越来越好,短缺问题更严重了。  
**[51:25] Speaker A:** We have a shortage of GPUs. Yes. Nvidia is known for divvying up the scarce allocation, not just based on high bidder, but rather on, "Hey, we want to make sure that these neoclouds exist. Let's give some to CoreWeave, let's give some to Crusoe, let's give some to Lambda." Why is it good for Nvidia?  
我们面临 GPU 短缺。是的。众所周知,Nvidia 在分配稀缺资源时,不只是看谁出价高,而是会考虑「嘿,我们要确保这些新兴云服务商能生存下来。给 CoreWeave 一些,给 Crusoe 一些,给 Lambda 一些。」这对 Nvidia 有什么好处?  
**[51:45] Speaker A:** First of all, would you agree with this characterization of fracturing the market?  
首先,你同意这种「分散市场」的说法吗?  
**[51:49] Speaker B:** No. No. Your premise is just wrong. We're sufficiently mindful about these things.  
不同意。你的前提就是错的。我们对这些事情考虑得很周全。  
**[51:59] Speaker B:** We're very mindful about these things. First of all, if you don't place a PO, all the talking in the world won't make a difference. Until we get a PO, what are we going to do?  
我们对这些事情非常谨慎。首先,如果你不下采购订单,说再多也没用。在我们收到订单之前,我们能做什么呢?  
**[52:12] Speaker B:** So the first thing is, we work really hard with everybody to get a forecast done, because these things take a long time to build, and the data centers take a long time to build.  
所以第一件事是,我们会非常努力地和每个人一起做好预测,因为这些东西需要很长时间来制造,数据中心也需要很长时间来建设。  
**[52:24] Speaker B:** We align ourselves with demand and supply and things like that through forecasting. Okay? That's job number one. Number two, we've tried to forecast with as many people as possible, but in the final analysis, you still have to place an order.  
我们通过预测来协调供需关系。明白吗?这是首要任务。第二,我们尽量和尽可能多的人做预测,但最终你还是得下订单。  
**[52:41] Speaker B:** Maybe, for whatever reason, you didn't place your order. What can I do? At some point, first in, first out. But beyond that,  
也许因为某些原因,你没有下订单。我能怎么办?到了某个时候,就是先到先得。但除此之外,  
**[52:52] Speaker B:** if you're not ready because your data center's not ready, or certain components aren't ready to enable you to stand up a data center, we might decide to serve another customer first.  
如果你还没准备好,因为你的数据中心还没建好,或者某些组件还没到位无法启动数据中心,我们可能会决定先服务另一个客户。  
**[53:04] Speaker B:** That's just maximizing the throughput of our own factory.  
这只是为了最大化我们自己工厂的产能利用率。  
**[53:10] Speaker B:** We might do some adjustments there. Aside from that, the prioritization is first in, first out. You've got to place a PO.  
我们可能会做一些调整。除此之外,优先级就是先到先得。你必须下采购订单。  
**[53:21] Speaker B:** If you don't place a PO... Now, of course, there are stories about that.  
如果你不下订单……当然,关于这个有很多传闻。  
**[53:27] Speaker A:** For example, all of this kind of started from an article about Larry and Elon having dinner with me where they begged for GPUs. That never happened. We absolutely had dinner. We absolutely had dinner, and it was a wonderful dinner. At no time did they beg for GPUs.  
比如,这一切都源于一篇文章,说 Larry 和 Elon 跟我吃饭时恳求我给他们 GPU。这从来没发生过。我们确实吃了饭,那是一顿很愉快的晚餐。但他们从来没有恳求过 GPU。  
**[53:51] Speaker A:** They just had to place an order. Once they place an order, we do our best to get the capacity to them. We're not complicated.  
他们只需要下订单就行了。一旦他们下了订单,我们就会尽力把产能交付给他们。我们没那么复杂。  
**[53:55] Speaker B:** Okay. So it sounds like there's a queue, and then based on whether your data center is ready and when you place a purchase order, you get them at a certain time. But it still doesn't sound like the highest bidder just gets it. Is there a reason to do it...?  
好的。所以听起来有一个队列,然后根据你的数据中心是否准备好以及你什么时候下采购订单,你会在特定时间拿到货。但听起来还是不是出价最高的人就能拿到。有什么理由这样做吗……?  
**[54:13] Speaker A:** We never do that.  
我们从不那样做。  
**[54:15] Speaker B:** Okay.  
我们从不那样做。  
**[54:15] Speaker A:** We never do.  
我们从不那样做。  
**[54:17] Speaker B:** Why not just do high bidder?  
因为那是糟糕的商业做法。你定好价格,然后人们决定买还是不买。  
**[54:17] Speaker A:** Because it's a bad business practice. You set your price and then people decide to buy it or not.  
因为那是糟糕的商业做法。你定好价格,然后人们决定买还是不买。  
**[54:31] Speaker A:** I understand that others in the chip industry change their prices when demand is higher, but we just don't.  
我知道芯片行业的其他公司会在需求高的时候涨价,但我们就是不这样做。  
**[54:39] Speaker A:** That's just never been a practice of ours. You can count on us. I prefer to be dependable, to be the foundation of the industry. You don't need to second-guess. If I quoted you a price, we quoted you a price. That's it. If demand goes through the roof, so be it.  
这从来不是我们的做法。你可以信赖我们。我更愿意成为可靠的那一方,成为这个行业的基石。你不需要去猜测。如果我给你报了价,那就是报价了。就这样。即使需求暴增,那也没关系。  
**[55:02] Speaker B:** On the other end, that's why you have a productive relationship with TSMC, right?  
从另一个角度看,这也是你们和 TSMC 保持良好合作关系的原因,对吧?  
**[55:05] Speaker A:** Yeah, Nvidia's been in business with them for, I guess, coming up on 30 years.  
是的,Nvidia 和他们合作了,我想,快 30 年了。  
**[55:14] Speaker A:** Nvidia and TSMC don't have a legal contract. There's always some rough justice. Sometimes I'm right, sometimes I'm wrong. Sometimes I got a better deal, sometimes I got a worse deal. But overall, the relationship is incredible. I can completely trust them. I can completely depend on them.  
Nvidia 和 TSMC 之间没有法律合同。总会有一些粗略的公平。有时我占便宜,有时我吃亏。有时我拿到更好的条件,有时条件差一些。但总体来说,这段关系非常好。我完全信任他们,完全可以依靠他们。  
**[55:37] Speaker A:** One of the things you can count on with Nvidia is that this year, Vera Rubin is going to be incredible. Next year, Vera Rubin Ultra will come.  
关于 Nvidia,你可以确信的一点是,今年 Vera Rubin 会非常出色。明年 Vera Rubin Ultra 会推出。  
**[55:46] Speaker A:** The year after that, Feynman will come. And the year after that,  
再下一年,Feynman 会推出。再下一年,  
**[55:48] Speaker A:** I haven't introduced the name yet. Every single year you can count on us.  
我还没公布名字。但每一年你都可以信赖我们。  
**[55:57] Speaker A:** You're going to have to go find another ASIC team in the world—pick your ASIC team—where you can say, "I can bet the farm, I can bet my entire business that you will be here for me every single year."  
你可以去找世界上任何一个 ASIC 团队——随便挑一个——看你能不能说:「我可以押上全部身家,押上整个生意,相信你每一年都会在这里支持我。」  
**[56:08] Speaker A:** Your token cost will decrease by an order of magnitude every single year. I can count on it like I can count on the clock.  
你的 token 成本每年会降低一个数量级。我可以像看钟表一样确定这一点。  
**[56:20] Speaker A:** I just said something about TSMC. For no other foundry in history can you possibly say that.  
我刚才说的关于 TSMC 的话,历史上没有其他任何晶圆厂能让你这么说。  
**[56:26] Speaker A:** You can say that about Nvidia today.  
但今天你可以这样评价 Nvidia。  
**[56:31] Speaker A:** You can count on us every single year. If you would like to buy a billion dollars worth of AI factory compute, no problem.  
你可以每年都信赖我们。如果你想买价值 10 亿美元的 AI 工厂算力,没问题。  
**[56:35] Speaker A:** If you'd like to buy a hundred million dollars, no problem. You'd like to buy $10 million, or just one rack, not a problem.  
如果你想买 1 亿美元的,没问题。你想买 1000 万美元的,或者只是一个机架,也没问题。  
**[56:43] Speaker A:** Or just one graphics card, okay, no problem.  
或者只是一块显卡,好的,没问题。  
**[56:49] Speaker A:** If you would like to place an order for a $100 billion AI factory, no problem.  
如果你想下单 1000 亿美元的 AI 工厂,没问题。  
**[56:54] Speaker A:** We're the only company in the world where you can say that today.  
我们是当今世界上唯一一家你可以这么说的公司。  
**[56:58] Speaker A:** I can say that about TSMC as well. I want to buy one, buy a billion, no problem.  
我对 TSMC 也可以这么说。我想买一个,买 10 亿个,都没问题。  
**[57:04] Speaker A:** We just have to go through the process of planning for it, and all the things that mature people do.  
我们只需要经历规划的流程,以及成熟的人该做的所有事情。  
**[57:12] Speaker A:** So I think this ability for Nvidia to be the foundation of the world's AI industry, this is a position that has taken us a couple of decades to arrive at. Enormous commitment, enormous dedication. The stability of our company, the consistency of our company, is really important.  
所以我认为,Nvidia 能够成为全球 AI 行业的基石,这个位置我们花了几十年才达到。巨大的投入,巨大的专注。我们公司的稳定性、一致性,真的非常重要。  
**[57:35] Speaker B:** Okay. I want to ask about China.  
好的。我想问问关于中国的问题。  
**[57:38] Speaker B:** I actually don't know what I think about whether it's good to sell chips to China or not, but I like to play devil's advocate against my guests.  
其实我自己也不知道该不该向中国出售芯片,但我喜欢和嘉宾唱反调。  
**[57:44] Speaker B:** So when Dario was on, who supports export controls, I asked him, why can't America and China both have a country of geniuses in the datacenter?  
所以当 Dario 来的时候,他支持出口管制,我就问他,为什么美国和中国不能都在数据中心里拥有一个天才之国?  
**[57:52] Speaker B:** But since you're on the opposite side, I'll ask you in the opposite way.  
但既然你站在相反的立场,我就反过来问你。  
**[57:58] Speaker B:** One way to think about it is, Anthropic actually announced a couple days ago Mythos Preview.  
可以这样想,Anthropic 几天前刚宣布了 Mythos Preview。  
**[58:02] Speaker B:** This model Mythos, they're not even releasing publicly because they say  
这个 Mythos 模型,他们甚至不打算公开发布,因为他们说  
**[58:05] Speaker A:** It has such cyber-offensive capabilities that we don't think the world is ready until we make sure these zero-days are patched up. But they say it found thousands of high-severity vulnerabilities across every major operating system, every browser.  
它具有如此强大的网络攻击能力，以至于我们认为在确保这些零日漏洞被修补之前，世界还没有准备好。但他们说它在每个主流操作系统、每个浏览器中都发现了数千个高危漏洞。  
**[58:18] Speaker A:** It found one in OpenBSD, which is this operating system that's been specifically designed to not have zero-days. It found one that's existed for 27 years.  
它在 OpenBSD 中发现了一个漏洞，而这个操作系统是专门设计来避免零日漏洞的。它发现的这个漏洞已经存在了 27 年。  
**[58:26] Speaker A:** So if Chinese companies and Chinese labs and the Chinese government had access to the AI chips to train a model like Claude Mythos with these cyber-offensive capabilities and run millions of instances of it with more compute, the question is, is that a threat to American companies, to American national security?  
所以如果中国公司、中国实验室和中国政府能够获得 AI 芯片，来训练像 Claude Mythos 这样具有网络攻击能力的模型，并用更多算力运行数百万个实例，问题就是：这对美国公司、对美国国家安全是否构成威胁？  
**[58:44] Speaker B:** First of all, Mythos was trained on fairly mundane capacity, and a fairly mundane amount of it, by an extraordinary company. The amount of capacity and the type of compute it was trained on is abundantly available in China.  
首先，Mythos 是用相当普通的算力训练的，而且算力量也很普通，只是训练它的公司非常出色。它训练所用的算力规模和类型，在中国是完全可以获得的。  
**[59:08] Speaker B:** So you just have to first realize that chips exist in China. They manufacture 60% of the world's mainstream chips, maybe more. It's a very large industry for them.  
所以你首先要意识到，中国是有芯片的。他们制造了全球 60% 的主流芯片，甚至可能更多。这对他们来说是一个非常大的产业。  
**[59:19] Speaker B:** They have some of the world's greatest computer scientists. As you know, most of the AI researchers in all of these AI labs are Chinese. They have 50% of the world's AI researchers.  
他们拥有世界上一些最顶尖的计算机科学家。你也知道，所有这些 AI 实验室里的大多数 AI 研究人员都是中国人。他们拥有全球 50% 的 AI 研究人员。  
**[59:39] Speaker B:** So the question is, considering all the assets they already have—they have an abundance of energy, they have plenty of chips, they've got most of the AI researchers—if you're worried about them, what is the best way to create a safe world?  
所以问题是，考虑到他们已经拥有的所有资源——他们有充足的能源，有大量的芯片，有大部分的 AI 研究人员——如果你担心他们，那么创造一个安全世界的最佳方式是什么？  
**[59:55] Speaker B:** Victimizing them, turning them into an enemy, likely isn't the best answer. They are an adversary. We want the United States to win.  
把他们当作受害者、把他们变成敌人，可能不是最好的答案。他们是竞争对手。我们希望美国获胜。  
**[01:00:16] Speaker B:** But I think having a dialogue and having research dialogue is probably the safest thing to do.  
但我认为进行对话、进行研究对话可能是最安全的做法。  
**[01:00:23] Speaker B:** This is an area that is glaringly missing because of our current attitude about China as an adversary. It is essential that our AI researchers and their AI researchers are actually talking.  
这是一个明显缺失的领域，因为我们目前把中国当作对手的态度。我们的 AI 研究人员和他们的 AI 研究人员实际进行交流是至关重要的。  
**[01:00:35] Speaker B:** It is essential that we try to both agree on what not to use the AI for. With respect to finding bugs in software,  
我们必须尝试就不应该将 AI 用于什么达成共识。关于在软件中查找漏洞这件事，  
**[01:00:49] Speaker A:** Of course, that's what AI is supposed to do. Is it going to find bugs in a lot of software? Of course. There are lots and lots of bugs. There are lots of bugs in the AI software.  
当然，这正是 AI 应该做的。它会在很多软件中发现漏洞吗？当然会。有大量大量的漏洞。AI 软件本身也有很多漏洞。  
**[01:01:03] Speaker A:** That's what AI is supposed to do, and I'm delighted that AI has reached a level where it could help us be so much more productive.  
这正是 AI 应该做的，我很高兴 AI 已经达到了这样的水平，可以帮助我们大幅提高生产力。  
**[01:01:08] Speaker A:** One of the things that is underemphasized is the richness of the ecosystem around cybersecurity, AI cybersecurity and AI security and AI privacy and AI safety.  
有一件被低估的事情，就是围绕网络安全、AI 网络安全、AI 安全、AI 隐私和 AI 安全性的生态系统是多么丰富。  
**[01:01:25] Speaker A:** There's a whole ecosystem of AI startups that are trying to create this future for us, where you have one AI agent that's incredible, surrounded by thousands of AI agents, keeping it safe, keeping it secure.  
有一整个 AI 初创公司生态系统正在努力为我们创造这样的未来：你有一个强大的 AI 智能体，周围有成千上万个 AI 智能体保护它的安全。  
**[01:01:46] Speaker A:** That future surely is going to happen. The idea that you're going to have an AI agent running around with nobody watching after it is kind of insane.  
这个未来肯定会实现。让一个 AI 智能体在没有任何监管的情况下到处运行，这种想法是相当疯狂的。  
**[01:01:58] Speaker A:** We know very well that this ecosystem needs to thrive.  
我们非常清楚这个生态系统需要蓬勃发展。  
**[01:02:02] Speaker A:** It turns out this ecosystem needs open source. This ecosystem needs open models. They need open stacks so that all of these AI researchers and all these great computer scientists can go build AI systems that are as formidable and can keep AI safe.  
事实证明，这个生态系统需要开源。这个生态系统需要开放模型。他们需要开放技术栈，这样所有这些 AI 研究人员和所有这些优秀的计算机科学家才能去构建同样强大的 AI 系统，来保证 AI 的安全。  
**[01:02:22] Speaker A:** So one of the things that we need to make sure that we do is we keep the open source ecosystem vibrant. That can't be ignored. A lot of that is coming out of China.  
所以我们需要确保做的一件事就是保持开源生态系统的活力。这一点不能被忽视。其中很多贡献来自中国。  
**[01:02:37] Speaker A:** We ought to not suffocate that. With respect to China, of course we want the United States to have as much computing as possible.  
我们不应该扼杀它。关于中国，我们当然希望美国拥有尽可能多的算力。  
**[01:02:50] Speaker A:** We're limited by energy, but we've got a lot of people working on that. We've got to not make energy a bottleneck for our country.  
我们受到能源的限制，但有很多人在解决这个问题。我们不能让能源成为我们国家的瓶颈。  
**[01:03:00] Speaker A:** But what we also want is to make sure that all the AI developers in the world are developing on the American tech stack, and making the contributions, the advancements of AI—especially when it's open source—available to the American ecosystem.  
但我们还想确保的是，世界上所有的 AI 开发者都在美国技术栈上开发，并且让 AI 的贡献和进步——尤其是开源的——能够为美国生态系统所用。  
**[01:03:14] Speaker A:** It would be extremely foolish to create two ecosystems: the open source ecosystem, and it only runs on a foreign tech stack, and a closed ecosystem that runs on the American tech stack. I think that would be a horrible  
创建两个生态系统将是极其愚蠢的：开源生态系统只在外国技术栈上运行，而封闭生态系统在美国技术栈上运行。我认为这对美国来说将是一个可怕的  
**[01:03:34] Speaker A:** Outcome for the United States. Since there are a lot of things, let me just triage the response. I think the concern, going back to the flop difference in the hacking, is yes, they have compute, but there's some estimates that because they're at 7nm—they don't have EUVs because of chip-making export controls—the amount of flops they're able to actually produce, they have one-tenth the amount of flops that the US has.  
结果。因为有很多事情要说，让我先分类回应一下。我认为关于黑客攻击中算力差异的担忧是，是的，他们有算力，但有一些估计认为，因为他们的制程是 7 纳米——由于芯片制造出口管制，他们没有 EUV 光刻机——他们实际能够产生的浮点运算量只有美国的十分之一。  
**[01:04:00] Speaker A:** So with that, could they eventually train a model like Mythos? Yes. But the question is, because we have more flops, American labs are able to get to these levels of capabilities first.  
所以说，他们最终能训练出像 Mythos 这样的模型吗？可以。但问题是，因为我们有更多的算力，美国实验室能够率先达到这些能力水平。  
**[01:04:12] Speaker A:** Because Anthropic got to it first, they say, "Okay, we're going to hold onto it for a month while all these American companies, we'll give them access to it. They're going to patch up all their vulnerabilities, and now we release it."  
因为 Anthropic 率先发现了这个漏洞,他们就说:「好,我们先保留一个月,在这期间把访问权限给所有这些美国公司,让他们修补好各自的漏洞,然后我们再公开发布。」  
**[01:04:22] Speaker A:** Furthermore, even if they train a model like this, the ability to deploy it at scale… If you had a cyber hacker, it's much more dangerous if they have a million of them versus a thousand of them. So that inference compute really matters a lot. In fact, the fact that they have so many AI researchers who are so good is the thing that makes it so scary, because what is it that makes those engineer researchers more productive? It's compute.  
而且,即使他们训练出这样的模型,大规模部署的能力也很关键……如果你有一个网络黑客,拥有一百万个显然比拥有一千个危险得多。所以推理算力真的非常重要。事实上,正是因为他们有这么多优秀的 AI 研究人员,才让这件事如此可怕,因为是什么让这些工程研究人员更高效呢?是算力。  
**[01:04:44] Speaker A:** If you talk to any AI lab in America, they say the thing that's bottlenecking them is compute. There are quotes from the DeepSeek founder, or Qwen leadership or whatever. They say the thing they're bottlenecked on is compute.  
如果你跟美国任何一家 AI 实验室聊,他们都会说瓶颈在算力。DeepSeek 创始人或者 Qwen 领导层也有类似的引述,他们说瓶颈就是算力。  
**[01:04:54] Speaker A:** So then the question is, isn't it better that we get American companies, because they have more compute, to get to the Mythos-level capabilities first, prepare our society for it, before China can get to it because they have less compute? We should always be first and we should always have more.  
那么问题就来了,让美国公司因为拥有更多算力而率先达到 Mythos 级别的能力,让我们的社会提前做好准备,这难道不是更好吗?总好过中国因为算力更少而后达到。我们应该始终领先,应该始终拥有更多。  
**[01:05:11] Speaker B:** But in order for that outcome you described to be true, you have to take it to the extremes. They have to have no compute.  
但要让你描述的那种结果成真,你必须把它推向极端。他们必须完全没有算力才行。  
**[01:05:26] Speaker A:** If they have some compute, the question is how much is needed? The amount of compute they have in China is enormous.  
如果他们有一些算力,问题就是需要多少?中国拥有的算力规模是巨大的。  
**[01:05:34] Speaker A:** You're talking about the country that is the second largest computing market in the world. If they want to aggregate their compute, they've got plenty of compute to aggregate.  
你说的可是全球第二大计算市场。如果他们想整合算力,他们有大量算力可以整合。  
**[01:05:44] Speaker B:** But is that true? People do these estimates and they're like, "SMIC is actually behind on the process nodes."  
但真的是这样吗?有人做过估算,他们会说:「SMIC 在制程节点上其实是落后的。」  
**[01:05:48] Speaker A:** I'm about to tell you.  
我正要告诉你。  
**[01:05:51] Speaker B:** Okay.  
好。  
**[01:05:52] Speaker A:** The amount of energy they have is incredible. Isn't that right? AI is a parallel computing problem, isn't it?  
他们拥有的能源量是惊人的,不是吗?AI 本质上是并行计算问题,对吧?  
**[01:05:58] Speaker A:** Why can't they just put 4x, 10x as many chips together because energy's free? They have so much energy. They have datacenters that are sitting completely empty, fully powered.  
为什么他们不能因为能源几乎免费,就把 4 倍、10 倍数量的芯片组合在一起?他们有那么多能源,有完全空置但已经通电的数据中心。  
**[01:06:11] Speaker A:** You know they have ghost cities, they have ghost datacenters too.  
你知道他们有鬼城,他们也有鬼数据中心。  
**[01:06:14] Speaker A:** They have so much infrastructure capacity. If they wanted to, they just gang up more chips, even if they're 7nm.  
他们有如此多的基础设施产能。如果他们想,完全可以把更多芯片组合起来,哪怕是 7nm 的芯片。  
**[01:06:20] Speaker A:** Their capacity of building chips is one of the largest in the world.  
他们的芯片制造能力是全球最大的之一。  
**[01:06:24] Speaker A:** The semiconductor industry knows that they monopolize mainstream chips.  
半导体行业都知道,他们垄断了主流芯片市场。  
**[01:06:30] Speaker A:** They have over-capacity, they have too much capacity.  
他们产能过剩,产能太多了。  
**[01:06:33] Speaker A:** So the idea that China won't be able to have AI chips is completely nonsense.  
所以认为中国无法拥有 AI 芯片的想法完全是无稽之谈。  
**[01:06:37] Speaker A:** Now, of course, if you ask me, would the United States be further ahead if the entire world had no compute at all?  
当然,如果你问我,假如全世界都完全没有算力,美国是不是会领先更多?  
**[01:06:45] Speaker A:** But that's just not an outcome. That's not a scenario that's true.  
但那根本不是现实结果,那不是真实的场景。  
**[01:06:51] Speaker A:** They have plenty of compute already.  
他们已经有足够的算力了。  
**[01:06:55] Speaker A:** The amount of threshold they need for the concern you're worried about, they've already reached that threshold and beyond.  
对于你担心的那个问题,他们需要达到的算力门槛,他们早就达到甚至超过了。  
**[01:06:59] Speaker A:** So I think you misunderstand that AI is a five-layer cake, and at the lowest layer is energy.  
所以我觉得你误解了,AI 是一个五层蛋糕,最底层是能源。  
**[01:07:04] Speaker A:** When you have an abundance of energy, it makes up for chips.  
当你拥有充足的能源,就可以弥补芯片的不足。  
**[01:07:10] Speaker A:** If you have an abundance of chips, it makes up for energy.  
如果你拥有充足的芯片,就可以弥补能源的不足。  
**[01:07:14] Speaker A:** For example, the United States is scarce on energy, which is the reason why Nvidia has to keep advancing our architecture and do this extreme co-design so that with the few chips that we ship—with the few chips,  
举个例子,美国的能源是稀缺的,这就是为什么 Nvidia 必须不断推进我们的架构,进行极致的协同设计,这样即使我们出货的芯片数量很少——芯片数量很少的情况下,  
**[01:07:35] Speaker A:** Because the amount of energy is so limited, our throughput per watt is off the charts. But if your amount of watts is completely abundant, it's free, what do you care about performance per watt for? You get plenty. You can use old chips to do it.  
因为能源供应非常有限,我们每瓦特的吞吐量是极高的。但如果你的瓦特数完全充足,甚至是免费的,你还在乎什么每瓦特性能呢?你有的是能源。用老芯片就能做到。  
**[01:07:51] Speaker A:** So 7nm chips are essentially Hopper. The ability for Hopper... I've got to tell you, today's models are largely trained on Hopper, Hopper generation.  
所以 7nm 芯片本质上就是 Hopper。Hopper 的能力……我得告诉你,今天的模型大部分都是在 Hopper 上训练的,Hopper 这一代。  
**[01:08:07] Speaker A:** So 7nm chips are plenty good. The abundance of energy is their advantage.  
所以 7nm 芯片已经足够好了。充足的能源供应就是他们的优势。  
**[01:08:12] Speaker A:** But then there's a question of whether they can actually manufacture enough chips.  
但接下来的问题是,他们是否真的能制造出足够多的芯片。  
**[01:08:18] Speaker A:** But they do. What's the evidence? Huawei just had the largest single year in the history of their company. How many chips did they ship?  
但他们做到了。证据是什么?Huawei 刚刚创下了公司历史上最大的单年业绩。他们出货了多少芯片?  
**[01:08:27] Speaker A:** A ton. Millions. Millions is way more than Anthropic has.  
大量。数百万片。数百万片远远超过 Anthropic 拥有的数量。  
**[01:08:35] Speaker B:** There's a question of how much logic SMIC can ship, and there's a question of how much memory—  
有一个问题是 SMIC 能出货多少逻辑芯片,还有一个问题是能出货多少内存——  
**[01:08:39] Speaker A:** I'm telling you what it is. They have plenty of logic, and they have plenty of HBM2 memory.  
我告诉你实际情况是什么。他们有充足的逻辑芯片,也有充足的 HBM2 内存。  
**[01:08:42] Speaker B:** Right. But as you know, the bottleneck often in training and doing inference on these models is the amount of bandwidth.  
对。但你知道,训练和推理这些模型时的瓶颈往往是带宽。  
**[01:08:51] Speaker B:** So if you have HBM2... I don't know the numbers offhand, but versus the newest thing you have, there could be almost an order of magnitude difference in memory bandwidth, which is huge.  
所以如果你用的是 HBM2……我不记得具体数字,但和你们最新的产品相比,内存带宽可能相差近一个数量级,这是巨大的差距。  
**[01:09:02] Speaker A:** Huawei is a networking company.  
Huawei 是一家网络公司。  
**[01:09:04] Speaker B:** But that doesn't change the fact that you need EUV for the most advanced HBM.  
不对。完全不对。你可以把它们组合在一起,就像我们用 NVL72 把它们组合在一起一样。  
**[01:09:04] Speaker A:** Not true. Not at all true. You could gang them together, just like we gang them together with NVL72.  
不对。完全不对。你可以把它们组合在一起,就像我们用 NVL72 把它们组合在一起一样。  
**[01:09:14] Speaker A:** They've already demonstrated silicon photonics, connecting all of this compute together into one giant supercomputer. Your premise is just wrong. The fact of the matter is, their AI development is going just fine.  
他们已经展示了硅光子技术,把所有这些算力连接成一台巨型超级计算机。你的前提就是错的。事实是,他们的 AI 发展进展得很好。  
**[01:09:33] Speaker A:** The best AI researchers in the world, because they're limited in compute, they also come up with extremely smart algorithms.  
世界上最优秀的 AI 研究人员,正因为算力受限,他们也会想出极其聪明的算法。  
**[01:09:39] Speaker A:** Remember, I just said that Moore's law is advancing about 25% per year.  
记住,我刚才说过摩尔定律每年推进大约 25%。  
**[01:09:45] Speaker A:** However, through great computer science, we could still improve algorithm performance by 10x.  
然而,通过出色的计算机科学,我们仍然可以将算法性能提升 10 倍。  
**[01:09:52] Speaker A:** What I'm saying is that great computer science is where the lever is.  
我想说的是,出色的计算机科学才是关键杠杆。  
**[01:09:58] Speaker A:** There is no question, MoE is a great invention. There's no question, all the incredible attention mechanisms reduce the amount of compute. We have got to acknowledge that most of the advances in AI came out of algorithm advances, not just the raw hardware.  
毫无疑问,MoE 是一项伟大的发明。毫无疑问,所有这些出色的注意力机制都减少了计算量。我们必须承认,AI 的大部分进步来自算法的进步,而不仅仅是原始硬件。  
**[01:10:19] Speaker A:** Now, if most advances came from algorithms and computer science and programming, tell me that their army of AI researchers is not their fundamental advantage.  
现在,如果大部分进步来自算法、计算机科学和编程,那你告诉我,他们庞大的 AI 研究人员队伍难道不是他们的根本优势吗?  
**[01:10:25] Speaker A:** We see it. DeepSeek is not an inconsequential advance. The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation.  
我们看到了。DeepSeek 不是一个无关紧要的进步。如果 DeepSeek 首先在 Huawei 上发布,那对我们国家来说将是一个可怕的结果。  
**[01:10:40] Speaker A:** Why is that? Because currently you can have a model like DeepSeek that can run on any accelerator, if it's open source.  
为什么?因为目前你可以有像 DeepSeek 这样的模型,如果它是开源的,就可以在任何加速器上运行。  
**[01:10:48] Speaker B:** Why would that stop being the case in the future?  
为什么未来就不会是这样了?  
**[01:10:50] Speaker A:** Suppose it doesn't. Suppose it's optimized for Huawei, suppose it's optimized for their architecture. It would put ours at a disadvantage.  
假设不是这样。假设它是为 Huawei 优化的,假设它是为他们的架构优化的。这会让我们处于劣势。  
**[01:10:58] Speaker B:** You described a situation that I perceive to be good news.  
你描述的情况,在我看来是个好消息。  
**[01:11:06] Speaker B:** A company developed software, developed an AI model, and it runs best on the American tech stack. I saw that as good news.  
一家公司开发了软件,开发了 AI 模型,而且它在美国技术栈上运行得最好。我认为这是好消息。  
**[01:11:15] Speaker B:** You set it up as a premise that it was bad news. I'm going to give you the bad news, that AI models around the world are developed and they run best on non-American hardware.  
但你把它设定为一个坏消息的前提。我要告诉你真正的坏消息:全球各地开发的 AI 模型在非美国硬件上运行得最好。  
**[01:11:27] Speaker B:** That is bad news for us. I guess I just don't see the evidence that there's these huge disparities that would prevent you from switching accelerators.  
那对我们来说才是坏消息。我只是看不到有什么证据表明存在巨大差异,会阻止你切换加速器。  
**[01:11:33] Speaker B:** American labs are running their models across all the clouds, across all the different accelerators—  
美国实验室在所有云平台上、所有不同的加速器上运行他们的模型——  
**[01:11:37] Speaker A:** I am the evidence. You take a model that's optimized for Nvidia and you try to run it on something else.  
我就是证据。你拿一个为 Nvidia 优化的模型,试着在别的硬件上运行。  
**[01:11:41] Speaker B:** But American labs do that.  
但美国实验室确实在这么做。  
**[01:11:44] Speaker A:** And they don't run better. Nvidia's success is perfect evidence. The fact that AI models are created on our stack, run best on our stack, how is that illogical to understand?  
但运行效果并不更好。Nvidia 的成功就是完美的证据。AI 模型在我们的技术栈上创建,在我们的技术栈上运行得最好,这有什么难理解的?  
**[01:11:58] Speaker B:** Anthropic's models are run on GPUs, they're run on Trainium, they're run on TPUs.  
Anthropic 的模型在 GPU 上运行,在 Trainium 上运行,也在 TPU 上运行。  
**[01:12:02] Speaker A:** A lot of work has to go into it to change. But go to the global south, go to the Middle East.  
要做出改变需要投入大量工作。但去全球南方看看,去中东看看。  
**[01:12:07] Speaker A:** Coming out of the box, if all of the AI models run best on somebody else's tech stack, you've got to be arguing some ridiculous claim right now that that's a good thing for the United States.  
开箱即用的情况下,如果所有 AI 模型在别人的技术栈上运行得最好,你现在居然要论证说这对美国是好事,这简直荒谬。  
**[01:12:18] Speaker B:** But I guess I don't understand the argument. Say Chinese companies get to the next Mythos first. They find all the security vulnerabilities in American software first, but they can do it on Nvidia hardware and they ship it to the global south. They do it on Nvidia hardware. How is that good?  
但我不太理解你的论点。假设中国公司率先开发出下一个 Mythos。他们率先发现美国软件中的所有安全漏洞,但他们可以在 Nvidia 硬件上做到这一点,然后把它运到全球南方。他们用的是 Nvidia 硬件。这怎么就好了?  
**[01:12:33] Speaker A:** Okay, it runs on Nvidia hardware— It's not good. It's not good.  
好吧,它在 Nvidia 硬件上运行——这不好。这不是好事。  
**[01:12:36] Speaker B:** Right.  
这不是好事。所以我们不能让它发生。  
**[01:12:36] Speaker A:** It's not good. So let's not let it happen.  
这不是好事。所以我们不能让它发生。  
**[01:12:39] Speaker B:** Why do you think it's perfectly fungible, that if you didn't ship them compute it would exactly be replaced by Huawei? They are behind, right? They have worse chips than you.  
你为什么认为这是完全可替代的,如果你不给他们提供算力,就会被 Huawei 完全替代?他们是落后的,对吧?他们的芯片比你们的差。  
**[01:12:46] Speaker A:** It's completely… There's evidence right now. Their chip industry's gigantic.  
这完全……现在就有证据。他们的芯片产业规模巨大。  
**[01:12:49] Speaker B:** You can just look at the FLOP or bandwidth or memory comparisons between the H200 and the Huawei 910C. It's like half to a third.  
你可以直接看 H200 和 Huawei 910C 之间的浮点运算、带宽或内存对比。大概是一半到三分之一。  
**[01:12:56] Speaker A:** They use more of it. They use twice as many. It seems like your argument is they have all this energy that's ready to go, right? And they need to fill it with chips.  
他们用更多的芯片。他们用两倍的数量。你的论点似乎是他们有大量能源准备就绪,对吧?他们需要用芯片来填充。  
**[01:13:03] Speaker B:** And they're good at manufacturing. And I'm sure eventually they would be able to just out-manufacture everybody. But there are these few critical years.  
而且他们擅长制造。我相信最终他们能够在制造上超越所有人。但有这么几年关键时期。  
**[01:13:10] Speaker A:** What is the critical year you're talking about?  
接下来这几年。我们会有能够执行所有网络攻击的模型。  
**[01:13:10] Speaker B:** These next few years. We've got these models that are going to be able to do all the cyber attacks.  
接下来这几年。我们会有能够执行所有网络攻击的模型。  
**[01:13:14] Speaker A:** In that case, if the next years are critical, then we have to make sure that all of the world's AI models are built on the American tech stack, in these critical years.  
如果接下来几年是关键时期,那我们就必须确保全世界的 AI 模型都建立在美国技术栈上,在这些关键年份里。  
**[01:13:22] Speaker B:** If they're built on the American tech stack, how would that prevent them, if they have more advanced capabilities, from launching the Mythos-equivalent cyber attacks?  
如果它们建立在美国技术栈上,如果它们具有更先进的能力,这怎么能阻止它们发动相当于 Mythos 级别的网络攻击?  
**[01:13:32] Speaker A:** There's no guarantee either way.  
无论哪种方式都没有保证。  
**[01:13:35] Speaker B:** But if you have it early, we can prepare for it. Listen, why are you causing one layer of the AI  
但如果你能提前拿到,我们就可以做好准备。听着,你为什么要让 AI 产业的某一层失去整个市场  
**[01:13:44] Speaker A:** industry to lose an entire market so that you could benefit another layer of the AI industry?  
就为了让 AI 产业的另一层受益?  
**[01:13:54] Speaker B:** There are five layers and every single layer has to succeed.  
AI 产业有五个层级,每一层都必须成功。  
**[01:13:59] Speaker B:** The layer that has to succeed most is actually the AI applications.  
其中最需要成功的那一层,其实是 AI 应用层。  
**[01:14:05] Speaker A:** Why are you so fixated on that AI model? That one company? For what reason?  
你为什么这么执着于那个 AI 模型?那一家公司?到底为什么?  
**[01:14:10] Speaker B:** Because those models make possible these incredibly offensive capabilities, and you need compute to run them. The energy, the chips, and the ecosystem of AI researchers make it possible.  
因为这些模型让那些极其强大的能力成为可能,而你需要算力来运行它们。能源、芯片,以及 AI 研究者的生态系统,这些都让一切成为可能。  
**[01:14:18] Speaker A:** A few months ago, Jane Street spent about 20,000 GPU hours training backdoors into three different language models. Then they challenged my audience to find the trigger phrases.  
几个月前,Jane Street 花了大约 20,000 GPU 小时在三个不同的语言模型中训练后门。然后他们向我的听众发起挑战,让大家找出触发短语。  
**[01:14:29] Speaker A:** I just caught up with Ricson who designed the puzzle about some of the solutions that Jane Street received.  
我刚刚和设计这个谜题的 Ricson 聊了聊 Jane Street 收到的一些解决方案。  
**[01:14:31] Speaker C:** If you think the base model was here and the backdoor model was here, you can kind of linearly interpolate the weights to adjust the strength of the backdoor, but you can also extrapolate it to make the backdoor even stronger. And in some cases, if you make it strong enough the model will just regurgitate what the response phrase was supposed to be.  
「如果你把基础模型看作这里,后门模型看作那里,你可以对权重进行线性插值来调整后门的强度,但你也可以外推让后门变得更强。在某些情况下,如果你把它做得足够强,模型就会直接吐出本该是响应短语的内容。」  
**[01:14:49] Speaker A:** So if you keep amplifying the difference between the base version and the backdoored version, eventually it should spit out the trigger phrase.  
所以如果你不断放大基础版本和后门版本之间的差异,最终它应该会吐出触发短语。  
**[01:14:57] Speaker A:** But this technique only worked on two out of the three models.  
但这个技术只在三个模型中的两个上有效。  
**[01:15:00] Speaker A:** Even Ricson isn't sure why it didn't work on the other.  
就连 Ricson 也不确定为什么它在另一个模型上不起作用。  
**[01:15:02] Speaker A:** Being able to verify that a model only does what you think it does is one of the most important open questions in AI security.  
能够验证一个模型只做你认为它会做的事,这是 AI 安全领域最重要的开放性问题之一。  
**[01:15:05] Speaker A:** If this is the kind of problem that excites you, Jane Street is hiring researchers and engineers. Go to janestreet.com/dwarkesh to learn more.  
如果这类问题让你兴奋,Jane Street 正在招聘研究员和工程师。访问 janestreet.com/dwarkesh 了解更多。  
**[01:15:15] Speaker B:** Okay, stepping back, it has to be the case that China is able to build enough 7nm capacity.  
好,退一步说,中国必须能够建立足够的 7nm 产能。  
**[01:15:21] Speaker B:** And remember, they're still stuck on 7nm while you'll move on to 3nm and then 2nm or 1.6nm with Feynman.  
记住,他们还停留在 7nm,而你们会推进到 3nm,然后是 2nm 或 1.6nm 的 Feynman 工艺。  
**[01:15:24] Speaker B:** So while you're on 1.6nm, they're still going to be on 7nm, and they have to produce enough of it to make up for the shortfall.  
所以当你们用上 1.6nm 时,他们还会停留在 7nm,而且他们必须生产足够多的 7nm 芯片来弥补差距。  
**[01:15:34] Speaker A:** They have so much energy that the more chips you give them, the more compute they'd have. So it comes out as a question of, ultimately they are getting more compute.  
他们有那么多能源,你给他们越多芯片,他们就有越多算力。所以归根结底的问题是,他们最终会获得更多算力。  
**[01:15:43] Speaker A:** Compute is an input to training and inference. Listen, I just think you speak in absolutes.  
算力是训练和推理的投入——听着,我只是觉得你说话太绝对了。  
**[01:15:48] Speaker B:** I think the United States ought to be ahead. The amount of compute in the United States is 100x more than anywhere else in the world. The United States ought to be ahead.  
我认为美国应该领先。美国的算力是世界其他任何地方的 100 倍。美国应该领先。  
**[01:16:00] Speaker B:** The United States is ahead. Nvidia builds the most advanced technologies.  
美国确实领先。Nvidia 打造最先进的技术。  
**[01:16:04] Speaker B:** We make sure that the US labs are the first to hear about it and have the first chance to buy it.  
我们确保美国的实验室最先听到消息,并有优先购买权。  
**[01:16:09] Speaker B:** And if they don't have enough money, we even invest in them.  
如果他们资金不够,我们甚至会投资他们。  
**[01:16:13] Speaker B:** The United States ought to be ahead. We want to do everything we can to make sure the United States is ahead. Number one point, do you agree?  
美国应该领先。我们想尽一切办法确保美国领先。第一点,你同意吗?  
**[01:16:22] Speaker B:** We're doing everything we can to do that.  
我们正在竭尽全力做到这一点。  
**[01:16:22] Speaker A:** But how is shipping chips to China keeping the US ahead if they're bottlenecked on compute?  
但如果美国自己在算力上遇到瓶颈,向中国出口芯片怎么能让美国保持领先呢?  
**[01:16:26] Speaker B:** No, no. We've got Vera Rubin for the United States. We have Vera Rubin for the United States.  
不不不,美国有 Vera Rubin。我们美国有 Vera Rubin。  
**[01:16:33] Speaker B:** Now, am I in the United States? Do you consider me part of the United States?  
那我现在在美国吗?你认为我是美国的一部分吗?  
**[01:16:38] Speaker A:** Yes.  
Nvidia 呢,你认为 Nvidia 是美国公司吗?好吧。  
**[01:16:38] Speaker B:** Nvidia. You consider Nvidia a United States company? Okay.  
Nvidia 呢,你认为 Nvidia 是美国公司吗?好吧。  
**[01:16:40] Speaker B:** Number one, why is it that we don't come up with a regulation that's more balanced so that Nvidia can win around the world instead of giving up the world?  
第一,为什么我们不制定一个更平衡的监管政策,让 Nvidia 能在全球市场获胜,而不是放弃全球市场?  
**[01:16:56] Speaker B:** Why would you want the United States to give up the world?  
你为什么希望美国放弃全球市场?  
**[01:17:00] Speaker B:** The chip industry is part of the American ecosystem.  
芯片产业是美国生态系统的一部分。  
**[01:17:03] Speaker B:** It's part of American technology leadership. It's part of the AI ecosystem. It's part of AI leadership.  
它是美国技术领导力的一部分,是 AI 生态系统的一部分,是 AI 领导力的一部分。  
**[01:17:08] Speaker B:** Why is it that your policy, your philosophy, leads to the United States giving up a vast part of the world's market?  
为什么你的政策、你的理念会导致美国放弃全球市场的很大一部分?  
**[01:17:16] Speaker A:** I guess the claim here is, Dario had this quote where he said that it's like Boeing bragging that we're selling North Korea nukes, but the missile casings are made by Boeing.  
我想这里的观点是,Dario 有句话说,这就像 Boeing 吹嘘说我们在向朝鲜出售核武器,但导弹外壳是 Boeing 制造的。  
**[01:17:27] Speaker A:** And that's somehow enabling the US technology stack. Fundamentally, you're giving them this capability.  
然后说这在某种程度上促进了美国的技术体系。但本质上,你是在给他们提供这种能力。  
**[01:17:34] Speaker A:** Comparing AI to anything that you just mentioned is lunacy.  
把 AI 和你刚才提到的任何东西相比都是荒谬的。  
**[01:17:37] Speaker A:** But AI is similar to enriched uranium, right? It can have positive uses, it can have negative uses. We still don't want to send enriched uranium to other countries.  
但 AI 类似于浓缩铀,对吧?它可以有积极用途,也可以有消极用途。我们仍然不想把浓缩铀送到其他国家。  
**[01:17:44] Speaker B:** Who's sending enriched—  
谁在运送浓缩——  
**[01:17:48] Speaker A:** The analogy is that enriched uranium is like compute.  
这个类比是说浓缩铀就像算力。  
**[01:17:51] Speaker B:** It's a lousy analogy. It's an illogical analogy. But if that compute can run a model that can do zero-day exploits against all American software, how is that not a weapon?  
这是个糟糕的类比,是个不合逻辑的类比。但如果那些算力可以运行一个能对所有美国软件进行零日漏洞攻击的模型,这怎么不算武器?  
**[01:18:04] Speaker A:** First of all, the way to solve that problem is to have dialogues with the researchers and dialogues with China, and dialogues with all the countries to make sure that people don't use technology in that way. That's a dialogue that has to happen. Okay?  
首先,解决这个问题的方法是与研究人员对话,与中国对话,与所有国家对话,确保人们不会以那种方式使用技术。这种对话必须进行,好吗?  
**[01:18:16] Speaker A:** Number one. Number two, we also need to make sure that the United States is ahead, that Vera Rubin, Blackwell, is available in the United States in abundance, mountains of it.  
第一点。第二点,我们还需要确保美国保持领先,确保 Vera Rubin、Blackwell 在美国大量供应,堆积如山。  
**[01:18:31] Speaker A:** Obviously, our results would show it. Abundance, tons of it.  
显然,我们的成果会证明这一点。大量供应,成吨的。  
**[01:18:36] Speaker A:** The amount of computing we have is great. We have amazing AI researchers here.  
我们拥有的算力规模很大。我们这里有出色的 AI 研究人员。  
**[01:18:40] Speaker A:** It's great. We ought to stay ahead. However, we also have to recognize that AI is not just a model. AI is a five-layer cake. The AI industry matters across every single layer, and we want the United States to win at every single layer, including the chip layer.  
很好。我们应该保持领先。但是,我们也必须认识到 AI 不仅仅是一个模型。AI 是一个五层蛋糕。AI 产业在每一层都很重要,我们希望美国在每一层都获胜,包括芯片层。  
**[01:18:59] Speaker A:** Conceding the entire market is not going to allow the United States to win the technology race long-term in the chip layer, in the computing stack. That is just a fact.  
放弃整个市场不会让美国在芯片层、在计算堆栈上长期赢得技术竞赛。这就是事实。  
**[01:19:10] Speaker B:** I guess then the crux comes down to, how does selling them chips now help us win in the long term?  
我想那么关键就在于,现在向他们出售芯片如何帮助我们长期获胜?  
**[01:19:19] Speaker A:** Tesla sold extremely good electric vehicles to China for a long time. iPhones are sold in China, extremely good. They didn't cause them lock-in. China will still make their version of EVs and they're dominating. Their smartphones are dominating.  
Tesla 长期向中国出售极其优秀的电动汽车。iPhone 在中国销售,极其优秀。它们并没有造成锁定效应。中国仍然会制造自己版本的电动汽车,而且正在占据主导地位。他们的智能手机也在占据主导地位。  
**[01:19:28] Speaker B:** When we started the conversation today, you acknowledged that Nvidia's position is very different. You used words like moat. The single  
今天对话开始时,你承认 Nvidia 的地位非常不同。你用了「护城河」这样的词。最  
**[01:19:40] Speaker A:** Most important thing to our company is the richness of our ecosystem, which is about developers. 50% of the AI developers are in China. The United States should not give that up.  
对我们公司来说最重要的是我们生态系统的丰富性,也就是开发者。全球 50% 的 AI 开发者在中国。美国不应该放弃这一点。  
**[01:19:53] Speaker A:** But we have a lot of Nvidia developers in the US, and that doesn't prevent American labs from also being able to use other accelerators in the future.  
但我们在美国也有大量 Nvidia 开发者,这并不妨碍美国实验室未来也能使用其他加速器。  
**[01:19:59] Speaker A:** In fact, right now they're using other accelerators as well, which is fine and great.  
事实上,现在他们也在使用其他加速器,这很好,没问题。  
**[01:20:03] Speaker B:** I don't see why that wouldn't be the case in China as well, if you sell them Nvidia chips, just the same way that Google can use TPUs and Nvidia—  
我不明白为什么在中国就不能是同样的情况,如果你卖给他们 Nvidia 芯片的话,就像 Google 可以同时使用 TPU 和 Nvidia——  
**[01:20:06] Speaker A:** We have to keep innovating and, as you probably know, our share is growing, not decreasing.  
我们必须持续创新,而且你可能知道,我们的市场份额在增长,不是在下降。  
**[01:20:12] Speaker A:** The premise that even if we competed in China, that we're going to lose that market anyways... You're not talking to somebody who woke up a loser.  
那种「即使我们在中国竞争,反正也会失去那个市场」的前提……你面对的不是一个一觉醒来就认输的人。  
**[01:20:26] Speaker A:** That loser attitude, that loser premise makes no sense to me.  
那种失败者的态度,那种失败者的前提,对我来说毫无意义。  
**[01:20:33] Speaker A:** We're not a car. We are not a car. The fact that I can buy this car brand one day and use another car brand another day, easy. Computing is not like that.  
我们不是汽车。我们不是汽车。你今天可以买这个品牌的车,明天用另一个品牌的车,很容易。但计算不是这样的。  
**[01:20:49] Speaker A:** There's a reason why the x86 deal exists. There's a reason why ARM is so sticky.  
x86 协议之所以存在是有原因的。ARM 之所以如此有粘性也是有原因的。  
**[01:20:55] Speaker A:** These ecosystems are hard to replace. It costs an enormous amount of time and energy, and most people don't want to do it.  
这些生态系统很难替代。需要投入大量的时间和精力,大多数人不愿意这么做。  
**[01:20:59] Speaker A:** So it's our job to continue to nurture that ecosystem, to keep advancing the technology so that we can compete in the marketplace.  
所以我们的工作就是继续培育这个生态系统,不断推进技术,这样我们才能在市场上竞争。  
**[01:21:10] Speaker A:** Conceding a marketplace based on the premise you described, I simply can't acknowledge that. It makes no sense. Because I don't think the United States is a loser.  
基于你描述的那个前提就放弃一个市场,我根本无法认同。这毫无意义。因为我不认为美国是失败者。  
**[01:21:21] Speaker A:** Our industry is not a loser. That losing proposition, that losing mindset, makes no sense to me.  
我们这个行业不是失败者。那种失败的命题,那种失败的心态,对我来说毫无意义。  
**[01:21:25] Speaker B:** Okay. I'll move on. I just want to make sure that—  
好的。我换个话题。我只是想确认——  
**[01:21:30] Speaker A:** You don't have to move on. I'm enjoying it.  
你不用换话题。我挺享受这个讨论的。  
**[01:21:37] Speaker B:** Okay, great. Then I won't. I appreciate that. But I think maybe the crux... and thanks for walking around the circles with me, because I think it helps bring out what the crux here is.  
好,太好了。那我就不换了。谢谢你的配合。不过我觉得可能问题的关键……也谢谢你陪我绕了这么多圈子,因为我觉得这有助于把问题的核心呈现出来。  
**[01:21:42] Speaker B:** The crux is you're going to extremes. Your argument starts from extremes. That if we give them any compute at all in this narrow moment, we will lose everything.  
关键在于你走向了极端。你的论证从极端开始。认为如果我们在这个关键时刻给他们任何算力,我们就会失去一切。  
**[01:21:52] Speaker A:** No, I think what my argument is—  
不,我认为我的论点是——  
**[01:21:56] Speaker A:** Those extremes, they're childish. Let me just make my argument for myself. The idea is not that there is some key threshold of compute. It's that any marginal compute is helpful. So if you have more compute, you can train a better model. And I just want you to acknowledge that any marginal sales for the American technology industry is beneficial.  
那些极端说法,很幼稚。让我自己来阐述我的论点。我的意思不是说存在某个算力的关键阈值。而是任何边际算力都是有帮助的。如果你有更多算力,你就能训练出更好的模型。我只是希望你承认,对美国科技行业来说,任何边际销售都是有益的。  
**[01:22:17] Speaker B:** I actually don't... If the AI models that run on those chips are capable of cyber offensive capabilities, or the chips are training models with cyber capabilities and running more instances of those models, it is not a nuclear weapon, but it enables a weapon of a kind.  
我其实不同意……如果运行在这些芯片上的 AI 模型具备网络攻击能力,或者这些芯片在训练具有网络攻击能力的模型并运行更多这类模型实例,它虽然不是核武器,但它使能了某种武器。  
**[01:22:31] Speaker A:** The logic that you use, you might as well say it to microprocessors and DRAMs. You might as well say it to electricity. But in fact we do have export controls on the technology that is relevant to making the most advanced DRAM. We have all kinds of export controls on China for all kinds of chip-making stuff. We sell a lot of DRAM and CPUs into China, and I think it's right.  
按照你的逻辑,你也可以对微处理器和 DRAM 这么说。你也可以对电力这么说。但事实上,我们确实对制造最先进 DRAM 的相关技术实施了出口管制。我们对中国的各种芯片制造技术都有各种出口管制。我们向中国销售大量 DRAM 和 CPU,我认为这是对的。  
**[01:22:50] Speaker B:** I guess this goes back to the fundamental question of, is AI different? If you have the kind of technology where they can find these zero-days in software, is that something where we want to minimize China's ability to get there first, to deploy it widely? We want the United States to be ahead. We can control that. How do we control that if the chips are already there and they're using them to train that model?  
我想这又回到了一个根本问题:AI 是不是不同?如果你拥有那种能在软件中找到零日漏洞的技术,我们是否应该尽量减少中国率先获得它、广泛部署它的能力?我们希望美国领先。我们可以控制这一点。但如果芯片已经在那里,他们正在用它训练那个模型,我们怎么控制?  
**[01:23:11] Speaker A:** We have tons of compute. We have tons of AI researchers. We're racing as fast as we can. Again, we have more nuclear weapons than anybody else, but we don't want to send enriched uranium anywhere.  
我们有大量算力。我们有大量 AI 研究人员。我们正在全速竞赛。再说一次,我们拥有的核武器比任何人都多,但我们不想把浓缩铀送到任何地方。  
**[01:23:20] Speaker B:** We're not enriched uranium. It's a chip, and it's a chip that they can make themselves.  
我们不是浓缩铀。这是芯片,而且是他们自己也能制造的芯片。  
**[01:23:28] Speaker A:** But there's a reason they're buying it from you. We have quotes from the founders of Chinese companies that say that they're bottlenecked on compute. Because our chips are better. On balance, our chips are better.  
但他们从你这里购买是有原因的。我们有中国公司创始人的引述,说他们在算力上遇到了瓶颈。因为我们的芯片更好。总体而言,我们的芯片更好。  
**[01:23:36] Speaker A:** There's just no question about it. In the absence of our chip... Can you acknowledge that Huawei had a record year?  
这是毫无疑问的。如果没有我们的芯片……你能承认 Huawei 创下了创纪录的一年吗?  
**[01:23:40] Speaker A:** Can you acknowledge that a whole bunch of chip companies have gone public? Can you acknowledge that?  
你能承认一大批芯片公司已经上市了吗?你能承认这一点吗?  
**[01:23:46] Speaker B:** Yes.  
是的。  
**[01:23:50] Speaker A:** Can you also acknowledge that we used to have a very large share in that market, and we no longer have a large share in that market?  
你能不能也承认,我们曾经在那个市场占有很大份额,而现在我们在那个市场的份额已经不大了?  
**[01:23:54] Speaker B:** We can also acknowledge that China is about 40% of the world's technology industry.  
我们也可以承认,中国占全球科技产业的约 40%。  
**[01:24:01] Speaker B:** To concede that market for the United States technology industry is a disservice to our country. It is a disservice to our national security.  
让美国科技产业放弃那个市场,对我们国家是一种伤害。这对我们的国家安全是一种伤害。  
**[01:24:10] Speaker B:** It is a disservice to our technology leadership, all for the benefit of one company.  
这对我们的技术领导地位是一种伤害,而这一切只是为了一家公司的利益。  
**[01:24:16] Speaker B:** It makes no sense to me.  
这对我来说毫无意义。  
**[01:24:18] Speaker A:** I guess I'm confused. It feels like you're making two different statements.  
我有点困惑。感觉你在做两种不同的陈述。  
**[01:24:21] Speaker A:** One is that we're going to win this competition with Huawei because our chips are going to be way better if we're allowed to compete.  
一种是,如果允许我们竞争,我们会在与 Huawei 的竞争中获胜,因为我们的芯片会好得多。  
**[01:24:24] Speaker A:** Another is that they would be doing the same exact thing without us anyway.  
另一种是,即使没有我们,他们也会做完全相同的事情。  
**[01:24:28] Speaker A:** How can both of those things be true at the same time?  
这两件事怎么可能同时为真?  
**[01:24:30] Speaker B:** It's obviously true. In the absence of a better choice, you'll take the only choice you have. How is that illogical? It's so logical.  
这显然是真的。在没有更好选择的情况下,你会选择你唯一拥有的选择。这怎么不合逻辑?这太合逻辑了。  
**[01:24:35] Speaker B:** The reason they want Nvidia chips is that they're better.  
他们想要 Nvidia 芯片的原因是它们更好。  
**[01:24:40] Speaker A:** Yeah.  
是的。  
**[01:24:41] Speaker B:** Better is more compute. More compute means you can train a better model.  
更好意味着更多算力。更多算力意味着你可以训练更好的模型。  
**[01:24:43] Speaker B:** No, it's just better. It's better because it's easier to program. We have a better ecosystem. But whatever the better is, whatever the better is... And of course we're going to send them compute.  
不,就是更好。它更好是因为更容易编程。我们有更好的生态系统。但无论更好是什么,无论更好是什么……当然我们会给他们提供算力。  
**[01:24:52] Speaker B:** So what? The fact of the matter is that we get to benefit.  
那又怎样?事实是我们会从中受益。  
**[01:24:59] Speaker B:** Don't forget, we get the benefit of American technology leadership.  
别忘了,我们会从美国的技术领导地位中受益。  
**[01:25:03] Speaker B:** We get the benefit of developers working on the American tech stack.  
我们会从开发者在美国技术栈上工作中受益。  
**[01:25:07] Speaker B:** We get the benefit, as those AI models diffuse out into the rest of the world, that the American tech stack is therefore the best for it. We can continue to advance and...  
我们会受益,因为当这些 AI 模型扩散到世界其他地方时,美国技术栈因此成为最佳选择。我们可以继续推进并……  
**[01:25:16] Speaker A:** diffuse American technology. That, I believe, is a positive. It's a very important part of American technology leadership.  
扩散美国技术。我认为这是积极的。这是美国技术领导地位非常重要的一部分。  
**[01:25:25] Speaker B:** Now, the policies that you're advocating resulted in the American telecommunications industry being policied out of basically the world, to the point where we don't control our own telecommunications anymore. I don't see that as smart.  
现在,你所倡导的政策导致美国电信行业基本上被政策赶出了全球市场,以至于我们不再控制自己的电信了。我不认为这是明智的。  
**[01:25:40] Speaker B:** It's a little narrow-minded, and it led to unintended consequences that I'm describing to you right now that you seem to have a very hard time understanding.  
这有点目光短浅,而且导致了我现在向你描述的意外后果,你似乎很难理解。  
**[01:25:48] Speaker A:** Okay, let's just step back. It seems like the crux here is there's a potential benefit and there's a potential cost. What we're trying to figure out is, is the benefit worth the cost? I guess I'm trying to get you to acknowledge the potential cost.  
好,我们先退一步。问题的核心在于,这件事既有潜在收益,也有潜在代价。我们要搞清楚的是:收益值不值得付出这个代价?我想让你承认一下这个潜在代价的存在。  
**[01:25:59] Speaker A:** Compute is an input to training powerful models. Powerful models do have powerful offensive capabilities, like cyber attacks.  
算力是训练强大模型的投入要素。强大的模型确实具备强大的攻击能力,比如网络攻击。  
**[01:26:09] Speaker A:** It is a good thing that American companies got to Mythos-level capabilities first, and then now they're going to hold off on those capabilities so that the American companies and American government can make their software more protected before that level of capability was announced.  
美国公司率先达到 Mythos 级别的能力,这是件好事。现在他们会暂缓释放这些能力,好让美国公司和美国政府在这种能力公开之前,把自己的软件防护做得更好。  
**[01:26:22] Speaker A:** If China had had more compute or more cloud compute, if they could have made a Mythos-level model earlier and deployed it widely, that would have been very bad.  
如果中国拥有更多算力或更多云计算资源,如果他们能更早做出 Mythos 级别的模型并广泛部署,那会非常糟糕。  
**[01:26:31] Speaker A:** One of the reasons that hasn't happened is that we have more compute thanks to companies like Nvidia in America. That is a cost of sending it to China.  
这种情况没有发生的原因之一,就是我们有更多算力,这要归功于美国的 Nvidia 这样的公司。把算力送到中国,这就是代价。  
**[01:26:40] Speaker A:** So let's leave the benefit aside for a second. Do you acknowledge that this is a potential cost?  
所以我们先把收益放一边。你承认这是一个潜在代价吗?  
**[01:26:45] Speaker B:** I'll also tell you the potential cost is we allow one of the most important layers of the AI stack, the chip layer, to concede an entire market—the second largest market in the world—so that they could develop scale, so that they could develop their own ecosystem, so that future AI models are optimized in a very different way than the American tech stack.  
我也告诉你潜在代价是什么:我们让 AI 技术栈中最重要的一层——芯片层——拱手让出一个完整的市场,世界第二大市场,让他们发展规模,让他们建立自己的生态系统,让未来的 AI 模型以一种与美国技术栈非常不同的方式进行优化。  
**[01:27:12] Speaker B:** As AI diffuses out into the rest of the world, their standards, their tech stack, will become  
随着 AI 扩散到世界其他地方,他们的标准、他们的技术栈,会变得  
**[01:27:21] Speaker A:** Superior to ours, because their models are open. I guess I just believe enough in Nvidia's kernel engineers and CUDA engineers to think that they could optimize—  
比我们的更优越,因为他们的模型是开源的。我只是足够相信 Nvidia 的内核工程师和 CUDA 工程师,相信他们能够优化——  
**[01:27:29] Speaker B:** AI is more than kernel optimization, as you know.  
当然,但你可以做很多事情,比如把模型蒸馏成非常适配你芯片的版本。  
**[01:27:29] Speaker A:** Of course, but there are so many things you can do, from distilling to a model that's well-fit for your chips.  
当然,但你可以做很多事情,比如把模型蒸馏成非常适配你芯片的版本。  
**[01:27:36] Speaker A:** We're going to do our best.  
我们会尽力而为。  
**[01:27:38] Speaker B:** You have all the software. It's just hard to imagine that there's a long-term lock-in to the Chinese ecosystem, even if they have a slightly better open source model for a while.  
你拥有所有的软件。很难想象会长期锁定在中国的生态系统里,即使他们在一段时间内有一个稍微好一点的开源模型。  
**[01:27:42] Speaker A:** China is the largest contributor to open source software in the world. Fact. China's the largest contributor to open models in the world. Fact. Today it's built on the American tech stack, Nvidia's. Fact. All five layers of the tech stack for AI are important.  
中国是世界上最大的开源软件贡献者。事实。中国是世界上最大的开源模型贡献者。事实。今天这些都建立在美国的技术栈上,Nvidia 的技术栈。事实。AI 技术栈的五层都很重要。  
**[01:28:03] Speaker A:** The United States ought to go win all five of them. They're all important. The one that is the most important, of course, is the AI application layer.  
美国应该在这五层上全部获胜。它们都很重要。当然,最重要的是 AI 应用层。  
**[01:28:14] Speaker A:** The layer that diffuses into society, the one that uses it most will benefit from this industrial revolution most.  
这一层会扩散到社会中,谁用得最多,谁就会从这场工业革命中获益最多。  
**[01:28:27] Speaker A:** But my point is that every layer has to succeed. If we scare this country into thinking that AI is somehow a nuclear bomb, so that everybody hates AI and everybody's afraid of AI, I don't know how you're helping the United States. You're doing it a disservice.  
但我的观点是,每一层都必须成功。如果我们吓唬这个国家,让大家觉得 AI 就像核弹一样,让所有人都讨厌 AI、害怕 AI,我不知道你这是在帮美国什么忙。你这是在帮倒忙。  
**[01:28:45] Speaker A:** If we scare everybody out of doing software engineering jobs because it's going to kill every software engineering job—and we don't have any software engineers as a result of that—we're doing a disservice to the United States.  
如果我们吓得大家都不敢做软件工程的工作,因为它会消灭所有软件工程师的岗位——结果我们就没有软件工程师了——我们这是在给美国帮倒忙。  
**[01:28:59] Speaker A:** If we scare everybody out of radiology so nobody wants to be a radiologist because computer vision is completely free and no AI is going to do a worse job than a radiologist, we misunderstand the difference between a job and a task.  
如果我们吓得大家都不敢学放射科,因为计算机视觉完全免费,而且 AI 做得不会比放射科医生差,那我们就是混淆了工作和任务的区别。  
**[01:29:14] Speaker A:** The job of a radiologist is patient care. The task is to read a scan.  
放射科医生的工作是患者护理。任务才是读片。  
**[01:29:19] Speaker A:** If we misunderstand that so profoundly and we scare everybody out of going to radiology school, we're not going to have enough radiologists and good enough healthcare.  
如果我们把这个理解得如此错误,吓得大家都不去读放射科,我们就不会有足够的放射科医生,也不会有足够好的医疗服务。  
**[01:29:29] Speaker A:** So I'm making the case that when you make a premise that is so extreme, everything goes—  
所以我要说的是,当你提出一个如此极端的前提,一切就会——  
**[01:29:40] Speaker A:** From zero or infinity, we end up scaring people in a way that's just not true. Life is not like that.  
从零到无穷,我们最终会以一种不真实的方式吓到人们。生活不是那样的。  
**[01:29:48] Speaker A:** Do we want the United States to be first? Of course we do. Do we need to be a leader in every layer of that stack? Of course we do.  
我们想让美国成为第一吗?当然想。我们需要在技术栈的每一层都领先吗?当然需要。  
**[01:29:59] Speaker A:** Of course we do. Today you're talking about Mythos because Mythos is important. Sure. That's fantastic.  
当然需要。今天你在谈论 Mythos,因为 Mythos 很重要。没错。这很好。  
**[01:30:07] Speaker A:** But in a few years' time, I'm making you the prediction that when we want the American tech stack, when we want American technology to be diffused around the world—out to India, out to the Middle East, out to Africa, out to Southeast Asia—when our country would like to export, because we would like to export our technology, we would like to export our standards, on that day, I want you and I to have that same conversation again.  
但我预测,几年后,当我们希望美国的技术栈、美国的技术能够扩散到全世界——印度、中东、非洲、东南亚——当我们国家想要出口技术、出口标准的时候,那一天,我希望你我能再进行一次同样的对话。  
**[01:30:38] Speaker A:** I will tell you exactly about today's conversation, about how your policy and what you imagined literally caused the United States to concede the second largest market in the world for no good reason at all.  
到时候我会明确告诉你,今天的这场对话,你的政策和你的设想是如何导致美国毫无理由地放弃了世界第二大市场。  
**[01:30:48] Speaker A:** We shouldn't concede it. If we lose it, we lose it. But why do we concede it?  
我们不应该主动放弃。如果我们输了,那就输了。但为什么要主动放弃呢?  
**[01:30:55] Speaker A:** Now nobody is advocating an all or nothing.  
现在没有人主张全有或全无的极端做法。  
**[01:31:02] Speaker A:** Nobody's advocating all or nothing, meaning we ship everything to China at all times. Nobody's advocating that.  
没有人主张全有或全无,也就是说随时把所有东西都运往中国。没人主张这样做。  
**[01:31:07] Speaker A:** We should always have the best technology here.  
我们应该始终把最好的技术留在国内。  
**[01:31:12] Speaker A:** We should always have the most technology here, and the first.  
我们应该始终拥有最多的技术,而且是最先拥有。  
**[01:31:16] Speaker A:** But we should also try to compete and win around the world.  
但我们也应该努力在全球范围内竞争并获胜。  
**[01:31:22] Speaker A:** Both of those things can simultaneously happen. It requires some amount of nuance, some amount of maturity instead of absolutes.  
这两件事可以同时发生。这需要一定的细致考量和成熟度,而不是绝对化的思维。  
**[01:31:29] Speaker A:** The world is just not absolutes.  
世界本来就不是非黑即白的。  
**[01:31:34] Speaker B:** Okay. The argument hinges on this. They've built models that are specified for the best chips that they make in a few years.  
好的。争论的关键在于:他们已经构建了专门针对未来几年最好芯片的模型。  
**[01:31:41] Speaker B:** Those chips get exported around the world. That sets the standard.  
这些芯片出口到世界各地,从而设定了标准。  
**[01:31:44] Speaker B:** Because of EUV export controls, as we said, you're going to move on to 1.6nm. They're still going to be on 7nm, even after a few years from now.  
由于 EUV 出口管制,正如我们所说,你们会发展到 1.6nm,而他们即使在几年后仍然停留在 7nm。  
**[01:31:54] Speaker B:** It may make sense that domestically they would prefer, "Hey, we've got so much energy, we can manufacture at scale. We'll still keep using 7nm." But on the exporting thing, their 7nm chips have to be competitive against your 1.6nm chips.  
在国内市场,他们可能会觉得「我们有充足的能源,可以大规模制造,继续使用 7nm 就行」。但在出口方面,他们的 7nm 芯片必须与你们的 1.6nm 芯片竞争。  
**[01:32:04] Speaker B:** Their models have to be so far optimized for  
他们的模型必须针对  
**[01:32:10] Speaker A:** The 7nm—that it's better to run their models on 7nm than to run their models on your 1.6nm.  
7nm 进行深度优化,优化到在 7nm 上运行他们的模型比在你们的 1.6nm 上运行还要好。  
**[01:32:17] Speaker B:** Can we just look at the facts then? Is Blackwell 50 times more advanced lithography than Hopper? Is it 50 times? Not even close.  
那我们能看看事实吗?Blackwell 的先进光刻技术是 Hopper 的 50 倍吗?是 50 倍吗?完全不是。  
**[01:32:24] Speaker A:** I just kept saying it over and over again. Moore's Law is dead. Between Hopper and Blackwell, from the transistors themselves, call it 75%.  
我一直在反复强调,摩尔定律已死。从 Hopper 到 Blackwell,就晶体管本身而言,大约提升了 75%。  
**[01:32:40] Speaker A:** It was three years apart, 75%. Blackwell is 50 times Hopper. My point is, architecture matters.  
它们相隔三年,提升 75%。但 Blackwell 的性能是 Hopper 的 50 倍。我的观点是,架构很重要。  
**[01:32:54] Speaker A:** Computer science matters. Semiconductor physics matters as well, but computer science matters.  
计算机科学很重要。半导体物理当然也很重要,但计算机科学同样重要。  
**[01:33:01] Speaker A:** The impact of AI largely comes from the computing stack,  
AI 的影响力很大程度上来自计算栈,  
**[01:33:06] Speaker A:** which is the reason why CUDA is so effective, which is the reason why CUDA is so beloved.  
这就是为什么 CUDA 如此高效,为什么 CUDA 如此受欢迎。  
**[01:33:12] Speaker A:** It's an ecosystem, a computing architecture that allows for so much flexibility that if you wanted to change an architecture completely—create something like MoE, create something like diffusion, create something that's disaggregated—you could do so. It's easy to do.  
它是一个生态系统,一个计算架构,提供了极大的灵活性,如果你想完全改变架构——创建像 MoE 这样的东西,创建像扩散模型这样的东西,创建分布式的架构——你都可以做到,而且很容易实现。  
**[01:33:29] Speaker A:** So the fact of the matter is, AI is about the stack above as much as it is about the architecture below. To the extent that we have architectures and software stacks that are optimized for our stack, for our ecosystem, it is obviously good,  
所以事实是,AI 既关乎上层的软件栈,也关乎底层的架构。如果我们的架构和软件栈是为我们自己的技术栈、我们的生态系统优化的,这显然是好事,  
**[01:33:46] Speaker A:** because we started the conversation today about how Nvidia's ecosystem is so rich.  
因为我们今天一开始就在讨论 Nvidia 的生态系统有多么丰富。  
**[01:33:51] Speaker A:** Why do people always love programming CUDA first? They do. They do. So do the researchers in China.  
为什么大家总是喜欢先用 CUDA 编程?确实是这样。中国的研究人员也是如此。  
**[01:33:58] Speaker A:** But if we are forced to leave China, if we're forced to leave China, first of all,  
但如果我们被迫离开中国市场,首先要说的是,  
**[01:34:06] Speaker A:** it's a policy mistake. Obviously it has backlash. It has turned out badly for the United States.  
这是一个政策失误。显然它产生了反作用,对美国来说结果很糟糕。  
**[01:34:19] Speaker A:** It enabled, it accelerated their chip industry. It forced all of their AI ecosystem to focus on their internal architectures. It's not too late, but  
它促成了、加速了中国芯片产业的发展。它迫使中国整个 AI 生态系统专注于自己的内部架构。现在还不算太晚,但  
**[01:34:29] Speaker A:** nonetheless it has already happened. You're going to see in the future,  
这一切已经发生了。未来你会看到,  
**[01:34:35] Speaker A:** they're not stuck at 7nm, obviously. They're good at manufacturing. They will continue to advance from 7nm and beyond. Now, is there a 10x difference between 5nm and  
他们显然不会停留在 7nm 制程。他们擅长制造,会继续从 7nm 往更先进的制程发展。那么 5nm 和  
**[01:34:50] Speaker A:** 7nm? The answer is no. Architecture matters. Networking matters. That's why Nvidia bought Mellanox. Networking matters. Energy matters. So all of that stuff matters.  
7nm 之间有 10 倍的差距吗?答案是没有。架构很重要。网络很重要,这就是 Nvidia 收购 Mellanox 的原因。网络很重要。能耗也很重要。所有这些因素都很重要。  
**[01:35:02] Speaker A:** It's not simplistic, like the way you're trying to distill it.  
这不像你试图简化的那样简单。  
**[01:35:06] Speaker B:** We can move on from China, but that actually raises an interesting question. We were discussing earlier these bottlenecks at TSMC and memory and so forth. So if we're in this world where you're already the majority of N3—and at some point you'll be N2 and you'll be a majority of that—do you see that you could go back to N7, the spare capacity at an older process node, and say, "Hey, the demand for AI is so great and our capacity to expand the leading edge is not meeting it, so we're going to make a Hopper or Ampere, but with everything we know about numerics today and all the other improvements you described"? Do you see that world happening before 2030?  
我们可以不谈中国了,但这引出了一个有趣的问题。我们之前讨论过 TSMC 和内存方面的瓶颈。如果在这个世界里,你们已经占据了 N3 制程的大部分产能——未来某个时候你们会用 N2 并占据其大部分产能——你觉得有可能回到 N7 制程,利用旧制程节点的闲置产能,然后说「嘿,AI 的需求太大了,我们扩展尖端制程的能力无法满足需求,所以我们要用今天对数值计算的所有认知以及你描述的所有其他改进,来制造一个 Hopper 或 Ampere」?你认为 2030 年之前会出现这种情况吗?  
**[01:35:41] Speaker A:** It's not necessary to. The reason for that is because with every generation, the architecture is more than just the transistor scale.  
没必要这样做。原因是每一代产品,架构不仅仅是晶体管规模的问题。  
**[01:36:03] Speaker A:** You're doing so much engineering and packaging and stacking, and the numerics and the system architecture. When you run out of capacity, to easily go back to another node… That's a level of R&D that no one could afford.  
你要做大量的工程工作,包括封装、堆叠,还有数值计算和系统架构。当产能不足时,要轻易回到另一个制程节点……那需要的研发投入是没人负担得起的。  
**[01:36:23] Speaker A:** We could afford to lean forward. I don't think we could afford to go back.  
我们有能力向前推进。我认为我们负担不起倒退。  
**[01:36:26] Speaker A:** Now, if the world simply says… If on that day, let's do the thought experiment, on that day we go, "Listen, we're just never going to have more capacity ever again." Would I go back and use 7nm? In a heartbeat, of course I would.  
现在,如果世界就是这么说……做个思想实验,如果有一天我们被告知「听着,我们永远不会再有更多产能了」。我会回去用 7nm 吗?当然会,毫不犹豫。  
**[01:36:42] Speaker B:** One question somebody I was talking to had is, why doesn't Nvidia run multiple different chip projects at the same time with totally different architecture? So you could do something like a Cerebras-style wafer scale. You could do a Dojo-style huge package. You could do one without CUDA. You have the resources and the engineering talent to do all of these in parallel. So why put all the eggs in one basket, given who knows where AI might go and architectures might go?  
有人问我一个问题,为什么 Nvidia 不同时推进多个完全不同架构的芯片项目?比如你们可以做一个 Cerebras 那种晶圆级规模的,可以做一个 Dojo 那种超大封装的,还可以做一个不带 CUDA 的。你们有资源也有工程人才可以并行做所有这些。那为什么要把所有鸡蛋放在一个篮子里,毕竟谁知道 AI 和架构会往哪个方向发展呢?  
**[01:37:04] Speaker A:** Oh, we could. It's just that we don't have a better idea. We could do all of those things.  
哦,我们可以做到。只是我们没有更好的想法。我们可以做所有那些事情。  
**[01:37:14] Speaker A:** It's just not better. We simulate it all in our simulator, provably worse. So we wouldn't do it.  
只是那些方案并不更好。我们在模拟器里模拟过所有这些,可证明地更差。所以我们不会那样做。  
**[01:37:22] Speaker A:** We're working on exactly the projects that we want to work on.  
我们正在做的正是我们想做的项目。  
**[01:37:32] Speaker A:** If the workload were to change dramatically—and I don't mean the algorithms, I actually mean the workload, and that depends on the shape of the market—we may decide to add other accelerators.  
如果工作负载发生剧烈变化——我指的不是算法,而是实际的工作负载,这取决于市场的形态——我们可能会决定增加其他加速器。  
**[01:37:49] Speaker A:** For example, recently we added Groq, and we're going to fold Groq into our CUDA ecosystem.  
例如,我们最近增加了 Groq,我们将把 Groq 整合到我们的 CUDA 生态系统中。  
**[01:38:00] Speaker A:** We're doing that now because the value of tokens has gone up so high that you could have different pricing of tokens.  
我们现在这样做是因为 token 的价值已经变得非常高,你可以对 token 采用不同的定价。  
**[01:38:07] Speaker A:** Back in the old days, just a couple years ago, tokens were either free or barely expensive.  
在过去,就在几年前,token 要么是免费的,要么几乎不值钱。  
**[01:38:11] Speaker A:** But now you can have different customers, and those customers want different answers.  
但现在你可以有不同的客户,这些客户想要不同的答案。  
**[01:38:17] Speaker A:** Because the customers make so much money—for example, our software engineers—if I can give them much more responsive tokens so that they're even more productive than they are today, I would pay for it.  
因为客户赚很多钱——比如我们的软件工程师——如果我能给他们响应更快的 token,让他们比现在更高效,我愿意为此付费。  
**[01:38:35] Speaker A:** But that market has only recently emerged.  
但这个市场是最近才出现的。  
**[01:38:41] Speaker A:** So I think we now have the ability to have the same model, based on the response time, have different segments.  
所以我认为现在我们有能力针对同一个模型,根据响应时间,划分不同的细分市场。  
**[01:38:46] Speaker A:** That's the reason why we decided to expand the Pareto frontier and create a segment of inference that is faster response time, even though it's lower throughput.  
这就是我们决定拓展帕累托前沿的原因,创造一个响应时间更快的推理细分市场,尽管吞吐量会更低。  
**[01:39:00] Speaker A:** Until now, higher throughput is always better.  
在此之前,更高的吞吐量一直被认为是更好的。  
**[01:39:05] Speaker A:** We think there could be a world where there could be very high ASP tokens, and even though the throughput is lower in the factory, the ASPs make up for it.  
我们认为可能会出现这样一个世界:token 的平均售价(ASP)非常高,即使工厂的吞吐量较低,高 ASP 也能弥补这一点。  
**[01:39:13] Speaker A:** That's the reason why we did it.  
这就是我们这么做的原因。  
**[01:39:17] Speaker A:** But otherwise, from an architecture perspective, if I had more money, I would put more behind Nvidia's architecture.  
但除此之外,从架构角度来看,如果我有更多资金,我会在 Nvidia 的架构上投入更多。  
**[01:39:21] Speaker B:** I think this idea of extremely premium tokens and just the disaggregation of the inference market is very interesting.  
我觉得这种极高价值 token 的概念,以及推理市场的细分化,都非常有意思。  
**[01:39:29] Speaker A:** The segmentation of it.  
对,市场的细分。  
**[01:39:34] Speaker B:** Yeah. Alright, final question. Suppose the deep learning revolution didn't happen. What would Nvidia be doing? Obviously games, but given—  
好的。最后一个问题。假设深度学习革命没有发生,Nvidia 会在做什么?显然会做游戏,但考虑到——  
**[01:39:48] Speaker A:** Accelerated computing, the same thing we've been doing all along.  
加速计算,和我们一直以来做的事情一样。  
**[01:39:55] Speaker A:** The premise of our company is that Moore's law is going to… General purpose computing  
我们公司的前提是,摩尔定律会……通用计算  
**[01:40:00] Speaker A:** It is good for a lot of things, but for a lot of computation it's not ideal. So we combined an architecture called a GPU, CUDA, to a CPU, so that we can accelerate the workload of the CPU.  
在很多方面都很好用,但对于大量计算任务来说并不理想。所以我们把一种叫 GPU 的架构,结合 CUDA,与 CPU 结合起来,这样就能加速 CPU 的工作负载。  
**[01:40:12] Speaker A:** Different kernels of code or algorithms could be offloaded onto our GPU. As a result, you speed up an application by 100x, 200x.  
不同的代码内核或算法可以卸载到我们的 GPU 上。结果就是,你能把应用程序的速度提升 100 倍、200 倍。  
**[01:40:24] Speaker A:** Where can you use that? Obviously engineering and science and physics, data processing, computer graphics, image generation, all kinds of things.  
这能用在哪里?显然是工程、科学、物理、数据处理、计算机图形学、图像生成,各种各样的领域。  
**[01:40:37] Speaker A:** Even if AI doesn't exist today, Nvidia would be very, very large.  
即使今天 AI 不存在,Nvidia 也会是一家非常非常大的公司。  
**[01:40:42] Speaker A:** The reason for that is fairly fundamental, which is that the ability for general purpose computing to continue to scale has largely run its course.  
原因很根本,那就是通用计算持续扩展的能力基本上已经走到尽头了。  
**[01:40:48] Speaker A:** And the only way... Not the only way, but the way to do that is through domain-specific acceleration.  
而唯一的方法……不是唯一的方法,但实现这一点的方式是通过特定领域的加速。  
**[01:40:55] Speaker A:** One of the domains that we started with was computer graphics, but there are many other domains. There's all kinds.  
我们最初涉足的领域之一是计算机图形学,但还有很多其他领域,各种各样的。  
**[01:41:09] Speaker A:** Particle physics and fluids, structured data processing, all kinds of different types of algorithms that benefit from CUDA.  
粒子物理、流体力学、结构化数据处理,各种不同类型的算法都能从 CUDA 中受益。  
**[01:41:14] Speaker A:** Our mission was really to bring accelerated computing to the world and advance the type of applications that general purpose computing can't do, and scale to the level of capability that helps break through certain fields of science.  
我们的使命实际上是把加速计算带给全世界,推进那些通用计算做不到的应用类型,并扩展到能够在某些科学领域实现突破的能力水平。  
**[01:41:36] Speaker A:** Some of the early applications were molecular dynamics, seismic processing for energy discovery, image processing of course, all of those kinds of fields where general purpose computing is just simply too inefficient to do so.  
一些早期应用包括分子动力学、用于能源勘探的地震处理、当然还有图像处理,所有这些领域通用计算都太低效了,根本做不到。  
**[01:41:50] Speaker A:** If there were no AI, I would be very sad.  
如果没有 AI,我会非常难过。  
**[01:41:57] Speaker A:** But because of the advances that we made in computing, we democratized deep learning.  
但正是因为我们在计算方面取得的进步,我们让深度学习变得普及了。  
**[01:42:06] Speaker A:** We made it possible for any researcher, any scientist, anywhere, any student, to be able to access a PC or a GeForce add-in card and do amazing science.  
我们让任何研究人员、任何科学家、任何地方的任何学生,都能够使用一台 PC 或一块 GeForce 显卡来做出色的科学研究。  
**[01:42:21] Speaker A:** That fundamental promise hasn't changed, not even a little bit.  
这个根本承诺从未改变,一点都没有。  
**[01:42:27] Speaker A:** If you watch GTC, there's the whole beginning part of it. None of it's AI. That whole part of it with computational lithography or our quantum chemistry work, data processing work, all of that stuff is unrelated to AI.  
如果你看 GTC 大会,开头整个部分都不是 AI。那整个部分讲的是计算光刻、我们的量子化学工作、数据处理工作,所有这些都和 AI 无关。  
**[01:42:45] Speaker A:** And it's still very important. I know that AI is very interesting and quite  
这一点仍然非常重要。我知道 AI 很有趣,也确实  
**[01:42:53] Speaker A:** Exciting, but there's a lot of people doing a lot of very important work that's not AI related, and tensors are not the only way that you compute it. We want to help everybody.  
令人兴奋,但有很多人在做大量非常重要的工作,这些工作与 AI 无关,而且张量也不是唯一的计算方式。我们希望帮助所有人。  
**[01:43:06] Speaker A:** Jensen, thank you so much.  
Jensen,非常感谢你。  
**[01:43:09] Speaker B:** You're welcome. I enjoyed it.  
我也是。  
**[01:43:09] Speaker A:** Me too.  
我也是。  

---

## Deep Dive Summary

### Topic 1: Will AI commoditize Nvidia's software advantage?
AI会让Nvidia的软件优势商品化吗？
_[00:00]_

**Q:** If software gets commoditized by AI, does Nvidia—which fundamentally makes software that others manufacture—also get commoditized?
**问：** 如果AI让软件商品化，那么本质上是让别人代工软件的Nvidia会不会也被商品化？

**A:** Speaker B reframes Nvidia's value proposition as the transformation of "electrons to tokens," arguing this process involves deep artistry and science that resists commoditization. While acknowledging Nvidia's manufacturing model—sending designs to TSMC, who coordinates with memory suppliers and ODMs—B emphasizes that making tokens "more valuable than another" requires ongoing invention whose journey is "far from over." The company's strategy is to do "as much as necessary and as little as possible," partnering extensively to enable the core transformation at "incredible capabilities." B explicitly adopts the questioner's framing as his "mental model" of Nvidia: electrons in, tokens out, with Nvidia orchestrating the middle layer through selective vertical integration and ecosystem partnerships. Despite efficiency improvements ahead, B doubts full commoditization because the underlying science remains incompletely understood and the engineering complexity of optimizing this transformation continues to deepen.
**答：** Speaker B把Nvidia的价值重新定义为"electrons to tokens"的转换过程，认为这个过程包含的工艺和科学深度很难被商品化。虽然承认Nvidia的制造模式——把设计发给TSMC，由TSMC协调内存供应商和ODM——但B强调让一个token"比另一个更有价值"需要持续的发明创造，这个旅程"远未结束"。公司的策略是做"必要的事，尽可能少做"，通过广泛的合作伙伴关系来实现核心转换的"incredible capabilities"。B明确表示认同提问者的框架作为他对Nvidia的"mental model"：输入electrons，输出tokens，Nvidia通过选择性的垂直整合和生态系统合作来协调中间层。尽管未来会有效率提升，但B怀疑会完全商品化，因为底层科学仍未被充分理解，优化这个转换过程的工程复杂度还在不断加深。

### Topic 2: Nvidia's ecosystem strategy and the five-layer AI cake
Nvidia的生态系统策略与AI五层架构
_[02:05]_

**Q:** How does Nvidia approach partnerships and what is their philosophy of doing 'as little as possible' while enabling the electron-to-token transformation?
**问：** Nvidia如何处理合作伙伴关系，以及他们在实现电子到token转换时"尽可能少做"的理念是什么？

**A:** Nvidia's strategy centers on building the largest possible ecosystem across what they call a "five-layer cake" of AI infrastructure, deliberately doing "as little as possible" by partnering extensively with supply chain, computer companies, application developers, and model makers. The parts Nvidia does handle internally are "insanely hard" and unlikely to be commoditized, allowing them to focus on core competencies while enabling partners across the stack. Contrary to conventional wisdom about AI commoditizing software tools, the speaker argues that "the number of agents is going to grow exponentially" alongside tool users, causing software companies to "skyrocket" as agents multiply the instances of tools like Excel, PowerPoint, Synopsys Design Compiler, and Cadence tools. The current limitation is that "agents aren't good enough at using their tools yet," but once they improve—either through companies building specialized agents or general agents becoming more capable—engineers will be "supported by a bunch of agents" exploring design spaces at unprecedented scale. This vision suggests a multiplicative effect where AI doesn't replace existing software infrastructure but dramatically increases its utilization through agent-mediated workflows.
**答：** Nvidia的策略核心是在AI基础设施的"五层架构"中构建最大规模的生态系统，通过与供应链、计算机公司、应用开发者和模型制造商广泛合作来"尽可能少做"。Nvidia自己承担的部分"极其困难"且不太可能被商品化，这让他们能专注于核心能力同时赋能整个技术栈的合作伙伴。与AI会让软件工具商品化的传统观点相反，演讲者认为"agent的数量会呈指数级增长"，工具使用者也会同步增长，这会让软件公司"飞速发展"，因为agent会成倍增加Excel、PowerPoint、Synopsys Design Compiler和Cadence等工具的使用实例。当前的限制是"agent还不够擅长使用工具"，但一旦它们改进——无论是通过公司构建专用agent还是通用agent变得更强——工程师将"得到一群agent的支持"，以前所未有的规模探索设计空间。这个愿景表明AI不会取代现有软件基础设施，而是通过agent中介的工作流大幅提升其使用率，产生乘数效应。

### Topic 3: AI agents will drive exponential growth in software tool usage
AI agents将推动软件工具使用量指数级增长
_[02:46]_

**Q:** Will enterprise software and tool makers be commoditized, or will AI agents actually increase demand for existing tools?
**问：** 企业软件和工具制造商会被商品化吗，还是AI agents反而会增加对现有工具的需求？

**A:** The speaker argues that AI agents will dramatically increase, not commoditize, demand for existing enterprise software tools. He distinguishes between "tool makers" like Excel, PowerPoint, and EDA companies (Cadence, Synopsys) versus workflow codification systems, predicting that "the number of agents is going to grow exponentially, and the number of tool users is going to grow exponentially." Using chip design as a concrete example, he envisions scenarios where "the number of instances of Synopsys Design Compiler is going to skyrocket" as engineers are "supported by a bunch of agents" that explore design spaces at unprecedented scale. The current limitation is that "agents aren't good enough at using their tools yet," but he expects this bottleneck to resolve through a combination of companies building specialized agents and general agents improving their tool-use capabilities. Rather than replacing human workflows, agents will amplify them, transforming today's engineer-constrained processes into massively parallel exploration systems that consume far more tool licenses.
**答：** 嘉宾认为AI agents不会让企业软件工具商品化，反而会大幅增加需求。他区分了"工具制造商"（如Excel、PowerPoint和EDA公司Cadence、Synopsys）和工作流固化系统，预测"agents数量和工具使用者数量都会指数级增长"。以芯片设计为例，他设想工程师会"被一群agents支持"，这些agents会以前所未有的规模探索设计空间，导致"Synopsys Design Compiler的实例数量会暴增"。当前的限制是"agents还不够擅长使用工具"，但他预期这个瓶颈会通过公司自建专用agents和通用agents提升工具使用能力来共同解决。Agents不会取代人类工作流程，而是放大它们，把今天受工程师数量限制的流程转变为大规模并行探索系统，消耗更多工具许可证。

### Topic 4: Supply chain commitments as competitive moat
供应链承诺作为竞争护城河
_[04:26]_

**Q:** Are Nvidia's massive purchase commitments with foundries, memory, and packaging suppliers their real moat for the next few years?
**问：** Nvidia 与晶圆厂、内存和封装供应商的大规模采购承诺是否是未来几年的真正护城河？

**A:** Nvidia's supply chain moat stems not just from explicit purchase commitments approaching $250 billion, but from their unique ability to inspire upstream investments through what the speaker calls "informing, inspiring, and aligning with CEOs" about AI's trajectory. Suppliers are willing to make these investments specifically for Nvidia because they trust the company's "capacity to buy their supply and sell it through" its massive downstream demand, which competitors cannot replicate. The speaker emphasizes that this isn't merely about locking up components—it's about "supply chain flow" and "churn," noting that "nobody is going to build a supply chain for an architecture if the business churn is low." Events like GTC serve a strategic function beyond marketing: bringing together the "full 360 degrees" ecosystem so upstream suppliers can directly witness the downstream demand and AI innovation that justifies their capital investments. The moat is therefore circular—Nvidia's market position enables supply commitments, which in turn sustain their ability to operate "at the scale we do them" when facing potential trillion-dollar revenue years.
**答：** Nvidia 的供应链护城河不仅来自接近 2500 亿美元的采购承诺，更源于其独特能力：通过「告知、激励和协调各行业 CEO」来推动上游投资。供应商愿意专门为 Nvidia 投资，是因为他们相信 Nvidia 有能力「购买他们的供应并通过下游销售出去」，而这种大规模下游需求是竞争对手无法复制的。发言人强调这不仅是锁定组件，而是关于「供应链流动」和「周转率」——「如果业务周转率低，没人会为某个架构建立供应链」。GTC 这样的活动具有战略功能：将「360 度全方位」的生态系统聚集在一起，让上游供应商直接看到下游需求和 AI 创新，从而证明他们的资本投资是合理的。这形成了一个循环护城河——Nvidia 的市场地位促成供应承诺，而供应承诺又支撑其在面对潜在万亿美元营收规模时的运营能力。

### Topic 5: Can upstream supply keep pace with 2x annual growth?
上游供应能否跟上每年2倍的增长？
_[08:32]_

**Q:** How can Nvidia continue doubling revenue and tripling compute output when they're already the majority customer on key manufacturing nodes?
**问：** 当Nvidia已经是关键制造节点的主要客户时，他们如何继续实现收入翻倍和计算输出三倍增长？

**A:** The speaker reframes supply constraints as a healthy market signal rather than a fundamental barrier to growth. He acknowledges that "at any instant, we could be limited" by various bottlenecks, even citing plumbers as a real-world example of unexpected constraints. However, he argues that having "instantaneous demand greater than the total supply" is actually "a good condition" because it indicates strong market pull. The key mechanism for overcoming bottlenecks is industry coordination: when "one particular component is too far away, the industry swarms it," suggesting that capital and resources naturally flow to resolve the most critical constraints. This perspective implies that while growth may face temporary friction from specific supply chain nodes, the overall trajectory can continue as long as no single bottleneck becomes insurmountably far from meeting demand.
**答：** 发言者将供应限制重新定义为健康的市场信号，而非增长的根本障碍。他承认"在任何时刻我们都可能受限"于各种瓶颈，甚至举例说水管工这种意想不到的限制因素确实发生过。但他认为"瞬时需求大于总供应"实际上是"一个好的状态"，因为这表明市场拉动力强劲。克服瓶颈的关键机制是产业协同：当"某个特定组件供应缺口太大时，整个产业就会蜂拥而至"解决它，这意味着资本和资源会自然流向最关键的约束点。这个视角暗示，虽然增长可能因特定供应链节点而遇到暂时摩擦，但只要没有单一瓶颈变得无法逾越，整体增长轨迹就能持续。

### Topic 6: Swarming bottlenecks and reshaping supply chains
集中解决瓶颈并重塑供应链
_[10:05]_

**Q:** How does the industry address supply chain bottlenecks like CoWoS packaging and HBM memory, and how does Nvidia influence these transformations?
**问：** 行业如何解决CoWoS封装和HBM内存等供应链瓶颈，Nvidia如何影响这些转变？

**A:** The speaker describes a systematic approach where the industry "swarms" supply chain bottlenecks when components fall too far behind demand, citing CoWoS packaging as a success story where capacity "doubled, doubled, doubled" over two years until supply stabilized. This swarming effect transformed what were once "specialty" technologies like CoWoS and HBM memory into "mainstream computing technology," prompting TSMC to scale packaging at the same pace as logic production. Nvidia's growing influence allows them to shape supply chains proactively, with the speaker recalling how they predicted today's AI demands five years ago and secured early partnerships with companies like Micron on LPDDR and HBM memory. The strategy has evolved from reactive problem-solving to "prefetching the bottlenecks years in advance," as demonstrated by multi-year investments with Lumentum, Coherent, and the silicon photonics ecosystem that fundamentally "reshaped the supply chain."
**答：** 讲者描述了一种系统性方法：当某个组件供应严重滞后时，整个行业会"集中火力"解决瓶颈。CoWoS封装就是成功案例，两年内产能"翻倍、翻倍、再翻倍"，最终供应趋于稳定。这种集中攻坚将CoWoS和HBM内存从"特殊技术"转变为"主流计算技术"，促使TSMC按照逻辑芯片的规模同步扩展封装产能。Nvidia影响力的增强让他们能主动塑造供应链，讲者回忆五年前就准确预测了今天AI需求的爆发，并提前与Micron等公司在LPDDR和HBM内存上建立合作。策略已从被动应对演变为"提前数年预判瓶颈"，例如与Lumentum、Coherent及硅光子生态系统的多年投资从根本上"重塑了供应链"。

### Topic 7: TSMC supply chain preparation and ecosystem shaping
Nvidia 如何为 TSMC 供应链做准备
_[12:23]_

**Q:** How is Nvidia preparing the supply chain to support scale through technology invention and capacity investment?
**问：** Nvidia 如何通过技术发明和产能投资来准备供应链以支持规模化？

**A:** Nvidia is taking a multi-layered approach to prepare the TSMC-centered supply chain for massive scale, combining technology development with strategic ecosystem investments. The company has partnered with TSMC on "COUPE" and invented technologies that they deliberately license to the broader supply chain to "keep it nice and open," ensuring no single bottleneck emerges. Beyond IP licensing, Nvidia is actively inventing "new workflows" and specialized equipment like "double-sided probing" for testing, which represents fundamental innovation in manufacturing processes. The strategy extends to direct financial involvement, where Nvidia invests in supply chain companies and helps them "scale up their capacity," essentially de-risking their expansion to meet Nvidia's future demand. This comprehensive approach—spanning patents, process innovation, equipment development, and capacity financing—shows Nvidia is "trying to shape the ecosystem" proactively rather than simply placing orders and hoping suppliers can deliver.
**答：** Nvidia 正在采取多层次策略来为以 TSMC 为中心的供应链做好大规模扩张准备，将技术开发与生态系统投资相结合。公司与 TSMC 合作开发了 COUPE 技术，并发明了一系列技术，然后有意将这些专利授权给更广泛的供应链，目的是"keep it nice and open"，确保不会出现单点瓶颈。除了知识产权授权，Nvidia 还在积极发明"new workflows"和专用设备，比如用于测试的"double-sided probing"，这代表了制造流程的根本性创新。这个策略还延伸到直接的财务参与，Nvidia 投资供应链公司并帮助它们"scale up their capacity"，本质上是在为这些公司的扩张降低风险，以满足 Nvidia 未来的需求。这种全面的方法——涵盖专利、工艺创新、设备开发和产能融资——表明 Nvidia 正在主动"shape the ecosystem"，而不是简单地下订单然后期待供应商能够交付。

### Topic 8: Labor shortages: plumbers, electricians, and radiologists
劳动力短缺：水管工、电工和放射科医生
_[12:57]_

**Q:** What are the real bottlenecks in scaling, and why are predictions about job displacement often wrong?
**问：** 扩展规模的真正瓶颈是什么，为什么关于工作替代的预测经常是错误的？

**A:** Speaker A argues that labor shortages in skilled trades represent harder scaling bottlenecks than manufacturing challenges, using "plumbers and electricians" as the most difficult example. They critique automation doomers who discourage career entry by predicting job elimination, pointing to how these warnings create self-fulfilling shortages rather than displacement. The radiologist case illustrates this pattern perfectly: ten years ago doomers claimed "radiology is going to be the first career to go," yet today "we're short of radiologists." The same dynamic threatens software engineering, where discouraging new entrants based on AI displacement fears could create critical talent gaps rather than prevent unemployment.
**答：** Speaker A 认为技能型工种的劳动力短缺比制造业扩张更难解决，其中"水管工和电工"是最困难的例子。他批评那些预测工作会被自动化取代、从而劝阻人们进入某些职业的末日论者，指出这些警告反而造成了人才短缺而非失业。放射科医生的案例最能说明问题：十年前末日论者宣称"radiology 将是第一个消失的职业"，但现在"我们却缺放射科医生"。同样的逻辑也威胁着软件工程领域——如果因为担心 AI 替代而劝阻新人入行，最终会造成严重的人才缺口而非失业潮。

### Topic 9: Scaling EUV machines and manufacturing capacity
扩展 EUV 机器和制造产能
_[13:58]_

**Q:** How quickly can manufacturing bottlenecks like EUV machines be scaled, and what does it take?
**问：** 像 EUV 机器这样的制造瓶颈能多快扩展，需要什么条件？

**A:** Speaker A argues that EUV manufacturing bottlenecks are fundamentally solvable within "two or three years" once a clear demand signal exists, emphasizing that "once you can build one, you can build ten, and once you can build ten, you can build a million." The key is not technical difficulty but rather convincing critical players in the supply chain—particularly TSMC, which can then drive ASML to scale production. Speaker A distinguishes between direct engagement with some suppliers and indirect influence through "critical pinch points" like TSMC, suggesting a strategic approach to supply chain coordination. Crucially, while these bottlenecks take 2-3 years to resolve, computing efficiency improvements are advancing much faster at "10x, 20x" and even "30x to 50x" from Hopper to Blackwell, meaning performance gains significantly outpace manufacturing constraints.
**答：** Speaker A 认为 EUV 制造瓶颈在有明确需求信号的情况下，"两到三年"内就能解决，强调"一旦能造一台，就能造十台，能造十台就能造一百万台"。关键不在于技术难度，而在于说服供应链中的关键角色——尤其是 TSMC，进而推动 ASML 扩大产能。Speaker A 区分了直接接触某些供应商和通过 TSMC 这样的"关键节点"间接施加影响的策略。重要的是，虽然这些瓶颈需要 2-3 年解决，但计算效率的提升速度要快得多，达到"10 倍、20 倍"，从 Hopper 到 Blackwell 甚至"30 到 50 倍"，意味着性能提升远超制造限制。

### Topic 10: Computing efficiency improvements vs capacity constraints
计算效率提升与产能限制
_[15:04]_

**Q:** Why don't manufacturing bottlenecks worry Nvidia compared to downstream issues like energy policy?
**问：** 为什么相比能源政策等下游问题，制造瓶颈不让 Nvidia 担心？

**A:** The speaker argues that manufacturing bottlenecks are temporary problems lasting only "two, three years" while Nvidia's efficiency gains are exponential and continuous. He emphasizes that computing efficiency improvements of "10x, 20x" and even "30x to 50x" from Hopper to Blackwell, combined with algorithmic innovations enabled by CUDA's flexibility, far outpace any capacity constraints. The real concern lies in "downstream" structural issues, particularly energy policy, because "you can't create an industry without energy." He frames this in the context of ambitious reindustrialization goals—bringing back chip manufacturing, building "AI factories," EVs, and robots—all of which require long-term energy infrastructure that takes much longer to address than adding "more chip capacity" or "more CoWoS capacity."
**答：** 讲者认为制造瓶颈只是暂时问题，通常"两三年"就能解决，而 Nvidia 的效率提升是指数级且持续的。他强调从 Hopper 到 Blackwell 实现了"30x 到 50x"的计算效率提升，加上 CUDA 灵活性带来的算法创新，这些进步远超任何产能限制的影响。真正令人担忧的是"下游"的结构性问题，特别是能源政策，因为"没有能源就无法创造产业"。他将此与美国再工业化的宏大目标联系起来——重振芯片制造、建设"AI factories"、电动车和机器人产业——所有这些都需要长期的能源基础设施，解决时间远超增加"芯片产能"或"CoWoS 产能"所需的周期。

### Topic 11: Energy as the real constraint for reindustrialization
能源是再工业化的真正制约因素
_[16:08]_

**Q:** Why is energy the critical long-term constraint for building new industries and AI factories?
**问：** 为什么能源是建设新产业和 AI 工厂的关键长期制约？

**A:** Speaker A frames manufacturing capacity challenges as fundamentally solvable on a 2-3 year timeline, dismissing concerns about both "chip capacity" and "CoWoS capacity" as temporary bottlenecks rather than structural constraints. The speaker positions themselves as "the expert" when the interviewer acknowledges lacking "technical knowledge to adjudicate" between conflicting perspectives from different guests. This exchange reveals a confident assessment that semiconductor supply chain constraints are tractable engineering problems, implicitly contrasting with energy as a deeper, less fungible limitation. The brief mention suggests that while fab capacity can be expanded through capital investment and time, energy infrastructure represents a more fundamental barrier to scaling AI compute and reindustrialization.
**答：** Speaker A 认为制造产能挑战在根本上是可以在 2-3 年时间线内解决的，他将「chip capacity」和「CoWoS capacity」都视为临时瓶颈而非结构性制约。当主持人承认自己缺乏「technical knowledge to adjudicate」不同嘉宾的矛盾观点时，Speaker A 将自己定位为「the expert」。这段对话揭示了一个自信的判断：半导体供应链的制约是可处理的工程问题，这隐含地与能源形成对比——能源是更深层、更难替代的限制因素。简短的讨论暗示，虽然晶圆厂产能可以通过资本投资和时间来扩张，但能源基础设施代表了扩展 AI 算力和再工业化的更根本障碍。

### Topic 12: TPU competition and training top AI models
TPU 竞争与顶级 AI 模型训练
_[16:23]_

**Q:** What does it mean for Nvidia that top models like Claude and Gemini were trained on TPUs?
**问：** Claude 和 Gemini 等顶级模型在 TPU 上训练，对 Nvidia 意味着什么？

**A:** Speaker A reframes the competitive landscape by emphasizing that Nvidia builds "accelerated computing, not a tensor processing unit," positioning their offering as fundamentally broader in scope. Rather than directly addressing the TPU threat, they argue that Nvidia's platform serves diverse computational domains including "molecular dynamics, quantum chromodynamics, data processing" and physics simulations, not just AI workloads. This strategic framing suggests that while TPUs may excel at training specific AI models, Nvidia's value proposition lies in being a general-purpose acceleration platform. The speaker acknowledges that "AI is the conversation today" but implies this is only one use case among many, subtly downplaying the significance of losing some flagship AI training workloads to Google's specialized hardware.
**答：** 发言人 A 重新定义了竞争格局，强调 Nvidia 构建的是 "accelerated computing"（加速计算）而非单纯的张量处理单元，将自身定位为范围更广的平台。面对 TPU 的竞争，他没有正面回应，而是列举 Nvidia 平台服务于多样化的计算领域，包括分子动力学、量子色动力学、数据处理以及物理模拟等，而不仅仅是 AI 训练。这种策略性表述暗示，虽然 TPU 可能在训练特定 AI 模型上表现出色，但 Nvidia 的价值主张在于通用加速平台的定位。发言人承认 "AI is the conversation today"，但同时暗示这只是众多应用场景之一，巧妙地淡化了在部分旗舰 AI 训练任务上输给 Google 专用硬件的影响。

### Topic 13: Accelerated computing vs tensor processing units
加速计算与张量处理单元
_[17:22]_

**Q:** How does Nvidia's accelerated computing platform differ from specialized TPUs in market reach and flexibility?
**问：** Nvidia 的加速计算平台在市场覆盖和灵活性方面与专用 TPU 有何不同？

**A:** The speaker positions Nvidia's accelerated computing approach as fundamentally broader than AI-specific TPUs, emphasizing that their "market reach is far greater than any TPU or ASIC can possibly have" because they accelerate applications across all domains, not just AI workloads. A critical differentiator is operational flexibility: Nvidia systems are "designed to be operated by other people," enabling deployment across every major cloud provider (Google, Amazon, Azure, OCI) and diverse use cases, whereas home-built systems like TPUs require organizations to be their own operators. The CUDA platform serves as both "a fantastic tensor processing unit" and a comprehensive solution handling "every life cycle of data processing, computing, AI," which allows customers ranging from cloud providers to enterprises like Eli Lilly for drug discovery to xAI for custom supercomputers. This architectural flexibility creates a self-reinforcing advantage: because Nvidia supports "every application in the world now," system builders have guaranteed customer demand across industries, making their "market opportunity just a lot larger" than specialized accelerators constrained to narrower use cases.
**答：** 讲者认为 Nvidia 的加速计算方案在市场覆盖面上远超专用的 TPU，核心原因是他们不仅服务 AI 场景，而是加速各个领域的应用。关键差异在于运维灵活性：Nvidia 系统「designed to be operated by other people」，因此能部署在所有主流云平台（Google、Amazon、Azure、OCI）并支持多样化场景，而 TPU 这类自研系统通常要求用户自己运维。CUDA 平台既是「fantastic tensor processing unit」，也能处理「every life cycle of data processing, computing, AI」全流程，这让客户范围从云服务商到 Eli Lilly 的药物研发、xAI 的定制超算都能覆盖。这种架构灵活性形成正向循环：因为 Nvidia 支持「every application in the world now」，系统构建者能确保跨行业的客户需求，使其市场机会远大于功能受限的专用加速器。

### Topic 14: Revenue sources and AI-specific optimization debate
收入来源和 AI 专用优化争论
_[19:48]_

**Q:** Given that Nvidia's revenue comes primarily from AI, are TPUs better optimized for matrix multiplies that dominate AI workloads?
**问：** 鉴于 Nvidia 的收入主要来自 AI，TPU 是否更适合主导 AI 工作负载的矩阵乘法？

**A:** The questioner challenges Nvidia's architecture by noting that its "spectacular revenue" of $60 billion per quarter comes from AI's unprecedented growth, not diverse applications. They argue that TPUs, with their "big systolic array," are purpose-built for the "very predictable matrix multiplies" that define AI workloads, while GPUs sacrifice die area for flexibility features like "warp schedulers" and thread switching that AI doesn't need. The implication is that as AI becomes the dominant compute workload, specialized architectures optimized for matrix operations should have an advantage over general-purpose GPUs. However, the response pushes back on this reductionist view, asserting that "matrix multiplies are an important part of AI, but they're not the only part," suggesting that AI workloads require more architectural versatility than the question assumes.
**答：** 提问者质疑 Nvidia 的架构选择，指出其每季度 600 亿美元的收入来自 AI 的空前增长，而非多样化应用。他们认为 TPU 的 "big systolic array" 专为 AI 中 "very predictable matrix multiplies" 设计，而 GPU 为了灵活性牺牲了芯片面积用于 "warp schedulers" 和线程切换等 AI 不需要的功能。言下之意是，随着 AI 成为主导计算负载，专用于矩阵运算的架构应该比通用 GPU 更有优势。但回应者反驳了这种简化观点，强调 "matrix multiplies are an important part of AI, but they're not the only part"，暗示 AI 工作负载需要的架构多样性超出提问者的假设。

### Topic 15: Programmability enables algorithmic innovation
可编程性推动算法创新
_[21:01]_

**Q:** Why is general programmability crucial for inventing new AI architectures and achieving performance leaps beyond Moore's Law?
**问：** 为什么通用可编程性对于发明新 AI 架构和实现超越摩尔定律的性能飞跃至关重要？

**A:** The speakers argue that general programmability is the fundamental enabler of rapid AI advancement because it allows researchers to invent entirely new algorithms and architectures beyond matrix multiplication, such as "hybrid SSM" models or systems that "fuse diffusion and autoregressive techniques." Moore's Law only delivers about 25% annual improvement, but Nvidia achieved a 50x energy efficiency leap from Hopper to Blackwell—something "you can't reasonably do with just Moore's Law"—by co-designing hardware with new algorithmic approaches like MoEs that are "parallelized, disaggregated, and distributed." The ability to "get down and come up with new kernels with CUDA" is essential for implementing these innovations, and Nvidia's advantage lies in being "an extreme co-design company" that can even "offload some of the computation into the fabric itself, like NVLink." Without programmable architectures, the industry would be constrained to incremental hardware improvements rather than the algorithmic breakthroughs that drive 10x or 100x performance gains.
**答：** 两位嘉宾认为，通用可编程性是 AI 快速发展的根本推动力，因为它让研究者能够发明全新的算法和架构，而不仅仅局限于矩阵乘法，比如 hybrid SSM 模型或者融合 diffusion 和 autoregressive 技术的系统。摩尔定律每年只能带来约 25% 的性能提升，但 Nvidia 从 Hopper 到 Blackwell 实现了 50 倍的能效飞跃——这种提升"单靠摩尔定律根本做不到"——关键在于硬件与新算法方法（如 MoE）的协同设计，这些方法需要并行化、解耦和分布式计算。通过 CUDA "深入编写新 kernel" 的能力对于实现这些创新至关重要，而 Nvidia 的优势在于是一家"极致的协同设计公司"，甚至可以"将部分计算卸载到 NVLink 这样的互连结构中"。如果没有可编程架构，行业只能依赖硬件的渐进式改进，而无法实现带来 10 倍或 100 倍性能提升的算法突破。

### Topic 16: CUDA and extreme co-design across the stack
CUDA 和跨堆栈的极致协同设计
_[23:15]_

**Q:** How does Nvidia's co-design approach across processors, fabric, libraries, and algorithms enable dramatic efficiency gains?
**问：** Nvidia 跨处理器、互联、库和算法的协同设计方法如何实现显著的效率提升？

**A:** Nvidia's competitive advantage stems from being an "extreme co-design company" that can simultaneously optimize across multiple layers of the computing stack. The speaker emphasizes that CUDA's programmability is essential infrastructure that enables this coordination—without it, they "wouldn't even know where to start" orchestrating changes across such diverse components. This approach allows Nvidia to push computation into unconventional locations, including "offload some of the computation into the fabric itself" through technologies like NVLink and into networking infrastructure with Spectrum-X. The ability to "affect change across the processors, the system, the fabric, the libraries, and the algorithm simultaneously" represents a holistic optimization strategy that competitors focusing on individual components cannot easily replicate.
**答：** Nvidia 的竞争优势源于其作为"extreme co-design company"的定位，能够同时优化计算堆栈的多个层面。发言人强调 CUDA 的可编程性是实现这种协调的关键基础设施——没有它，他们甚至"wouldn't even know where to start"来协调如此多样化的组件。这种方法让 Nvidia 能够将计算推送到非常规位置，包括通过 NVLink 等技术"offload some of the computation into the fabric itself"，以及通过 Spectrum-X 将计算卸载到网络基础设施中。能够"affect change across the processors, the system, the fabric, the libraries, and the algorithm simultaneously"代表了一种整体优化策略，专注于单个组件的竞争对手很难复制。

### Topic 17: Crusoe cloud infrastructure and KV caching optimization
Crusoe 云基础设施和 KV 缓存优化
_[23:51]_

**Q:** How does Crusoe optimize inference performance through cross-user KV caching and what hardware do they offer?
**问：** Crusoe 如何通过跨用户 KV 缓存优化推理性能，他们提供什么硬件？

**A:** Crusoe differentiates itself not just through early access to cutting-edge hardware like "NVIDIA's Blackwell and Blackwell Ultra platforms" and the upcoming "Vera Rubin deployment," but through architectural innovations in inference optimization. While standard inference engines perform KV caching within a single user's session, Crusoe implements cross-user, cross-GPU KV caching that allows "a thousand agents" sharing the same system prompt to reuse a single computed KV cache across "every single GPU in the cluster." This optimization becomes particularly valuable as AI systems become "more agentic" with "much longer prefixes" needed for tool use and file access. The performance gains are substantial: Crusoe demonstrated "up to 10x faster time-to-first-token and up to 5x better throughput than vLLM" in recent benchmarks, positioning their infrastructure as suitable for both inference and training workloads without requiring users to switch cloud providers.
**答：** Crusoe 的差异化不仅体现在能够率先提供 NVIDIA Blackwell、Blackwell Ultra 以及即将推出的 Vera Rubin 等尖端硬件，更在于其推理优化的架构创新。传统推理引擎只在单个用户会话内做 KV 缓存，而 Crusoe 实现了跨用户、跨 GPU 的 KV 缓存共享机制：当一千个 agent 使用相同的 system prompt 时，Crusoe 只需计算一次 KV cache，就能让集群中的所有 GPU 复用。这种优化在 AI 系统变得更具 agent 特性、需要更长的 prefix 来使用工具和访问文件时尤为重要。性能提升非常显著：Crusoe 在基准测试中实现了比 vLLM "快 10 倍的首 token 时间"和"5 倍的吞吐量提升"，使其基础设施同时适用于推理和训练工作负载，用户无需切换云服务商。

### Topic 18: CUDA's role for hyperscalers writing custom kernels
CUDA 对自研内核的超大规模厂商的价值
_[25:00]_

**Q:** If hyperscalers like Anthropic, Google, and OpenAI write their own kernels and use alternative accelerators, is CUDA still critical for frontier AI?
**问：** 如果 Anthropic、Google 和 OpenAI 等超大规模厂商自己写内核并使用其他加速器，CUDA 对前沿 AI 还重要吗？

**A:** The speakers distinguish between two customer segments with fundamentally different needs: academic researchers who "just needed to run PyTorch with CUDA" versus hyperscalers who "have the resources to write their own kernels" to extract that critical "last 5% of performance." They observe that major AI labs are indeed moving away from CUDA dependencies—Anthropic and Google run TPUs and Trainium, while OpenAI uses Triton to compile "down to CUDA C++" with custom replacements for cuBLAS and NCCL that work across accelerators. Despite this migration among frontier labs, Speaker B reframes CUDA's value proposition: it remains "a rich ecosystem" where "building on CUDA first is incredibly smart," suggesting CUDA's role has shifted from being an irreplaceable runtime to being the optimal starting point for development before optimization and porting.
**答：** 两位嘉宾区分了两类用户的需求差异：学术研究者只需要"run PyTorch with CUDA"就够了，而超大规模厂商"有资源自己写内核"来榨取关键的"最后 5% 性能"。他们观察到主要 AI 实验室确实在摆脱 CUDA 依赖——Anthropic 和 Google 跑 TPU 和 Trainium，OpenAI 虽然用 GPU 但通过 Triton 编译到"CUDA C++"层面，用自研栈替代 cuBLAS 和 NCCL，并且能编译到其他加速器。尽管前沿实验室在迁移，Speaker B 重新定义了 CUDA 的价值：它仍是"丰富的生态系统"，"先在 CUDA 上构建是非常明智的"，暗示 CUDA 的角色已从不可替代的运行时转变为开发的最佳起点，之后再优化和移植。

### Topic 19: CUDA ecosystem richness and framework support
CUDA 生态系统的丰富性和框架支持
_[26:11]_

**Q:** What makes CUDA's ecosystem valuable for developers building AI systems?
**问：** 是什么让 CUDA 生态系统对构建 AI 系统的开发者有价值？

**A:** The speaker argues that CUDA's value proposition centers on ecosystem maturity and comprehensive framework support, which reduces debugging complexity in large-scale AI systems. Nvidia actively contributes to emerging frameworks, with "huge amounts of Nvidia technology" embedded in Triton's backend, and supports a rapidly expanding landscape including vLLM, SGLang, and new reinforcement learning frameworks like verl and NeMo RL as "post-training and reinforcement learning" explodes. The critical advantage is reliability: when building on CUDA, developers can trust that failures are "more likely in your code and not in the mountain of code underneath," allowing them to isolate problems in their own logic rather than questioning the foundational stack. While acknowledging "we still have lots of bugs ourselves," the speaker emphasizes that CUDA is "so well wrung out" through extensive real-world use that it provides a dependable foundation for complex AI development.
**答：** 讲者认为 CUDA 的核心价值在于生态成熟度和全面的框架支持，这能降低大规模 AI 系统的调试复杂度。Nvidia 积极贡献于新兴框架，在 Triton 后端嵌入了 "huge amounts of Nvidia technology"，并支持快速扩张的生态，包括 vLLM、SGLang，以及随着 "post-training and reinforcement learning" 爆发而出现的 verl 和 NeMo RL 等强化学习框架。关键优势是可靠性：基于 CUDA 开发时，开发者可以相信问题 "more likely in your code and not in the mountain of code underneath"，能够将故障定位在自己的逻辑而非底层技术栈。虽然承认 "we still have lots of bugs ourselves"，但讲者强调 CUDA 经过大量实际应用已经 "so well wrung out"，为复杂 AI 开发提供了可信赖的基础。

### Topic 20: Install base and cross-platform availability advantages
安装基数和跨平台可用性优势
_[27:31]_

**Q:** Why does Nvidia's install base and cloud presence matter for developers and AI companies?
**问：** 为什么 Nvidia 的安装基数和云端存在对开发者和 AI 公司很重要？

**A:** The speaker argues that Nvidia's competitive moat rests on three interconnected pillars, with install base being critical because developers fundamentally need their software to "run on a whole bunch of other computers." Nvidia's "several hundred million GPUs" spanning multiple generations (A10, A100, H100, H200, L and P series) create immediate utility—once you develop software or a model, "it's going to be useful everywhere" from robotics hardware to cloud infrastructure. The universal cloud presence makes Nvidia "genuinely unique" by eliminating platform lock-in concerns: AI companies uncertain about "which cloud service provider" they'll partner with can develop once and deploy anywhere, including on-premises. This combination of ecosystem richness, massive install base, and deployment versatility makes "CUDA invaluable" as Nvidia's "great treasure"—developers gain write-once, run-anywhere capability across an enormous existing footprint.
**答：** 讲者认为 Nvidia 的竞争优势建立在三个相互关联的支柱上，其中安装基数至关重要，因为开发者本质上需要他们的软件能够在大量其他计算机上运行。Nvidia 拥有数亿块 GPU，跨越多个世代（A10、A100、H100、H200、L 系列、P 系列），这创造了即时价值——一旦开发出软件或模型，它就能在任何地方发挥作用，从机器人硬件到云基础设施。无处不在的云端部署使 Nvidia 具有独特性，消除了平台锁定的顾虑：AI 公司即使不确定会与哪家云服务商合作，也可以一次开发、随处部署，包括本地部署。生态系统的丰富性、庞大的安装基数和部署的灵活性，共同使 CUDA 成为 Nvidia 的核心资产——开发者获得了在巨大现有基础上一次编写、到处运行的能力。

### Topic 21: Whether CUDA advantages matter to hyperscaler customers
CUDA 优势对超大规模客户是否重要
_[29:17]_

**Q:** Do CUDA's ecosystem advantages matter to hyperscalers who can build their own software stacks, and can Nvidia sustain high margins if customers can bypass the CUDA moat?
**问：** 对于有能力自建软件栈的超大规模云服务商，CUDA 的生态优势还重要吗？如果客户能绕过 CUDA 护城河，Nvidia 还能保持高利润率吗？

**A:** The speakers question whether Nvidia's CUDA moat remains defensible against its largest customers who generate most of its revenue and possess the capability to build custom software stacks. Speaker B highlights that tasks like writing efficient kernels for attention or MLP operations have "tight verification loops" that make them amenable to automated optimization, potentially enabling hyperscalers to develop their own solutions. Speaker A acknowledges that Nvidia may retain advantages through "great price performance" and superior hardware specs, but raises a critical concern about margin sustainability. Historically, Nvidia has maintained "over 70%" margins across AI hardware and software specifically because of the CUDA ecosystem lock-in, but if major customers can "afford to build instead of the CUDA moat," the competitive landscape may shift toward pure hardware specifications—flops and memory bandwidth per dollar—rather than software ecosystem advantages. This could fundamentally compress Nvidia's industry-leading margins even if they remain a preferred vendor.
**答：** 两位讨论者质疑 Nvidia 的 CUDA 护城河在面对其最大客户时是否还能守住。Speaker B 指出，像为 attention 或 MLP 操作编写高效 kernel 这类任务具有"紧密的验证循环"，很适合自动化优化，这让超大规模客户有可能开发自己的解决方案。Speaker A 承认 Nvidia 可能仍凭借"出色的性价比"和更优的硬件规格保持优势，但提出了一个关键担忧：利润率的可持续性。历史上，Nvidia 在 AI 硬件和软件领域保持着"超过 70%"的利润率，正是因为 CUDA 生态系统的锁定效应。但如果主要客户能够"负担得起自建而非依赖 CUDA 护城河"，竞争格局可能会转向纯硬件指标——每美元的算力和内存带宽——而非软件生态优势。即使 Nvidia 仍是首选供应商，这种转变也可能从根本上压缩其行业领先的利润率。

### Topic 22: Nvidia's engineering support and architecture optimization
Nvidia 的工程支持和架构优化
_[30:28]_

**Q:** How does Nvidia's deep engineering support help AI labs optimize performance on their GPUs?
**问：** Nvidia 的深度工程支持如何帮助 AI 实验室在其 GPU 上优化性能？

**A:** Nvidia deploys an extensive engineering team to AI labs because their GPU architectures require specialized expertise to extract maximum performance, unlike general-purpose CPUs. The speaker uses a vivid analogy: CPUs are "like a Cadillac" that anyone can drive comfortably, while Nvidia's accelerators are "like F1 racers" where reaching 100 mph is easy but pushing to the limit demands deep architectural knowledge. This expertise gap is significant—Nvidia's optimization work routinely delivers "2x" to "3x" performance improvements, sometimes even 50%, by creating specialized kernels using AI-assisted techniques. These gains compound dramatically across large GPU fleets of Hoppers and Blackwells, making Nvidia's engineering support a critical competitive advantage that the speaker believes will remain necessary "for quite some time."
**答：** Nvidia 向 AI 实验室派驻大量工程师，因为他们的 GPU 架构需要专业知识才能榨取最大性能，这点和通用 CPU 很不一样。讲者用了个生动的比喻：CPU "like a Cadillac"，谁都能开得舒服；而 Nvidia 的加速器 "like F1 racers"，开到 100 英里/小时很容易，但要推到极限就需要深厚的架构理解。这个专业门槛带来的差距很明显——Nvidia 的优化工作经常能带来 "2x" 到 "3x" 的性能提升，有时甚至达到 50%，方法是用 AI 辅助技术创建专门的 kernel。这些提升在大规模 Hopper 和 Blackwell GPU 集群上会产生巨大的复合效应，使得 Nvidia 的工程支持成为关键竞争优势，讲者认为这种需求在 "quite some time" 内都会持续存在。

### Topic 23: Performance per TCO claims and benchmark challenges
每 TCO 性能声明和基准测试挑战
_[32:09]_

**Q:** Why does Nvidia claim the best performance per TCO and challenge competitors like TPU and Trainium to prove their cost advantages?
**问：** 为什么 Nvidia 声称拥有最佳的每 TCO 性能,并挑战 TPU 和 Trainium 等竞争对手证明其成本优势?

**A:** Nvidia asserts that its computing stack delivers "the best performance per TCO in the world, bar none," directly linking this efficiency to revenue generation where doubling performance translates to doubled revenues. The speakers challenge competitors to substantiate their cost claims through public benchmarks, specifically pointing to Dylan's InferenceMAX platform that remains unused by TPU and Trainium providers. They express frustration that competitors like Trainium claim "40% cost advantages" and TPUs tout superior economics, yet refuse to participate in MLPerf or other transparent benchmarking exercises. The argument rests on "first principles" reasoning that competitor claims "make absolutely zero sense," suggesting Nvidia's market success stems fundamentally from superior TCO rather than other factors.
**答：** Nvidia 宣称其计算平台拥有"世界上最佳的每 TCO 性能",并强调这种效率直接转化为收入增长——性能翻倍意味着收入翻倍。发言人挑战竞争对手通过公开基准测试来证明其成本优势,特别指出 Dylan 的 InferenceMAX 平台已经开放使用,但 TPU 和 Trainium 都不愿参与。他们对竞争对手的做法感到不解:Trainium 声称有"40% 的成本优势",TPU 也宣传经济性更好,却都拒绝参加 MLPerf 等透明的基准测试。从"第一性原理"来看,竞争对手的说法"完全说不通",Nvidia 认为自己的市场成功根本上源于卓越的 TCO 表现。

### Topic 24: Customer distribution and external vs internal cloud usage
客户分布和外部与内部云使用情况
_[33:27]_

**Q:** How much of Nvidia's cloud business serves external customers versus internal hyperscaler use?
**问：** Nvidia 的云业务中有多少服务于外部客户而非超大规模云服务商的内部使用？

**A:** The speaker clarifies that while 60% of Nvidia's customers are the top five cloud providers, the vast majority of that business serves external customers rather than the hyperscalers' internal operations. He provides specific examples: "most of Nvidia in AWS is for external customers, not internal use," while Azure and OCI customers are "obviously all" or entirely external. The speaker attributes this customer preference to Nvidia's exceptional "reach" and ability to "bring them all of the great customers in the world" who are already built on Nvidia infrastructure. He describes a self-reinforcing "flywheel" driven by install base, architectural programmability, ecosystem richness, and the existence of "tens of thousands" of AI companies globally. The rhetorical question posed to AI startups—"what architecture would you choose?"—underscores Nvidia's dominant position as the default choice for new AI ventures seeking maximum compatibility and ecosystem support.
**答：** 发言人澄清说，虽然 Nvidia 60% 的客户是五大云服务商，但这些业务绝大部分服务于外部客户，而非云服务商自己的内部使用。他举了具体例子：AWS 上的 Nvidia 业务「大部分是为外部客户服务，而非内部使用」，而 Azure 和 OCI 的客户「显然全部」或完全是外部客户。发言人将云服务商的这种偏好归因于 Nvidia 卓越的「覆盖范围」以及能够「为他们带来全世界所有优质客户」的能力，而这些客户本身就构建在 Nvidia 基础设施之上。他描述了一个由安装基数、架构可编程性、生态系统丰富度以及全球「数万家」AI 公司共同驱动的自我强化「飞轮效应」。他向 AI 初创公司提出的反问——「你会选择什么架构？」——凸显了 Nvidia 作为新兴 AI 企业寻求最大兼容性和生态支持时的默认选择地位。

### Topic 25: The Nvidia flywheel: install base, ecosystem, and abundance
Nvidia 的飞轮效应：装机量、生态系统与供应充足
_[34:26]_

**Q:** What creates Nvidia's competitive flywheel for AI startups and customers?
**问：** Nvidia 在 AI 领域的竞争飞轮是如何形成的？

**A:** Nvidia's competitive moat stems from three self-reinforcing advantages that create a powerful flywheel effect. The speaker emphasizes that customers naturally gravitate toward "the most abundant" architecture with "the largest installed base" and "a rich ecosystem," all of which Nvidia dominates. Performance metrics drive this dominance: their "perf per dollar" delivers the "lowest cost tokens," while their industry-leading "perf per watt" means a one-gigawatt data center generates maximum revenue by producing the most tokens possible. The flywheel closes because Nvidia has "the most customers in the world," making their infrastructure the default choice for companies looking to rent compute capacity. This creates a virtuous cycle where abundance drives adoption, adoption expands the ecosystem, and the ecosystem reinforces abundance.
**答：** Nvidia 的竞争优势源于三个相互强化的因素，形成了强大的飞轮效应。客户会自然选择「最充足」的架构、「最大的装机量」以及「丰富的生态系统」，而 Nvidia 在这三方面都占据主导地位。性能指标是核心驱动力：他们的性价比（perf per dollar）能提供「最低成本的 token」，而业界领先的能效比（perf per watt）意味着一个 1 吉瓦的数据中心能通过生成最多的 token 来实现收益最大化。飞轮的闭环在于 Nvidia 拥有「全球最多的客户」，使其基础设施成为需要租用算力的公司的默认选择。这形成了一个良性循环：供应充足推动采用率，采用率扩大生态系统，生态系统又反过来强化供应优势。

### Topic 26: Market structure and foundation lab compute concentration
市场结构和基础模型实验室的算力集中度
_[35:25]_

**Q:** Is compute actually distributed across thousands of AI companies, or concentrated in a few foundation labs like Anthropic and OpenAI who can build alternative stacks? Why is Anthropic using TPUs if Nvidia has superior performance?
**问：** 算力实际上是分布在数千家 AI 公司中，还是集中在 Anthropic 和 OpenAI 等少数基础模型实验室中，这些实验室可以构建替代方案？如果 Nvidia 性能更优，为什么 Anthropic 使用 TPU？

**A:** Speaker B challenges the assumption that compute is evenly distributed, arguing that even through five hyperscalers, the actual users are concentrated in "big foundation labs" like Anthropic and OpenAI who have the resources and technical capability to make different accelerators work. Speaker A strongly rejects this premise, insisting "your premise is wrong" and emphasizing the issue is "too important to AI" and "too important to the future of science" to let the misconception stand. The exchange culminates in a pointed question about Anthropic's recent "multi-gigawatt deal with Broadcom and Google for TPUs" representing the "majority of their compute," which directly contradicts the claim that Nvidia's superior price-performance would naturally dominate. Speaker A's vehement pushback suggests the market structure is more distributed or competitive than the concentration narrative implies, though the specific correction is deferred.
**答：** Speaker B 质疑算力分布的假设，认为即使通过五大云服务商，实际使用者也集中在 Anthropic 和 OpenAI 这样的「大型基础模型实验室」，它们有资源和技术能力让不同的加速器工作。Speaker A 强烈反对这个前提，坚称「你的前提是错的」，并强调这个问题「对 AI 太重要了」、「对科学的未来太重要了」，不能让这种误解继续。讨论最后聚焦在一个尖锐的问题上：Anthropic 最近宣布与 Broadcom 和 Google 达成「multi-gigawatt 级别的 TPU 交易」，占其「大部分算力」，这直接与 Nvidia 性能优势应该主导市场的说法矛盾。Speaker A 的激烈反驳暗示市场结构比算力集中叙事所描述的更分散或更有竞争性，但具体的纠正被推迟了。

### Topic 27: AI companies' accelerator choices beyond Nvidia
AI公司在Nvidia之外的加速器选择
_[36:42]_

**Q:** Why are major AI companies using TPUs and other accelerators instead of only Nvidia GPUs?
**问：** 为什么主要AI公司会使用TPU和其他加速器，而不是只用Nvidia GPU？

**A:** Speaker B argues that the apparent diversification away from Nvidia is largely illusory, driven almost entirely by one company: Anthropic. He emphasizes that "without Anthropic, why would there be any TPU growth at all? It's 100% Anthropic" and the same applies to AWS Trainium adoption. While acknowledging that companies like OpenAI are exploring AMD partnerships and building their own Titan accelerator, he maintains "they're vastly Nvidia" in their actual compute infrastructure. Speaker B frames alternative accelerator experiments as necessary market validation—"if they don't try these other things, how would they know how good ours is?"—rather than genuine competitive threats. He points to the high failure rate of ASIC projects and notes that "you still have to build something better than Nvidia," suggesting confidence that Nvidia's position remains secure despite exploration of alternatives.
**答：** Speaker B认为AI公司看似在分散使用不同加速器，但实际上这主要是由Anthropic一家公司推动的假象。他强调"without Anthropic"的话，TPU和AWS Trainium根本不会有增长，"It's 100% Anthropic"。虽然OpenAI等公司在尝试AMD合作和自研Titan加速器，但他们的计算基础设施"vastly Nvidia"。Speaker B将这些替代方案的尝试视为必要的市场验证——"if they don't try these other things, how would they know how good ours is"——而非真正的竞争威胁。他指出很多ASIC项目最终被取消，并表示"you still have to build something better than Nvidia"，显示出对Nvidia地位的信心。

### Topic 28: Economics of custom ASICs vs Nvidia
定制ASIC与Nvidia的经济性对比
_[38:09]_

**Q:** Can custom ASICs compete with Nvidia on cost given margin structures?
**问：** 考虑到利润率结构，定制ASIC能否在成本上与Nvidia竞争？

**A:** Speaker B argues that the economic case for custom ASICs is weaker than commonly assumed because margin structures don't favor them as much as expected. While Nvidia operates at roughly "70%" margins, ASIC providers like Broadcom also command "65%" margins, meaning the actual cost savings are minimal—only about 5 percentage points. Beyond margins, B emphasizes that building something "better than Nvidia" is extremely difficult given Nvidia's unmatched scale and velocity in delivering "big leaps, every single year." The counterargument that ASICs only need to be "not more than 70% worse" to justify the margin difference is dismissed because the margin gap itself is too narrow to compensate for Nvidia's technical advantages. B suggests that for ASICs to make sense, "Nvidia's got to be missing something, seriously," implying that only significant gaps in Nvidia's offering would justify the ASIC investment.
**答：** Speaker B认为定制ASIC的经济优势被高估了，因为利润率结构并不像预期那样有利。虽然Nvidia的利润率约为70%，但Broadcom等ASIC供应商的利润率也高达65%，实际成本节省只有5个百分点左右。除了利润率问题，B强调考虑到Nvidia无与伦比的规模和每年都能实现"big leaps"的迭代速度，要造出比Nvidia更好的产品极其困难。有人认为ASIC只需要"不比Nvidia差70%以上"就能抵消利润率差异，但B驳斥了这一观点，因为利润率差距太小，无法弥补Nvidia的技术优势。B暗示只有当"Nvidia's got to be missing something, seriously"时，即Nvidia的产品存在重大缺陷时，投资ASIC才有意义。

### Topic 29: Nvidia's missed opportunity with Anthropic
Nvidia错失投资Anthropic的机会
_[39:06]_

**Q:** Why didn't Nvidia invest in Anthropic early when Google and AWS did?
**问：** 为什么Nvidia没有像Google和AWS那样早期投资Anthropic？

**A:** The speaker acknowledges missing the Anthropic investment opportunity due to two factors: a failure to "deeply internalize" that foundation AI labs required massive capital commitments that traditional VCs wouldn't provide, and Nvidia's lack of financial capacity at the time to make "multi-billion dollar" investments. Google and AWS secured their positions by investing heavily upfront in exchange for compute commitments, a strategic move Nvidia couldn't match then. The speaker frames this as both a conceptual miss—not recognizing that AI labs "really had no other options" for funding—and a practical constraint, noting "even if I understood it, I don't think we would've been in a position to do that." However, the lesson has been learned: Nvidia has since invested heavily in OpenAI and later joined Anthropic when the opportunity arose, with the speaker emphasizing they won't "make that same mistake again."
**答：** 发言人承认错失投资Anthropic的机会源于两个因素：一是没有深刻理解基础AI实验室需要巨额资本投入，而传统VC不会提供这种规模的资金；二是当时Nvidia没有足够的财力进行数十亿美元级别的投资。Google和AWS通过前期大额投资换取了计算资源使用承诺，这是Nvidia当时无法做到的战略布局。发言人将此归结为认知上的失误——没有意识到AI实验室"真的别无选择"——以及实际能力的限制。不过这个教训已经吸取：Nvidia后来重金投资了OpenAI，并在有机会时加入了Anthropic的投资，发言人强调不会再犯同样的错误。

### Topic 30: Nvidia's current investments in OpenAI and Anthropic
Nvidia对OpenAI和Anthropic的投资时机
_[41:06]_

**Q:** Why didn't Nvidia invest earlier when valuations were lower and they had the cash?
**问：** 为什么Nvidia没有在估值更低时提前投资？

**A:** Speaker A explains that Nvidia invested "as soon as we could have" and would have done it earlier if possible, but the company lacked both the investment sensibility and recognition of necessity at the time. Nvidia had "never invested outside the company" at such scale and initially assumed these AI labs "could just go raise from VCs, for God's sakes, like all companies do." The speaker acknowledges a fundamental misunderstanding: "what they were trying to do couldn't have been done through VCs"—a realization that came too late for Anthropic, which had to seek funding elsewhere. Despite this missed opportunity and implicit regret, the speaker expresses genuine satisfaction that both companies succeeded, stating "Anthropic's existence is great for the world" and crediting OpenAI's "genius" for recognizing early on that their ambitions required unconventional funding structures.
**答：** Speaker A坦承Nvidia是"as soon as we could have"才投资的，如果可能会更早行动，但当时公司既没有大规模对外投资的意识，也没有意识到必要性。Nvidia此前"never invested outside the company"达到这种规模，最初以为这些AI实验室"could just go raise from VCs"就像所有公司一样。Speaker A承认了一个根本性的误判："what they were trying to do couldn't have been done through VCs"——这个认知来得太晚，导致Anthropic不得不寻求其他资金来源。尽管错失了机会并隐含遗憾，Speaker A仍真诚地表示两家公司的成功令人欣慰，认为"Anthropic's existence is great for the world"，并赞扬OpenAI的"genius"在于早期就认识到他们的雄心需要非常规的融资结构。

### Topic 31: Why Nvidia Supports Cloud Providers Instead of Becoming One
Nvidia为何支持云服务商而非自己做云
_[43:21]_

**Q:** Why doesn't Nvidia become a hyperscaler and rent compute directly instead of supporting middlemen like CoreWeave?
**问：** 为什么Nvidia不直接成为超大规模云服务商出租算力，而是支持CoreWeave这样的中间商？

**A:** Nvidia follows a core philosophy of doing "as much as needed, as little as possible," focusing resources only on work that wouldn't happen without them. The speaker argues that foundational infrastructure like NVLink, CUDA (developed over 20 years while losing money), domain-specific libraries, and cuLitho for computational lithography would never have been created if Nvidia hadn't taken those risks. In contrast, cloud services represent a market where "if I didn't do it, somebody would show up," making it unnecessary for Nvidia to enter directly. Instead, Nvidia strategically invests in enabling "neoclouds" and "AI clouds" like CoreWeave, Nscale, and Nebius that wouldn't exist without their support, allowing these companies to convert CapEx into OpEx for AI labs while Nvidia focuses on platform innovation that only they can deliver.
**答：** Nvidia遵循一个核心理念："做必须做的事，但尽可能少做"，只把资源投入到没有他们就不会发生的工作上。发言人认为，像NVLink、CUDA（亏损开发了20年）、各领域专用库以及计算光刻的cuLitho这些基础设施，如果Nvidia不冒险去做，根本不会有人做出来。相比之下，云服务是一个"如果我不做，也会有人出现"的市场，Nvidia没必要亲自进入。Nvidia的策略是投资扶持CoreWeave、Nscale、Nebius这些"新云"和"AI云"，让它们帮AI实验室把资本支出转化为运营支出，而Nvidia专注于只有自己能做的平台创新。

### Topic 32: Nvidia's philosophy of not picking winners
Nvidia不挑选赢家的哲学
_[46:25]_

**Q:** Why does Nvidia invest in all foundation model companies instead of picking winners?
**问：** 为什么Nvidia投资所有基础模型公司而不是挑选赢家？

**A:** Nvidia deliberately avoids picking winners among foundation model companies, choosing instead to "invest in all of them" as both a business imperative and a matter of humility. The speaker grounds this philosophy in Nvidia's own unlikely survival story: among 60 3D graphics companies at the start, Nvidia "would be at the top of that list not to make it" because their architecture was "precisely wrong" and "impossible thing for developers to support." This near-death experience taught them that predicting winners is futile—even well-reasoned first principles can lead to the wrong solution initially. By supporting the entire ecosystem rather than selecting favorites, Nvidia ensures its architecture can "connect with as many industries as possible" and enables the broader vision of building "the planet" on AI and "the American tech stack." The approach reflects both strategic pragmatism ("it's imperative to our business") and philosophical conviction that it's "not our job" to determine which companies succeed.
**答：** Nvidia刻意避免在基础模型公司中挑选赢家，而是选择"投资所有公司"，这既是业务需要也是一种谦逊态度。演讲者用Nvidia自身不太可能存活的故事来支撑这一理念：在最初的60家3D图形公司中，Nvidia"会排在最不可能成功的名单前列"，因为他们的架构"完全错误"且"开发者不可能支持"。这次濒死经历让他们明白预测赢家是徒劳的——即使基于良好的第一性原理推理，最初也可能得出错误方案。通过支持整个生态系统而非挑选偏好对象，Nvidia确保其架构能够"连接尽可能多的行业"，并实现让"整个星球"建立在AI和"美国技术栈"之上的更大愿景。这种做法既体现了战略务实性（"对我们的业务至关重要"），也体现了哲学信念——判断哪些公司会成功"不是我们的工作"。

### Topic 33: Supporting neoclouds without picking favorites
支持新兴云服务商而不偏袒
_[48:31]_

**Q:** How does Nvidia support neoclouds like CoreWeave without playing favorites?
**问：** Nvidia如何在支持CoreWeave等新兴云服务商的同时不偏袒任何一方？

**A:** Nvidia's support for neoclouds is demand-driven rather than proactive favoritism—the key distinction is that these companies "need to want to exist" and must initiate the relationship by asking for help. The speaker emphasizes a threshold-based approach: neoclouds must demonstrate their own "business plan, expertise" and "passion" before Nvidia considers involvement. Support is conditional on the neocloud having substantive "capabilities themselves," with Nvidia stepping in specifically when "they need some investment in order to get it off the ground." This framework allows Nvidia to enable neoclouds that wouldn't otherwise exist while maintaining that they're responding to qualified requests rather than arbitrarily choosing winners in the market.
**答：** Nvidia对新兴云服务商的支持是需求驱动而非主动偏袒——关键区别在于这些公司必须"自己想要存在"并主动寻求帮助。发言人强调了一种门槛机制：新兴云服务商必须先展示自己的"商业计划、专业能力"和"热情"，Nvidia才会考虑介入。支持是有条件的，这些公司必须具备实质性的"自身能力"，Nvidia只在"他们需要投资来启动业务"时才介入。这个框架让Nvidia能够扶持那些原本无法存在的新兴云服务商，同时保持自己是在响应合格请求，而不是在市场上任意挑选赢家。

### Topic 34: Nvidia's approach to financing and investment strategy
Nvidia 的融资和投资策略
_[49:11]_

**Q:** Does Nvidia want to be in the financing business, and when does it make strategic investments?
**问：** Nvidia 是否想进入融资业务,以及何时进行战略投资?

**A:** Nvidia explicitly avoids becoming a financier, preferring to "work with all the people in the financing business" rather than compete with them. The company's philosophy is to "keep our business model as simple as possible" and focus on its core competencies while supporting its ecosystem. However, Nvidia makes selective strategic investments when companies critically need their participation and when the speaker has deep conviction in their future—exemplified by OpenAI's $30 billion pre-IPO investment. The investment criteria centers on necessity rather than opportunity: they invest "because they need us to do it," not to maximize investment activity. The guiding principle is minimalism: "we're not trying to do as much as possible, we're trying to do as little as possible," reflecting a disciplined approach that reserves capital deployment for situations where Nvidia's involvement is essential to companies the world needs to exist.
**答：** Nvidia 明确表示不想成为金融机构,更愿意"与所有金融机构合作"而非与之竞争。公司的理念是"保持商业模式尽可能简单",专注于核心业务并支持生态系统。但 Nvidia 会在特定情况下进行战略投资:当公司迫切需要他们参与,且对其前景有深刻信念时——OpenAI 的 300 亿美元 pre-IPO 投资就是典型案例。投资标准围绕"必要性"而非机会主义:"because they need us to do it",而不是为了最大化投资活动。指导原则是极简主义:"trying to do as little as possible",体现了一种克制的策略——只在 Nvidia 的参与对世界所需的公司至关重要时才动用资本。

### Topic 35: GPU allocation and shortage management
GPU 分配和短缺管理
_[51:13]_

**Q:** How does Nvidia allocate scarce GPUs among customers during shortages?
**问：** 在短缺期间 Nvidia 如何在客户之间分配稀缺的 GPU?

**A:** Nvidia's allocation strategy is fundamentally "first in, first out" based on purchase orders rather than strategic market fragmentation or highest bidder wins. The speaker emphasizes that "if you don't place a PO, all the talking in the world won't make a difference," rejecting the premise that Nvidia deliberately divvies up supply to favor neoclouds or specific players. The process begins with extensive forecasting to align supply and demand given long manufacturing and data center build times, but ultimately requires customers to commit with actual orders. Nvidia may adjust delivery timing based on operational readiness—if a customer's "data center's not ready, or certain components aren't ready," they might serve another customer first to "maximize the throughput of our own factory." The speaker directly refutes viral narratives, clarifying that stories about executives like Larry and Elon "begging for GPUs" at dinner are false: "at no time did they beg for GPUs. They just had to place an order."
**答：** Nvidia 的分配策略本质上是基于采购订单的"先到先得"，而非战略性的市场分割或价高者得。发言人强调"如果你不下 PO，说再多也没用"，否认了 Nvidia 刻意将供应分配给 neocloud 或特定玩家的说法。流程始于大量的需求预测工作，因为制造和数据中心建设都需要很长时间，但最终需要客户用实际订单来承诺。Nvidia 可能会根据运营准备情况调整交付时间——如果客户的"数据中心还没准备好，或某些组件还没到位"，他们可能会优先服务其他客户以"最大化自己工厂的吞吐量"。发言人直接反驳了病毒式传播的叙事，澄清关于 Larry 和 Elon 在晚餐上"乞求 GPU"的故事是假的："他们从未乞求过 GPU，只是需要下订单"。

### Topic 36: Nvidia's fixed pricing philosophy and long-term customer trust
Nvidia 的固定定价理念与长期客户信任
_[53:55]_

**Q:** Why doesn't Nvidia use dynamic pricing or auction-based allocation during high demand periods?
**问：** 为什么 Nvidia 在需求旺盛时不采用动态定价或竞价分配机制?

**A:** Nvidia deliberately avoids dynamic pricing despite extreme demand, viewing it as "a bad business practice" that would undermine their role as a dependable industry foundation. The company maintains a fixed-price model where "you set your price and then people decide to buy it or not," refusing to adjust prices upward even when "demand goes through the roof." This approach contrasts with other chip companies that "change their prices when demand is higher," but Nvidia prioritizes being a partner "you can count on" where customers don't need to "second-guess" quoted prices. The speaker frames this as a strategic choice to build long-term trust and position Nvidia as the stable infrastructure layer of the AI industry, suggesting this same reliability philosophy extends to their relationship with manufacturing partner TSMC.
**答：** Nvidia 在需求极度旺盛时刻意避免动态定价,认为这是「不好的商业做法」,会削弱他们作为可靠行业基础的角色。公司坚持固定价格模式,即「你定好价格,然后客户决定买不买」,即使「需求爆表」也拒绝上调价格。这种做法与其他会「在需求高时调价」的芯片公司形成对比,但 Nvidia 优先考虑成为客户「可以依赖」的合作伙伴,让客户不必「猜测」报价是否会变。发言人将此定位为建立长期信任的战略选择,把 Nvidia 打造成 AI 行业稳定的基础设施层,并暗示这种可靠性理念同样延伸到他们与制造合作伙伴 TSMC 的关系中。

### Topic 37: Nvidia-TSMC partnership and product roadmap reliability
Nvidia 与 TSMC 的合作关系及产品交付可靠性
_[55:02]_

**Q:** What makes Nvidia's relationship with TSMC unique and how reliable is Nvidia's product delivery?
**问：** Nvidia 与 TSMC 的合作关系有何独特之处?Nvidia 的产品交付可靠性如何?

**A:** Nvidia's nearly 30-year relationship with TSMC operates without legal contracts, relying instead on "rough justice" where both parties accept that deals sometimes favor one side or the other, building what the speaker calls "incredible" trust and dependability. The speaker positions Nvidia as uniquely reliable in delivering annual product generations—citing the roadmap from Vera Rubin to Vera Rubin Ultra to Feynman—and challenges listeners to "find another ASIC team" that can guarantee year-over-year delivery with token costs decreasing "by an order of magnitude every single year." Nvidia's scalability spans from single graphics cards to "$100 billion AI factory" orders, a flexibility the speaker claims "no other company in the world" can match today. This reliability, described as taking "a couple of decades" to establish, positions Nvidia as the foundational infrastructure for the AI industry, with TSMC being the only foundry partner that enables such consistent execution at scale.
**答：** Nvidia 与 TSMC 近 30 年的合作关系没有法律合同约束,而是基于「rough justice」原则——双方接受有时这方占优、有时那方占优的现实,从而建立起极高的信任和依赖。发言人强调 Nvidia 在年度产品迭代上的可靠性是独一无二的,从 Vera Rubin 到 Vera Rubin Ultra 再到 Feynman 的路线图清晰可期,并挑战听众「去找另一个 ASIC 团队」能保证每年交付且 token 成本「降低一个数量级」。Nvidia 的规模弹性从单张显卡到「1000 亿美元 AI 工厂」订单都能承接,这种能力在当今世界「没有其他公司」能做到。这种可靠性是「几十年」积累的结果,使 Nvidia 成为 AI 产业的基础设施,而 TSMC 是唯一能支撑这种规模化稳定交付的代工厂合作伙伴。

### Topic 38: China export controls and AI capabilities
对中国的出口管制和 AI 能力
_[57:35]_

**Q:** Should the US restrict AI chip exports to China given potential cyber-offensive capabilities of advanced models?
**问：** 鉴于先进模型潜在的网络攻击能力,美国是否应该限制向中国出口 AI 芯片?

**A:** Speaker B argues that export controls are ineffective because China already possesses the fundamental resources needed to develop advanced AI capabilities. He points out that Anthropic's Mythos model, which demonstrated extraordinary cyber-offensive capabilities including finding vulnerabilities that existed "for 27 years" in security-focused systems, was trained on "fairly mundane capacity" that is "abundantly available in China." China's existing advantages are substantial: they "manufacture 60% of the world's mainstream chips," employ "50% of the world's AI researchers," and have "an abundance of energy." Rather than export controls creating safety, B suggests that "victimizing them, turning them into an enemy" is counterproductive, though he acknowledges they remain "an adversary" and affirms wanting "the United States to win."
**答：** 嘉宾 B 认为出口管制实际上无效,因为中国已经具备开发先进 AI 能力所需的基础资源。他指出 Anthropic 的 Mythos 模型展现了极强的网络攻击能力,甚至能发现安全系统中存在了 27 年的漏洞,但这个模型的训练算力其实"相当普通",而这种算力在中国"大量存在"。中国的现有优势很明显:生产全球 60% 的主流芯片,拥有全球 50% 的 AI 研究人员,还有充足的能源。B 认为,与其通过出口管制"把他们变成受害者和敌人",不如寻找其他方式创造安全的世界,尽管他承认中国是"对手",美国应该保持领先。

### Topic 39: US-China AI research dialogue and cooperation
美中 AI 研究对话与合作
_[59:55]_

**Q:** What is the best approach to managing AI safety concerns with China as an adversary?
**问：** 在中国作为对手的情况下,管理 AI 安全问题的最佳方法是什么?

**A:** The speakers argue that despite China being an adversary, maintaining research dialogue is "the safest thing to do" rather than complete antagonism. They emphasize that AI researchers from both countries must talk and agree on prohibited uses, while acknowledging the U.S. should "win" the competition. A critical insight is that AI safety depends on a "richness of the ecosystem" where thousands of AI agents monitor and secure powerful AI systems, and this ecosystem fundamentally "needs open source" and "open models." The speakers warn against suffocating China's contributions to open source AI, noting that "a lot of that is coming out of China," because restricting it would undermine the very safety infrastructure needed to keep AI secure. They frame this as a pragmatic tradeoff: maintaining U.S. computational advantage while preserving the collaborative open-source ecosystem essential for AI safety.
**答：** 两位嘉宾认为,尽管中国是对手,但保持研究对话才是「最安全的做法」,而非完全对抗。他们强调美中两国的 AI 研究人员必须交流并就禁止用途达成共识,同时承认美国应该在竞争中「获胜」。一个关键洞察是,AI 安全依赖于「丰富的生态系统」,需要成千上万个 AI agent 来监控和保护强大的 AI 系统,而这个生态系统从根本上「需要 open source」和「open models」。嘉宾警告不要扼杀中国对开源 AI 的贡献,因为「很多都来自中国」,限制它会破坏保障 AI 安全所需的基础设施。他们将此框定为务实的权衡:在保持美国计算优势的同时,维护对 AI 安全至关重要的开源协作生态系统。

### Topic 40: Open source AI ecosystem and tech stack strategy
开源 AI 生态系统和技术栈策略
_[01:02:37]_

**Q:** Why is maintaining a unified open source AI ecosystem on American tech stack important for US interests?
**问：** 为什么在美国技术栈上维护统一的开源 AI 生态系统对美国利益很重要?

**A:** The speaker argues that the US should ensure "all the AI developers in the world are developing on the American tech stack" to keep open source AI advancements within the American ecosystem. The critical strategic error would be creating "two ecosystems" where open source AI runs only on foreign infrastructure while closed systems remain on American infrastructure—described as "a horrible outcome for the United States." While acknowledging China's compute capacity, the speaker notes their chip manufacturing limitations at 7nm without EUV technology result in "one-tenth the amount of flops that the US has." This computational advantage means "American labs are able to get to these levels of capabilities first," even though Chinese labs could eventually train comparable models. The strategy prioritizes maintaining US energy capacity to avoid making it "a bottleneck" while leveraging the flops advantage to stay ahead in the AI race.
**答：** 讲者认为美国应该确保"全世界所有 AI 开发者都在 American tech stack 上开发"，这样才能让开源 AI 的进步留在美国生态系统内。最大的战略失误是形成"两个生态系统"——开源 AI 只在外国基础设施上运行，而闭源系统留在美国基础设施上，这会是"对美国非常糟糕的结果"。虽然承认中国有算力，但讲者指出由于中国芯片制造停留在 7nm 且没有 EUV 技术，实际产生的算力"只有美国的十分之一"。这种算力优势意味着"美国实验室能够率先达到这些能力水平"，尽管中国实验室最终也能训练出类似模型。这个策略的重点是保持美国能源供应不成为"瓶颈"，同时利用算力优势在 AI 竞赛中保持领先。

### Topic 41: Anthropic's early access and vulnerability patching strategy
Anthropic 的提前访问与漏洞修补策略
_[01:04:12]_

**Q:** How does early access to AI models help American companies prepare for security threats?
**问：** 提前访问 AI 模型如何帮助美国公司应对安全威胁？

**A:** The speaker argues that Anthropic's strategy of holding new models for a month while giving American companies early access allows them to "patch up all their vulnerabilities" before public release, creating a defensive advantage. This approach is grounded in the belief that compute is the critical bottleneck—both for developing dangerous capabilities and for deploying them at scale, where "a million" cyber hackers are far more dangerous than "a thousand." The speaker emphasizes that compute availability directly determines AI researcher productivity and that "any AI lab in America" as well as Chinese labs like DeepSeek and Qwen cite compute as their primary constraint. The strategic implication is that American companies, having more compute resources, should leverage this advantage to reach "Mythos-level capabilities first" and prepare society for the risks before adversaries with less compute can catch up, maintaining a persistent lead in both capability development and defensive readiness.
**答：** 讲者认为 Anthropic 的策略是在公开发布新模型前保留一个月，期间让美国公司提前访问以"修补所有漏洞"，从而建立防御优势。这一策略基于一个核心观点：算力是关键瓶颈——既决定了危险能力的开发，也决定了大规模部署的可能性，因为拥有"一百万"个网络黑客比"一千个"危险得多。讲者强调算力直接决定 AI 研究人员的生产力，"美国的任何 AI 实验室"以及 DeepSeek、Qwen 等中国实验室都将算力视为主要限制因素。战略含义是：美国公司拥有更多算力，应该利用这一优势率先达到"Mythos 级别的能力"，在算力较少的对手赶上之前为社会做好准备，从而在能力开发和防御准备方面始终保持领先。

### Topic 42: China's compute capacity and infrastructure advantages
中国的算力和基础设施优势
_[01:04:54]_

**Q:** Does China have sufficient compute resources despite chip restrictions?
**问：** 尽管受到芯片限制，中国是否有足够的算力资源？

**A:** The speaker argues that China has more than sufficient compute capacity to develop advanced AI, dismissing concerns about chip export restrictions. China is "the second largest computing market in the world" with enormous aggregate compute potential, and crucially, AI is fundamentally "a parallel computing problem" where energy abundance matters more than cutting-edge chips. The speaker points to China's massive infrastructure overcapacity—"ghost datacenters" that are "fully powered" but empty—which enables them to compensate for less advanced process nodes by simply "gang up more chips, even if they're 7nm." China's semiconductor manufacturing capacity "monopolize mainstream chips" with "over-capacity," meaning the threshold needed for AI development concerns "they've already reached that threshold and beyond." The speaker emphasizes that AI is "a five-layer cake, and at the lowest layer is energy," suggesting China's energy infrastructure advantage fundamentally changes the compute equation regardless of chip sophistication.
**答：** 讲者认为中国拥有充足的算力来发展先进AI，芯片出口限制的担忧被夸大了。中国是"全球第二大计算市场"，拥有巨大的算力聚合潜力，而且AI本质上是"并行计算问题"，能源充裕比尖端芯片更重要。讲者指出中国存在大量基础设施过剩——"ghost datacenters"完全供电却空置——这使得他们可以通过"堆叠更多芯片，哪怕是7nm"来弥补制程劣势。中国的半导体制造产能"垄断主流芯片"且存在"产能过剩"，意味着AI发展所需的算力门槛"他们早已达到甚至超越"。讲者强调AI是"五层蛋糕，最底层是能源"，暗示中国的能源基础设施优势从根本上改变了算力竞争的逻辑，芯片先进程度反而不是决定性因素。

### Topic 43: Energy abundance compensating for chip limitations
能源优势如何弥补芯片代差
_[01:06:59]_

**Q:** How does China's energy abundance offset their disadvantage in advanced chip manufacturing?
**问：** 中国的能源优势如何抵消先进芯片制造上的劣势？

**A:** The speaker frames AI infrastructure as a "five-layer cake" where energy sits at the foundation, creating a fundamental tradeoff between chip efficiency and energy availability. He argues that the U.S. faces energy scarcity, forcing Nvidia to pursue "extreme co-design" and maximize "throughput per watt," while countries with abundant cheap energy can compensate by using older generation chips at scale. Specifically, he notes that "7nm chips are essentially Hopper" and that "today's models are largely trained on Hopper generation," suggesting that older process nodes remain capable enough when energy constraints are removed. He points to Huawei's record-breaking year as evidence that China can manufacture sufficient quantities of these less advanced chips, implying that energy abundance combined with manufacturing scale creates a viable alternative path to AI capability despite lacking cutting-edge semiconductor technology.
**答：** 讲者将AI基础设施比作"五层蛋糕"，能源是最底层，芯片效率和能源可用性之间存在根本性的权衡关系。他认为美国面临能源稀缺，迫使Nvidia追求"极致协同设计"来最大化每瓦性能，而拥有充足廉价能源的国家可以通过大规模使用老一代芯片来弥补。具体来说，他指出"7nm芯片基本等同于Hopper"，而且"当今的模型大多是在Hopper这一代上训练的"，这表明当能源限制被解除时，较老的制程节点仍然足够强大。他以Huawei创纪录的年度业绩为证据，说明中国能够制造足够数量的这些非尖端芯片，暗示能源优势加上制造规模可以开辟一条通往AI能力的可行路径，即使缺乏最先进的半导体技术。

### Topic 44: Memory bandwidth and networking solutions debate
内存带宽和网络解决方案争论
_[01:08:18]_

**Q:** Can China overcome memory bandwidth limitations through networking and silicon photonics?
**问：** 中国能否通过网络和硅光子技术克服内存带宽限制？

**A:** Speaker A argues that China's semiconductor limitations are overstated, pointing to Huawei shipping "millions" of chips in their largest year ever, far exceeding what companies like Anthropic deploy. While Speaker B raises concerns about memory bandwidth bottlenecks—noting that HBM2 versus newer memory could show "almost an order of magnitude difference"—Speaker A counters that networking solutions can compensate, emphasizing that Huawei is fundamentally "a networking company." A specifically dismisses the need for EUV lithography for advanced memory, arguing chips can be "gang[ed] together" using techniques like NVL72, and that Huawei has "already demonstrated silicon photonics" to create unified supercomputer architectures. The debate concludes with A's assertion that compute constraints actually drive innovation, as "the best AI researchers" develop "extremely smart algorithms" when hardware-limited.
**答：** Speaker A 认为中国的半导体限制被夸大了，指出 Huawei 在其历史最大年份出货了"数百万"芯片，远超 Anthropic 等公司的部署量。虽然 Speaker B 担心内存带宽瓶颈——指出 HBM2 与新一代内存可能有"近一个数量级的差异"——但 Speaker A 反驳说网络解决方案可以弥补，强调 Huawei 本质上是"一家网络公司"。A 明确否认先进内存需要 EUV 光刻技术，认为芯片可以像 NVL72 那样"组合在一起"，而且 Huawei "已经展示了 silicon photonics" 来创建统一的超算架构。争论最后 A 断言算力限制实际上推动创新，因为"最优秀的 AI 研究人员"在硬件受限时会开发出"极其聪明的算法"。

### Topic 45: Algorithm advances vs hardware advantages
算法进步与硬件优势的对比
_[01:09:33]_

**Q:** Are algorithmic improvements more important than raw hardware performance in AI development?
**问：** 在 AI 发展中，算法改进是否比硬件性能更重要？

**A:** The speaker argues that algorithmic innovation is the primary driver of AI progress, far outpacing hardware improvements. While Moore's law delivers roughly 25% annual performance gains, "great computer science" can achieve 10x improvements through algorithmic breakthroughs. The speaker emphasizes that "most of the advances in AI came out of algorithm advances, not just the raw hardware," citing examples like MoE (Mixture of Experts) and attention mechanisms that dramatically reduce computational requirements. This leads to a strategic conclusion: if algorithms matter more than hardware, then "their army of AI researchers is their fundamental advantage," referring to competitors' talent pools. The speaker uses DeepSeek as a concrete example of this dynamic, warning that if such algorithmic advances emerge on alternative hardware platforms like Huawei first, it represents "a horrible outcome for our nation," highlighting the geopolitical implications of algorithmic leadership.
**答：** 演讲者认为算法创新是 AI 进步的主要驱动力，远超硬件改进的贡献。虽然摩尔定律每年带来约 25% 的性能提升，但通过 "great computer science" 可以实现 10 倍的算法性能改进。演讲者强调 "most of the advances in AI came out of algorithm advances"，而不仅仅是硬件本身，并举例说明 MoE 和 attention 机制等创新大幅降低了计算需求。这引出一个战略性结论：如果算法比硬件更重要，那么竞争对手的 "army of AI researchers" 就是他们的根本优势所在。演讲者以 DeepSeek 为具体案例，警告如果这类算法突破首先出现在 Huawei 等替代硬件平台上，将是 "a horrible outcome for our nation"，凸显了算法领先地位的地缘政治意义。

### Topic 46: DeepSeek optimization and tech stack dominance
DeepSeek 优化与技术栈主导权
_[01:10:25]_

**Q:** Why does it matter which hardware platform AI models are optimized for?
**问：** AI 模型针对哪个硬件平台优化为什么重要？

**A:** Speaker A argues that hardware optimization creates strategic lock-in effects, warning that if DeepSeek or future models "run best on somebody else's tech stack," it would disadvantage American competitiveness globally. A points to "Nvidia's success" as evidence that models optimized for specific hardware create meaningful performance advantages, making it harder to switch accelerators despite theoretical portability. Speaker B challenges this premise by noting that American labs like Anthropic already "run on GPUs, they're run on Trainium, they're run on TPUs," suggesting cross-platform deployment is feasible with engineering effort. The core disagreement centers on whether optimization advantages are durable enough to matter strategically: A sees "coming out of the box" performance in the global south and Middle East as critical for ecosystem dominance, while B questions whether hardware platform matters if adversaries can still use American chips to develop superior models first.
**答：** Speaker A 认为硬件优化会产生战略性的锁定效应，警告如果 DeepSeek 或未来模型「在别人的技术栈上运行得最好」，会削弱美国的全球竞争力。A 以「Nvidia 的成功」为证据，说明针对特定硬件优化的模型会带来显著性能优势，尽管理论上可移植，但切换加速器很困难。Speaker B 质疑这个前提，指出像 Anthropic 这样的美国实验室已经「在 GPU、Trainium 和 TPU 上运行」，说明通过工程努力可以实现跨平台部署。核心分歧在于优化优势是否具有持久的战略意义：A 认为在全球南方和中东地区「开箱即用」的性能对生态系统主导地位至关重要，而 B 质疑如果对手仍然可以用美国芯片率先开发出更优模型，硬件平台是否真的重要。

### Topic 47: Chip substitutability and critical timeline debate
芯片可替代性和关键时间线争论
_[01:12:18]_

**Q:** Can Chinese chips substitute for American chips in the critical next few years?
**问：** 在关键的未来几年中，中国芯片能否替代美国芯片？

**A:** Speaker B challenges the assumption of perfect chip fungibility, arguing that Chinese chips like the Huawei 910C are measurably inferior to American hardware like the H200, operating at "half to a third" the performance in FLOP, bandwidth, and memory. While acknowledging China's manufacturing scale advantage and that they can compensate by using "twice as many" chips, B emphasizes that "these few critical years" before advanced AI models enable widespread cyber attacks represent a window where American technological superiority matters. Speaker A counters that China's chip industry is already "gigantic" with evidence of substitution happening now, and questions why one layer of the AI industry should "lose an entire market" to benefit another layer. The core disagreement centers on whether the performance gap creates meaningful strategic advantage during this critical period, or whether China's scale and energy infrastructure make the gap irrelevant.
**答：** Speaker B 质疑芯片完全可替代的假设，指出中国芯片如 Huawei 910C 在性能上明显落后于美国的 H200，在 FLOP、带宽和内存方面只有"一半到三分之一"的表现。虽然承认中国在制造规模上有优势，可以通过使用"两倍数量"的芯片来弥补，但 B 强调在先进 AI 模型能够发动大规模网络攻击之前的"这几年关键时期"，美国的技术优势至关重要。Speaker A 反驳说中国芯片产业已经"非常庞大"，有证据表明替代正在发生，并质疑为什么要让 AI 产业的一个层面"失去整个市场"来让另一个层面受益。核心分歧在于：性能差距是否在关键时期创造了实质性战略优势，还是中国的规模和能源基础设施使这个差距变得无关紧要。

### Topic 48: AI industry layers and market strategy
AI 行业分层与市场策略
_[01:13:44]_

**Q:** Why prioritize one layer of the AI industry over another in export policy?
**问：** 为什么在出口政策中要优先考虑 AI 行业的某一层级？

**A:** Speaker B argues against sacrificing entire markets to benefit a single layer of the AI industry, emphasizing that the industry consists of "five layers" that all need to succeed together. Rather than fixating on protecting one specific AI model or company, B contends that "the layer that has to succeed most is actually the AI applications" layer. This perspective challenges the logic of export restrictions that might protect model developers at the expense of broader market access. The speaker questions the rationale behind policy decisions that prioritize one company or model over the health of the entire AI ecosystem, suggesting a more holistic approach to industry strategy is needed.
**答：** Speaker B 反对为了某一层级的利益而牺牲整个市场，强调 AI 行业包含"五个层级"，所有层级都需要共同成功。他认为不应该只盯着保护某个特定的 AI model 或公司，而是应该关注"最需要成功的其实是 AI applications 这一层"。这个观点挑战了那些可能为了保护模型开发者而牺牲更广泛市场准入的出口限制政策。Speaker B 质疑为什么要把某一家公司或某个模型置于整个 AI 生态系统健康发展之上，主张需要采取更全面的行业策略。

### Topic 49: Jane Street backdoor detection puzzle and AI security verification
Jane Street 后门检测谜题与 AI 安全验证
_[01:14:18]_

**Q:** How did Jane Street's backdoor detection challenge work and what does it reveal about AI security verification?
**问：** Jane Street 的后门检测挑战如何运作，它揭示了 AI 安全验证的什么问题？

**A:** Jane Street invested approximately 20,000 GPU hours to train backdoors into three language models and challenged participants to discover the trigger phrases. Ricson, the puzzle designer, explained that one successful detection technique involved weight interpolation and extrapolation—by amplifying the difference between base and backdoored model weights, teams could make "the backdoor even stronger" until the model would "regurgitate what the response phrase was supposed to be." However, this weight extrapolation method only succeeded on two of the three models, and even Ricson couldn't explain why it failed on the third. This outcome underscores a fundamental challenge in AI security: "being able to verify that a model only does what you think it does is one of the most important open questions," as current detection techniques remain unreliable and incomplete even when researchers know backdoors exist.
**答：** Jane Street 投入约 20,000 GPU 小时在三个语言模型中植入后门，然后挑战参与者找出触发短语。设计者 Ricson 介绍了一种成功的检测技术：通过权重插值和外推，放大基础模型与后门模型之间的差异，可以让"后门变得更强"，直到模型"吐出预设的响应短语"。但这种权重外推方法只在三个模型中的两个上成功，第三个失败的原因连 Ricson 也无法解释。这个结果凸显了 AI 安全的根本挑战：即使研究人员明知后门存在，现有检测技术仍然不可靠且不完整，"验证模型是否只做你认为它会做的事"仍是最重要的开放性问题之一。

### Topic 50: China's 7nm chip production capacity and US technological lead
中国 7nm 芯片产能与美国算力优势
_[01:15:15]_

**Q:** Can China build enough 7nm capacity while the US moves to more advanced nodes, and should the US maintain its compute advantage?
**问：** 当美国推进到更先进制程时，中国能否建立足够的 7nm 产能？美国是否应该保持算力领先？

**A:** Speaker B argues that China will inevitably build sufficient 7nm capacity to compensate for restrictions, even as the US advances to "1.6nm with Feynman" while China remains on 7nm. However, B emphasizes that maintaining US technological leadership is paramount, noting that US compute capacity is already "100x more than anywhere else in the world." B outlines Nvidia's strategy of ensuring "US labs are the first to hear about it and have the first chance to buy it," and even investing in labs that lack funding. Speaker A challenges this position by questioning how shipping chips to China supports US leadership "if they're bottlenecked on compute," highlighting the tension between commercial interests and strategic advantage. The exchange reveals a fundamental disagreement about whether providing any compute to China undermines US dominance, with B asserting absolute commitment to keeping the US ahead while A questions the consistency of that policy.
**答：** Speaker B 认为，即使美国推进到 "1.6nm Feynman" 制程而中国仍停留在 7nm，中国也必然能建立足够的 7nm 产能来弥补限制带来的缺口。但 B 强调保持美国技术领先地位至关重要，指出美国的算力已经是 "世界其他地区的 100 倍"。B 介绍了 Nvidia 的策略：确保 "美国实验室最先得知并有优先购买权"，甚至对资金不足的实验室进行投资。Speaker A 则质疑这一立场，追问如果中国 "受限于算力"，那么向中国出口芯片如何能支持美国保持领先。这段对话揭示了一个核心矛盾：向中国提供任何算力是否会削弱美国优势。B 坚称会竭尽全力保持美国领先，而 A 则质疑这一政策的一致性。

### Topic 51: Debate on shipping chips to China and US competitiveness
向中国出口芯片与美国竞争力的争论
_[01:16:22]_

**Q:** Does shipping chips to China undermine US leadership, or does restricting exports harm American companies' global competitiveness?
**问：** 向中国出口芯片是削弱美国领导地位，还是限制出口会损害美国公司的全球竞争力？

**A:** Speaker B argues that overly restrictive export controls force American companies to "give up the world" and undermine US technology leadership rather than protect it. He frames Nvidia as part of "the American ecosystem" and "American technology leadership," questioning why policy would prevent a US company from winning in global markets. B advocates for "more balanced" regulations that allow Nvidia to compete internationally while maintaining US advantages like the Vera Rubin system domestically. Speaker A counters with a national security framing, citing Dario's analogy that selling advanced chips to China is like "Boeing bragging that we're selling North Korea nukes, but the missile casings are made by Boeing," suggesting the technology itself poses strategic risks regardless of who profits from the sale.
**答：** Speaker B 认为过于严格的出口管制会迫使美国公司"放弃全球市场"，反而削弱而非保护美国的技术领先地位。他将 Nvidia 定位为"美国生态系统"和"美国技术领导力"的一部分，质疑为什么政策要阻止美国公司在全球市场竞争。B 主张制定"更平衡"的监管政策，让 Nvidia 能在国际上竞争，同时在国内保持像 Vera Rubin 这样的优势。Speaker A 则从国家安全角度反驳，引用 Dario 的类比：向中国出售先进芯片就像"Boeing 吹嘘我们在向朝鲜卖核武器，但导弹外壳是 Boeing 造的"，暗示无论谁从销售中获利，技术本身都构成战略风险。

### Topic 52: AI-as-weapon analogy: enriched uranium vs compute
AI 武器类比：浓缩铀与算力
_[01:17:16]_

**Q:** Is comparing AI compute to enriched uranium a valid analogy, and can compute enable offensive cyber capabilities?
**问：** 将 AI 算力比作浓缩铀是否恰当，算力是否会促成网络攻击能力？

**A:** Speaker A invokes Dario's analogy comparing selling compute to adversaries as equivalent to "Boeing bragging that we're selling North Korea nukes," arguing that AI compute is "similar to enriched uranium" in having dual-use potential. Speaker B dismisses this as "a lousy analogy" and "illogical," but acknowledges the security concern by asking how compute running models capable of "zero-day exploits against all American software" wouldn't constitute a weapon. Speaker A's response pivots away from export controls, instead advocating for two strategies: first, establishing "dialogues with the researchers and dialogues with China" to prevent misuse through diplomatic engagement, and second, ensuring American technological superiority by making next-generation chips like "Vera Rubin, Blackwell" available domestically "in abundance, mountains of it." The exchange reveals a fundamental disagreement about whether compute restrictions are appropriate, with one side emphasizing containment through control and the other prioritizing openness coupled with overwhelming domestic advantage.
**答：** Speaker A 引用 Dario 的类比，认为向对手出售算力相当于 "Boeing 向朝鲜卖核弹还自夸"，并主张 AI 算力 "类似浓缩铀"，具有双重用途。Speaker B 认为这个类比 "很糟糕" 且 "不合逻辑"，但承认安全担忧，反问如果算力能运行可执行 "针对所有美国软件的零日漏洞攻击" 的模型，怎么不算武器。Speaker A 的回应转向了两个策略而非出口管制：首先通过 "与研究人员对话、与中国对话" 等外交手段防止技术滥用；其次确保美国技术优势，让 Vera Rubin、Blackwell 等下一代芯片在国内 "大量供应，堆积如山"。这段对话揭示了关于算力限制是否合适的根本分歧：一方强调通过管控来遏制，另一方主张开放加上压倒性的国内优势。

### Topic 53: Multi-layer AI industry and chip market strategy
多层 AI 产业与芯片市场策略
_[01:18:16]_

**Q:** Why does Nvidia argue that winning across all five layers of the AI stack requires competing globally in the chip market?
**问：** 为什么 Nvidia 认为在 AI 技术栈的五个层面获胜需要在全球芯片市场竞争？

**A:** The speaker argues that while the US should maintain leadership through abundant access to advanced chips like "Vera Rubin" and "Blackwell," AI competitiveness cannot be reduced to model development alone. They frame AI as a "five-layer cake" where the US must "win at every single layer, including the chip layer" to maintain long-term technological advantage. The core strategic concern is that "conceding the entire market" in chips would undermine US competitiveness in the foundational computing stack, regardless of strengths in AI research or model development. This perspective positions global chip market participation as essential infrastructure for sustained leadership across the full AI value chain, not just a commercial consideration.
**答：** 发言者认为，虽然美国应该通过充足的 Vera Rubin 和 Blackwell 等先进芯片供应来保持领先地位，但 AI 竞争力不能简化为模型开发。他们将 AI 比作"五层蛋糕"，美国必须"在包括芯片层在内的每一层都获胜"才能保持长期技术优势。核心战略关切在于，在芯片市场"让出整个市场"会削弱美国在基础计算栈的竞争力，无论在 AI 研究或模型开发上有多强。这一观点将全球芯片市场参与定位为维持整个 AI 价值链领导地位的必要基础设施，而非仅仅是商业考量。

### Topic 54: Tesla and iPhone precedent: does selling advanced tech create lock-in?
Tesla 和 iPhone 先例：出售先进技术是否产生锁定效应？
_[01:19:10]_

**Q:** Do examples like Tesla EVs and iPhones in China prove that selling advanced technology doesn't create lasting competitive advantage?
**问：** Tesla 电动车和 iPhone 在中国的例子是否证明出售先进技术不会产生持久竞争优势？

**A:** Speaker A argues that historical precedents like Tesla and iPhone sales to China demonstrate that selling advanced technology doesn't guarantee lock-in, as China now dominates EVs and smartphones despite accessing these products. However, A distinguishes Nvidia's position by emphasizing that "the richness of our ecosystem" centered on developers is fundamentally different—with 50% of AI developers in China, the U.S. risks losing this critical developer base through export restrictions. A counters concerns about competitive disadvantage by noting that American labs already use multiple accelerators beyond Nvidia, suggesting the market naturally supports diversity. The speaker rejects defeatist assumptions with a competitive stance: "You're not talking to somebody who woke up a loser," asserting that Nvidia's share is "growing, not decreasing" and that continued innovation, not market withdrawal, is the path forward.
**答：** Speaker A 用 Tesla 和 iPhone 在中国的案例说明，出售先进技术并不必然产生锁定效应——中国现在在电动车和智能手机领域占据主导地位。但他强调 Nvidia 的情况不同，核心在于"生态系统的丰富性"，特别是开发者群体——中国拥有 50% 的 AI 开发者，美国不应通过出口管制放弃这一关键资源。对于竞争劣势的担忧，A 指出美国实验室本身就在使用多种加速器，市场天然支持多样性。他以"你面对的不是一个天生的失败者"的姿态拒绝悲观假设，强调 Nvidia 的市场份额在增长而非下降，持续创新才是正道。

### Topic 55: Nvidia's ecosystem stickiness and developer lock-in
Nvidia 生态系统粘性与开发者锁定
_[01:20:12]_

**Q:** Is Nvidia's computing ecosystem fundamentally stickier than cars or phones, and does competing in China risk losing that advantage?
**问：** Nvidia 的计算生态系统是否比汽车或手机更具粘性，在中国竞争是否会失去这一优势？

**A:** The speaker rejects the premise that competing in China means inevitable market loss, calling it a "loser attitude" and "loser mindset" that contradicts American competitiveness. He draws a fundamental distinction between computing platforms and consumer products like cars, arguing that computing ecosystems have inherent switching costs that create stickiness. He points to historical precedents like "the x86 deal" and ARM's market position as evidence that "these ecosystems are hard to replace" because they require "an enormous amount of time and energy" to switch. Rather than conceding markets based on pessimistic assumptions, he frames Nvidia's strategy as continuing to "nurture that ecosystem" and "keep advancing the technology" to maintain competitive advantage through technical superiority and developer lock-in.
**答：** 发言人拒绝接受在中国竞争就意味着必然失去市场的前提，称这是一种"loser attitude"和"loser mindset"，与美国的竞争力相悖。他明确区分了计算平台和汽车等消费品的本质差异，认为计算生态系统具有内在的切换成本，形成了粘性。他以 x86 和 ARM 的市场地位为例，说明"这些生态系统很难被替代"，因为切换需要"enormous amount of time and energy"。他认为 Nvidia 的策略不是基于悲观假设放弃市场，而是通过持续培育生态系统和技术进步，利用技术优势和开发者锁定来保持竞争力。

### Topic 56: Marginal compute advantage vs offensive AI capabilities
边际算力优势与 AI 攻击能力
_[01:21:25]_

**Q:** Does any marginal increase in compute help train better models, and should that concern outweigh commercial benefits?
**问：** 算力的任何边际增长是否都有助于训练更好的模型，这种担忧是否应超过商业利益？

**A:** Speaker A argues that the concern isn't about a "key threshold of compute" but rather that "any marginal compute is helpful" for training better models, and therefore any chip sales to China incrementally strengthen their AI capabilities. Speaker B counters that if AI models gain "cyber offensive capabilities" like finding zero-days in software, the chips become enablers of "a weapon of a kind," making this fundamentally different from general-purpose technology. Speaker A pushes back by noting we already have extensive export controls on advanced DRAM manufacturing technology and other chip-making equipment, yet still sell CPUs and DRAM to China, suggesting a precedent for nuanced export policy. The core disagreement centers on whether AI represents a categorically different technology—one where marginal compute advantages in offensive capabilities justify stricter controls than traditional computing hardware, even at the cost of commercial benefits to the American technology industry.
**答：** Speaker A 认为问题的核心不在于某个算力门槛，而是「任何边际算力都有帮助」——更多算力就能训练更好的模型，因此任何对华芯片销售都会增强中国的 AI 能力。Speaker B 反驳说，如果 AI 模型具备「网络攻击能力」，比如能找到软件中的 zero-day 漏洞，那么芯片就成了「某种武器的使能者」，这与通用技术有本质区别。Speaker A 指出美国已经对先进 DRAM 制造技术和芯片制造设备实施了广泛的出口管制，但仍然向中国销售 CPU 和 DRAM，这说明可以有更细致的出口政策。双方分歧的核心在于：AI 是否是一种本质上不同的技术——在攻击能力方面，边际算力优势是否重要到需要比传统计算硬件更严格的管制，即使这会损害美国科技产业的商业利益。

### Topic 57: Export controls scope: AI chips vs CPUs, DRAM, and electricity
出口管制范围：AI 芯片与 CPU、DRAM 和电力
_[01:22:50]_

**Q:** Is AI fundamentally different from other technologies, or should export control logic apply equally to CPUs, DRAM, and electricity?
**问：** AI 是否与其他技术根本不同，还是出口管制逻辑应同样适用于 CPU、DRAM 和电力？

**A:** Speaker A argues AI chips warrant export controls similar to nuclear materials, drawing a parallel that "we have more nuclear weapons than anybody else, but we don't want to send enriched uranium anywhere." The core justification is strategic advantage: Chinese companies are "bottlenecked on compute" because "our chips are better," as evidenced by quotes from Chinese founders themselves. Speaker B challenges this analogy by emphasizing that chips are fundamentally different from enriched uranium since "it's a chip that they can make themselves," questioning whether the control logic should extend to general-purpose technologies. The debate centers on whether AI's potential capabilities—like finding zero-days in software—make it categorically different enough to justify controls that wouldn't apply to CPUs, DRAM, or electricity, with A emphasizing maintaining U.S. technological lead and B highlighting the practical limits of export restrictions on reproducible technology.
**答：** Speaker A 认为 AI 芯片应该像核材料一样实施出口管制，类比说"我们拥有最多核武器，但不会向任何地方输送浓缩铀"。核心理由是战略优势：中国公司"受限于算力"，因为"我们的芯片更好"，这一点从中国创始人的引述中得到证实。Speaker B 质疑这个类比，强调芯片与浓缩铀本质不同，因为"这是芯片，他们自己能造"，质疑管制逻辑是否应该延伸到通用技术。争论焦点在于 AI 的潜在能力（比如发现软件零日漏洞）是否让它在本质上足够特殊，从而合理化那些不适用于 CPU、DRAM 或电力的管制措施——A 强调保持美国技术领先，B 则指出对可复制技术实施出口限制的实际局限性。

### Topic 58: China's domestic chip industry growth and US market share loss
中国芯片产业崛起与美国市场份额流失
_[01:23:28]_

**Q:** Has restricting Nvidia's sales led to Huawei's record year and growth of Chinese chip companies, costing US market share?
**问：** 限制 Nvidia 销售是否导致 Huawei 创纪录业绩和中国芯片公司增长，使美国失去市场份额？

**A:** Speaker A argues that export restrictions have backfired by enabling domestic Chinese competitors, pointing out that "Huawei had a record year" and "a whole bunch of chip companies have gone public" while the US "used to have a very large share in that market" but no longer does. Speaker A emphasizes that Chinese companies were previously "bottlenecked on compute" and buying American chips "because our chips are better," suggesting restrictions removed a superior competitor from the market. Speaker B counters that China represents "about 40% of the world's technology industry" and that conceding this market would be "a disservice to our country" and "our national security," arguing the restrictions serve broader strategic interests beyond any single company's sales. The exchange reveals a fundamental tension between short-term market share loss and long-term strategic positioning in technology competition.
**答：** Speaker A 认为出口管制适得其反，反而扶持了中国本土竞争对手。他指出 "Huawei 创下历史最佳业绩"，"一大批芯片公司上市"，而美国在该市场的份额已经大幅下降。Speaker A 强调中国公司此前受限于算力，购买美国芯片是 "因为我们的芯片更好"，暗示管制反而让更优秀的竞争者退出了市场。Speaker B 反驳说中国占全球科技产业的 40%，放弃这个市场会损害美国国家安全和技术领先地位，认为限制措施服务于更广泛的战略利益，而非单一公司的销售业绩。这段对话揭示了短期市场份额损失与长期技术竞争战略定位之间的根本矛盾。

### Topic 59: Nvidia chips vs Huawei competition logic
Nvidia 芯片与华为竞争的逻辑
_[01:24:21]_

**Q:** How can Nvidia claim their chips are better while also saying China would develop AI without them anyway?
**问：** Nvidia 如何能既声称自己的芯片更好，又说中国无论如何都会发展 AI？

**A:** Speaker B argues there's no contradiction between Nvidia chips being superior and China developing AI regardless of access to them. The logic is straightforward: "in the absence of a better choice, you'll take the only choice you have." Nvidia's advantage comes from being "better" in multiple dimensions—more compute power for training better models, easier programmability, and a stronger ecosystem. Even if the U.S. restricts chip exports, China will pursue AI development with whatever alternatives are available, just with inferior tools. Speaker B concludes pragmatically that selling to China allows the U.S. to "benefit" economically while China would advance AI capabilities either way.
**答：** Speaker B 认为 Nvidia 芯片更优越和中国无论如何都会发展 AI 这两点并不矛盾。逻辑很简单："in the absence of a better choice, you'll take the only choice you have"（没有更好的选择时，你就用仅有的选择）。Nvidia 的优势是多维度的——更强的算力可以训练更好的模型、更易编程、生态系统更完善。即使美国限制芯片出口，中国也会用替代方案继续发展 AI，只是工具会差一些。Speaker B 务实地总结说，向中国销售芯片能让美国在经济上 "benefit"（获益），而中国无论如何都会推进 AI 能力。

### Topic 60: Benefits of American tech stack diffusion
美国技术栈全球扩散的战略价值
_[01:24:52]_

**Q:** What are the benefits to America of exporting compute and technology leadership globally?
**问：** 向全球输出算力和技术领导力对美国有什么好处？

**A:** Speaker B argues that America gains strategic advantages when its technology diffuses globally, specifically through "developers working on the American tech stack" and AI models spreading worldwide on American infrastructure. This creates a self-reinforcing cycle where "the American tech stack is therefore the best for it" as adoption increases. However, B warns that restrictive policies could backfire, pointing to how the telecommunications industry was "policied out of basically the world" through overregulation, resulting in America losing control of its own telecom infrastructure. B characterizes the opposing view as "narrow-minded" for failing to anticipate these "unintended consequences," suggesting that protectionist technology policies can paradoxically weaken rather than strengthen American technological dominance.
**答：** Speaker B 认为美国技术的全球扩散能带来战略优势，特别是通过让全球"开发者使用 American tech stack"以及 AI 模型在美国基础设施上传播。这形成了一个正向循环：随着采用率提升，"American tech stack 自然成为最优选择"。但 B 警告说限制性政策可能适得其反，他以电信行业为例——过度监管导致美国企业"被政策挤出全球市场"，最终美国连自己的电信基础设施都失去了控制权。B 批评对方观点"目光短浅"，没有预见到这些"意外后果"，暗示保护主义的技术政策反而会削弱而非增强美国的技术主导地位。

### Topic 61: Telecommunications policy failures as precedent
电信政策失败作为前车之鉴
_[01:25:40]_

**Q:** How did restrictive policies harm American telecommunications industry leadership?
**问：** 限制性政策如何损害了美国电信行业的领导地位？

**A:** Speaker B argues that restrictive policies were "narrow-minded" and produced "unintended consequences" that undermined U.S. telecommunications competitiveness, though the specific mechanisms aren't detailed in this brief exchange. Speaker A attempts to reframe the debate around cost-benefit analysis, seeking acknowledgment of potential downsides to the policies being discussed. The tension reveals a disagreement about whether the other party is fully grasping the historical lesson: that well-intentioned restrictions can backfire strategically. Speaker B's frustration suggests they view the telecommunications precedent as a clear cautionary tale that should inform current policy debates, while Speaker A pushes for explicit articulation of tradeoffs rather than assumed conclusions.
**答：** Speaker B认为限制性政策"narrow-minded"（目光短浅），导致了"unintended consequences"（意料之外的后果），削弱了美国电信行业的竞争力，但这段对话中没有展开具体机制。Speaker A试图将讨论重新聚焦到成本收益分析上，希望对方承认政策可能带来的负面影响。这种张力反映出双方对历史教训理解的分歧：善意的限制措施可能在战略上适得其反。Speaker B的挫败感表明，他认为电信行业的前车之鉴应该为当前政策辩论提供明确警示，而Speaker A则坚持要求明确阐述权衡取舍，而非预设结论。

### Topic 62: Costs of China accessing advanced compute
中国获得先进算力的成本
_[01:25:48]_

**Q:** What are the security risks if China had developed Mythos-level AI capabilities earlier with more compute?
**问：** 如果中国更早获得算力并开发出 Mythos 级别的 AI 能力，会带来什么安全风险？

**A:** Speaker A frames compute access as a direct security tradeoff: powerful models enable "powerful offensive capabilities, like cyber attacks," and it was "a good thing that American companies got to Mythos-level capabilities first" because they could then hold back deployment while hardening defenses. The core argument is that if China had accessed more compute or cloud resources earlier, they could have "made a Mythos-level model earlier and deployed it widely," which "would have been very bad" from a security standpoint. Speaker A attributes the current advantage partly to American compute leadership "thanks to companies like Nvidia" and frames sending compute to China as accepting this security cost. Speaker B counters by highlighting a different cost: export controls risk allowing "the chip layer" to "concede an entire market—the second largest market in the world," enabling China to build independent scale and ecosystems where "future AI models are optimized in a very different way than the American tech stack."
**答：** Speaker A 认为算力获取是一个直接的安全权衡：强大模型会带来"网络攻击等强大的进攻能力"，而"美国公司率先达到 Mythos 级别的能力是件好事"，因为他们可以在公开前加固防御。核心论点是，如果中国更早获得更多算力或云计算资源，就可能"更早开发出 Mythos 级别的模型并广泛部署"，这"会非常糟糕"。Speaker A 将当前优势部分归功于"像 Nvidia 这样的公司"带来的美国算力领先地位，并认为向中国输送算力就是接受这种安全成本。Speaker B 则提出另一种成本：出口管制可能让"芯片层""放弃整个市场——全球第二大市场"，使中国建立独立的规模和生态系统，导致"未来的 AI 模型以与美国技术栈非常不同的方式优化"。

### Topic 63: Risk of conceding chip market to China
向中国让出芯片市场的风险
_[01:26:45]_

**Q:** What happens if export controls allow China to develop their own chip ecosystem and AI stack at scale?
**问：** 如果出口管制让中国大规模发展自己的芯片生态系统和 AI 技术栈会怎样？

**A:** Speaker B warns that export controls create a strategic risk by forcing China to develop an independent chip ecosystem in "the second largest market in the world," giving them the scale needed to build a fundamentally different AI stack. This could lead to "future AI models" being "optimized in a very different way than the American tech stack," with Chinese standards potentially becoming superior as "AI diffuses out into the rest of the world" due to their open model approach. Speaker A counters with confidence in American technical capabilities, citing "Nvidia's kernel engineers and CUDA engineers" and techniques like model distillation to optimize for specific chips. However, Speaker B pushes back that "AI is more than kernel optimization," suggesting the competitive threat extends beyond low-level technical performance to encompass broader ecosystem and standards competition.
**答：** Speaker B 警告说，出口管制会带来战略风险，因为这会迫使中国在"全球第二大市场"独立发展芯片生态系统，获得足够规模来构建完全不同的 AI 技术栈。这可能导致"未来的 AI 模型"以"与美国技术栈非常不同的方式进行优化"，随着"AI 扩散到世界其他地区"，中国的标准可能因为开放模型策略而变得更具优势。Speaker A 则对美国技术能力有信心，提到"Nvidia 的 kernel 工程师和 CUDA 工程师"以及模型蒸馏等技术可以针对特定芯片优化。但 Speaker B 反驳说"AI 不仅仅是 kernel 优化"，暗示竞争威胁不只是底层技术性能，而是涉及更广泛的生态系统和标准之争。

### Topic 64: China's open source dominance and tech stack competition
中国的开源主导地位与AI技术栈竞争
_[01:27:29]_

**Q:** Why does China's leadership in open source software and models matter for the five layers of AI stack competition?
**问：** 中国在开源软件和模型方面的领先地位对AI技术栈五层竞争有何影响？

**A:** Speaker A argues that China's position as "the largest contributor to open source software" and "open models in the world" represents a strategic advantage that the U.S. must take seriously across all five layers of the AI technology stack. While acknowledging that China's current ecosystem is "built on the American tech stack, Nvidia's," the speaker emphasizes that "all five layers of the tech stack for AI are important" and the United States should compete to "win all five of them." The speaker identifies the AI application layer as "the most important" because it's "the layer that diffuses into society" and determines which country will "benefit from this industrial revolution most." The argument concludes with a warning that treating AI like "a nuclear bomb" through fear-mongering does the United States "a disservice" by undermining its competitive position rather than strengthening it.
**答：** Speaker A强调中国作为全球最大的开源软件和开源模型贡献者，这一事实对美国构成了战略挑战。虽然目前中国的AI生态系统仍然建立在美国的技术栈（特别是Nvidia）之上，但他认为AI技术栈的五层都很重要，美国必须在每一层都保持竞争力。其中最关键的是应用层，因为这是真正渗透到社会中的层面，谁在这一层占据优势，谁就能从这场工业革命中获益最多。他警告说，如果把AI当作核弹一样对待，用恐惧来限制发展，反而会削弱美国的竞争地位。

### Topic 65: Dangers of AI fear-mongering
AI 恐慌论的危害
_[01:28:27]_

**Q:** How does treating AI like a nuclear bomb or scaring people about job displacement harm American competitiveness?
**问：** 把 AI 当作核弹或吓唬人们会失业如何损害美国竞争力？

**A:** The speaker argues that extreme AI narratives—whether treating it "like a nuclear bomb" or claiming it will eliminate entire professions—actively harm American competitiveness by discouraging talent from entering critical fields. He distinguishes between tasks and jobs, using radiology as a concrete example: while AI may automate "the task" of reading scans, "the job of a radiologist is patient care," meaning the profession remains essential even as specific tasks evolve. If fear-mongering drives people away from software engineering or radiology training, the U.S. will face talent shortages that weaken both its technological capabilities and healthcare system. The core problem is that "extreme" premises—whether predicting "zero or infinity"—distort reality and create self-fulfilling prophecies where fear itself becomes the obstacle, not the technology.
**答：** 讲者认为极端的 AI 叙事——无论是把它"当作核弹"还是声称它会消灭整个职业——会阻碍人才进入关键领域，从而实际损害美国竞争力。他区分了任务和工作：以放射科为例，虽然 AI 可能自动化"读片这个任务"，但"放射科医生的工作是病人护理"，这意味着即使具体任务演变，这个职业仍然不可或缺。如果恐慌论把人们吓得不敢学软件工程或放射科，美国将面临人才短缺，削弱技术能力和医疗系统。核心问题在于"极端"的前提——无论预测"归零还是无限"——都扭曲了现实，制造出自我实现的预言：恐惧本身成了障碍，而非技术。

### Topic 66: Radiologist example: jobs vs tasks
放射科医生案例：工作与任务的区别
_[01:29:40]_

**Q:** What's the difference between a radiologist's job (patient care) and task (reading scans) in the AI era?
**问：** 在 AI 时代，放射科医生的工作（患者护理）和任务（读片）有什么区别？

**A:** The speaker argues against binary thinking that creates unnecessary fear, emphasizing that reality operates in gradients rather than extremes "from zero or infinity." He asserts that American technological leadership requires dominance across multiple layers of the technology stack, not just isolated components. The discussion pivots to a forward-looking prediction about technology export, where the speaker envisions American technology, standards, and platforms like Mythos being "diffused around the world" to regions including India, the Middle East, Africa, and Southeast Asia. He frames this as a strategic imperative for maintaining influence, suggesting that today's technological investments will determine tomorrow's ability to shape global standards and markets.
**答：** 讲者反对非黑即白的思维方式，认为这种"从零到无穷"的极端化会制造不必要的恐慌，而现实是渐进式的。他强调美国要保持科技领先地位，需要在技术栈的每一层都具备领导力，而不是只在某个孤立环节占优。话题随后转向技术输出的前瞻性预测：讲者设想美国的技术、标准和像 Mythos 这样的平台将"扩散到全世界"，包括印度、中东、非洲和东南亚等地区。他将此视为维持影响力的战略要务，暗示今天的技术投资将决定未来塑造全球标准和市场的能力。

### Topic 67: Winning all five layers of AI stack
赢得 AI 技术栈的全部五层
_[01:30:07]_

**Q:** Why must the US lead in all five layers of the AI stack and export technology globally rather than conceding markets?
**问：** 为什么美国必须在 AI 技术栈的全部五层领先并向全球出口技术，而不是放弃市场？

**A:** The speaker argues that the US is unnecessarily "conceding the second largest market in the world for no good reason" through overly restrictive export policies, which will undermine America's ability to diffuse its technology and standards globally. He emphasizes a critical distinction: "nobody is advocating all or nothing," meaning the US can maintain technological leadership at home while still competing internationally. The position calls for "some amount of nuance, some amount of maturity instead of absolutes"—the US should "always have the best technology here" and "the most technology here, and the first," but simultaneously "try to compete and win around the world." He warns that current policies will have consequences when America wants to export its tech stack to India, the Middle East, Africa, and Southeast Asia, suggesting that premature market concession today will cost strategic influence tomorrow.
**答：** 演讲者认为，美国正在通过过度限制性的出口政策"毫无理由地放弃世界第二大市场"，这将削弱美国向全球推广其技术和标准的能力。他强调一个关键区别："没有人主张全有或全无"，意思是美国可以在保持国内技术领先的同时参与国际竞争。他的立场呼吁"一定程度的细致和成熟，而不是绝对化"——美国应该"始终拥有最好的技术"和"最多的技术，并且是最先的"，但同时也要"努力在全球范围内竞争并获胜"。他警告说，当美国想要向印度、中东、非洲和东南亚出口其技术栈时，当前的政策将产生后果，暗示今天过早放弃市场将在明天付出战略影响力的代价。

### Topic 68: Nuance vs absolutes in export policy
出口管制政策需要细致权衡而非绝对化
_[01:31:22]_

**Q:** Why is a nuanced approach better than absolute all-or-nothing export controls?
**问：** 为什么出口管制需要细致权衡的方法，而不是全有或全无的绝对化政策？

**A:** Speaker A argues that effective export policy requires rejecting binary thinking in favor of "some amount of nuance, some amount of maturity instead of absolutes." The core insight is that multiple competing objectives can be balanced simultaneously rather than forcing an either-or choice. This perspective acknowledges that "the world is just not absolutes," suggesting that real-world policy challenges involve tradeoffs that cannot be resolved through categorical rules. Speaker B's response hints at the practical tension: adversaries are already designing around current restrictions by building models "specified for the best chips that they make in a few years," which implies that overly rigid controls may be circumvented while more calibrated approaches could better address evolving threats.
**答：** Speaker A 认为有效的出口政策需要摒弃非黑即白的思维，采用"一定程度的细致权衡和成熟度，而不是绝对化"的方式。核心观点是多个相互竞争的目标可以同时平衡，而不必强制做出非此即彼的选择。这种视角承认"世界本身就不是绝对的"，表明现实世界的政策挑战涉及无法通过绝对规则解决的权衡取舍。Speaker B 的回应暗示了实际困境：对手已经在针对当前限制进行设计，为"未来几年他们制造的最好芯片"构建模型，这意味着过于僵化的管制可能被规避，而更精细的方法可能更好地应对不断演变的威胁。

### Topic 69: 7nm vs 1.6nm chip competitiveness
7nm 与 1.6nm 芯片的竞争力对比
_[01:31:34]_

**Q:** Can Chinese 7nm chips and optimized models compete against American 1.6nm chips in global exports?
**问：** 中国的 7nm 芯片和优化模型能否在全球出口中与美国的 1.6nm 芯片竞争？

**A:** Speaker B argues that due to EUV export controls, China will remain on 7nm while the US advances to 1.6nm, creating a fundamental export competitiveness challenge. While domestically China might leverage abundant energy and manufacturing scale to continue using 7nm chips, the global export market requires their chips and models to compete directly against more advanced American lithography. The critical test is whether Chinese models can be "so far optimized" for 7nm that they outperform American models running on 1.6nm chips. However, Speaker A challenges the premise by pointing out that lithographic advancement between generations is not as dramatic as assumed—Blackwell is "not even close" to 50 times more advanced than Hopper, with transistor improvements being only around 75%, suggesting "Moore's Law is dead." This implies the competitiveness gap may be narrower than the nanometer numbers suggest, potentially making optimization-driven competition more viable.
**答：** Speaker B 认为由于 EUV 出口管制，中国将停留在 7nm 而美国推进到 1.6nm，这造成了根本性的出口竞争挑战。虽然在国内市场，中国可以利用充足的能源和规模化制造继续使用 7nm 芯片，但在全球出口市场上，他们的芯片和模型必须直接与更先进的美国光刻技术竞争。关键问题是中国的模型能否针对 7nm "优化到极致"，以至于在 7nm 上的表现超过美国模型在 1.6nm 芯片上的表现。不过 Speaker A 质疑了这个前提，指出代际之间的光刻进步并没有想象中那么大——Blackwell 相比 Hopper "远不到" 50 倍的进步，晶体管改进只有约 75%，表明 "Moore's Law is dead"。这意味着竞争力差距可能比纳米数字显示的要小，通过优化驱动的竞争可能更具可行性。

### Topic 70: Architecture matters more than lithography
架构比光刻更重要
_[01:32:24]_

**Q:** Why is Blackwell 50x faster than Hopper despite only 75% transistor improvement, and why does architecture matter more than process node?
**问：** 为什么 Blackwell 比 Hopper 快 50 倍，尽管晶体管只改进了 75%，为什么架构比制程节点更重要？

**A:** The speaker argues that Moore's Law is effectively dead by pointing to the stark gap between transistor-level gains and system-level performance: Blackwell achieved "50 times" the performance of Hopper despite only a "75%" improvement in transistors over three years. This dramatic multiplier demonstrates that "architecture matters" and "computer science matters" as much as semiconductor physics. The real leverage comes from "the computing stack" and software ecosystems like CUDA, which the speaker describes as "so effective" and "so beloved" because it provides the flexibility to implement radically different architectures—whether MoE, diffusion models, or disaggregated systems—without friction. The implication is that AI progress depends on co-optimizing hardware architecture with software infrastructure, not just shrinking transistors.
**答：** 演讲者通过一个鲜明对比来论证 Moore's Law 已经失效：Blackwell 的性能是 Hopper 的 50 倍，但晶体管层面三年内只提升了 75%。这种巨大的性能倍增说明"架构比光刻更重要"，计算机科学和半导体物理同样关键。真正的杠杆来自"计算栈"和 CUDA 这样的软件生态系统，它之所以"如此有效"且"备受喜爱"，是因为提供了灵活性——无论是 MoE、diffusion 模型还是分布式架构都能轻松实现。核心观点是 AI 进步依赖硬件架构与软件基础设施的协同优化，而不仅仅是缩小晶体管。

### Topic 71: CUDA ecosystem advantages
CUDA 生态系统的优势
_[01:33:29]_

**Q:** Why do researchers globally, including in China, prefer programming on CUDA first?
**问：** 为什么全球研究人员，包括中国的研究人员，都更喜欢首先在 CUDA 上编程？

**A:** The speaker argues that AI success depends on "the stack above as much as it is about the architecture below," emphasizing that software ecosystems matter as much as hardware. Nvidia's competitive advantage stems from having "architectures and software stacks that are optimized" specifically for their ecosystem. The speaker observes that researchers universally, including those in China, "always love programming CUDA first" as a testament to the platform's richness. This preference reflects the depth of Nvidia's ecosystem rather than just hardware performance, creating a powerful network effect that makes CUDA the default choice for AI development globally.
**答：** 发言人认为 AI 的成功取决于"上层软件栈和下层架构同等重要"，强调软件生态系统与硬件同样关键。Nvidia 的竞争优势在于拥有"专门为其生态系统优化的架构和软件栈"。发言人观察到，包括中国研究人员在内的全球研究者都"总是喜欢首先在 CUDA 上编程"，这证明了该平台的丰富性。这种偏好反映的是 Nvidia 生态系统的深度，而不仅仅是硬件性能，形成了强大的网络效应，使 CUDA 成为全球 AI 开发的默认选择。

### Topic 72: Consequences of forcing Nvidia out of China
强迫 Nvidia 退出中国的后果
_[01:33:58]_

**Q:** How have export controls backfired by accelerating China's domestic chip industry and forcing their AI ecosystem away from American architectures?
**问：** 出口管制如何适得其反，加速了中国国内芯片产业并迫使其 AI 生态系统远离美国架构？

**A:** The speaker argues that forcing Nvidia out of China was "a policy mistake" that has "turned out badly for the United States" by producing unintended consequences. Rather than containing China's chip capabilities, the export controls "accelerated their chip industry" and "forced all of their AI ecosystem to focus on their internal architectures," creating a self-sufficient competitor. The speaker dismisses the notion that China is permanently stuck at 7nm nodes, noting "they're good at manufacturing" and will continue advancing beyond current limitations. He emphasizes that raw process node advantage is overrated—"is there a 10x difference between 5nm and 7nm? The answer is no"—because other factors like architecture, networking (citing Nvidia's Mellanox acquisition), and energy efficiency matter more. The speaker warns against oversimplifying the competitive landscape, suggesting the damage has already been done though "it's not too late" to reverse course.
**答：** 发言人认为强迫 Nvidia 退出中国是"政策失误"，给美国带来了"糟糕的结果"和意外后果。出口管制非但没有遏制中国的芯片能力，反而"加速了他们的芯片产业"，并"迫使他们整个 AI 生态系统专注于内部架构"，培养出了一个自给自足的竞争对手。发言人驳斥了中国会永久停留在 7nm 节点的观点，指出"他们擅长制造"，会继续突破当前限制。他强调单纯的制程节点优势被高估了——"5nm 和 7nm 之间有 10 倍差距吗？答案是没有"——因为架构、网络（提到 Nvidia 收购 Mellanox）和能效等其他因素更重要。发言人警告不要过度简化竞争格局，暗示损害已经造成，尽管"还不算太晚"可以扭转局面。

### Topic 73: Using older process nodes for AI chips
使用旧制程节点生产 AI 芯片
_[01:35:06]_

**Q:** Could Nvidia go back to N7 process to meet AI demand if leading-edge capacity is constrained?
**问：** 如果前沿产能受限，Nvidia 能否回到 N7 制程来满足 AI 需求？

**A:** The speaker argues that returning to older process nodes like 7nm is economically impractical because modern chip generations involve far more than transistor scaling—they integrate advances in "packaging and stacking" and "numerics and the system architecture" that would require prohibitive R&D investment to backport. The company "could afford to lean forward" with new generations but "couldn't afford to go back" to redesign older nodes with current knowledge. However, in a hypothetical scenario where leading-edge capacity permanently stops expanding, the speaker acknowledges they would use 7nm "in a heartbeat" as a last resort. The core tension is between the theoretical feasibility of using older nodes and the practical reality that each generation's architectural innovations are deeply coupled to its process technology, making backward adaptation a massive engineering undertaking that doesn't make business sense unless all other options are exhausted.
**答：** 嘉宾认为回到 7nm 这样的旧制程节点在经济上不可行，因为现代芯片的每一代不仅仅是晶体管缩放，还涉及「封装和堆叠」以及「数值计算和系统架构」的进步，要把这些技术倒推回旧节点需要巨额研发投入。公司「有能力向前推进」新一代产品，但「没能力往回走」去重新设计旧节点。不过，在一个假设场景中——如果前沿产能永久停止扩张，嘉宾承认会「毫不犹豫」使用 7nm 作为最后手段。核心矛盾在于：理论上可以使用旧节点，但实际上每一代的架构创新都与其制程技术深度耦合，向后适配是一项巨大的工程任务，除非所有其他选项都用尽，否则在商业上说不通。

### Topic 74: Why Nvidia doesn't pursue multiple parallel architectures
为什么 Nvidia 不同时开发多种架构
_[01:36:42]_

**Q:** Why doesn't Nvidia run multiple different chip projects simultaneously like wafer-scale or alternative architectures?
**问：** 为什么 Nvidia 不同时运行多个不同的芯片项目，比如 Cerebras 的晶圆级架构或 Tesla Dojo 那样的大封装方案？

**A:** Nvidia's decision not to pursue multiple parallel chip architectures simultaneously isn't driven by resource constraints—the speaker acknowledges "we could" do all of these approaches including wafer-scale designs like Cerebras or massive packages like Dojo. The fundamental constraint is strategic conviction rather than capability: they "don't have a better idea" than their current architectural direction. This reveals a philosophy of focused execution where having the engineering talent and resources to explore alternatives is insufficient justification without a compelling technical or market thesis for why those alternatives would be superior. The brevity and confidence of the response suggests Nvidia views architectural diversification as a hedge against uncertainty rather than a path to innovation, and they're betting their current approach is the right one.
**答：** Nvidia 不同时开发多种芯片架构，并不是因为资源不足——发言人承认他们完全有能力做 Cerebras 那样的晶圆级设计或 Dojo 那样的大封装方案。真正的限制是战略信念而非技术能力：他们目前「没有更好的想法」。这反映出一种专注执行的哲学——即使拥有工程人才和资源去探索其他方向，如果没有令人信服的技术或市场理由证明那些替代方案更优，就不足以成为行动的理由。这个简短而自信的回答表明，Nvidia 将架构多元化视为对不确定性的对冲，而非创新路径，他们坚信当前的方向是正确的。

### Topic 75: Hardware acceleration strategy and simulator validation
硬件加速策略与模拟器验证
_[01:37:14]_

**Q:** Why doesn't Nvidia pursue certain hardware alternatives?
**问：** 为什么Nvidia不追求某些硬件替代方案？

**A:** Nvidia uses rigorous simulation to evaluate hardware alternatives before committing resources, and many proposed approaches are "provably worse" when tested in their simulator. The company maintains a disciplined focus on projects that demonstrate clear advantages, avoiding speculative hardware development that doesn't show measurable improvement. Their strategy is deliberately selective—they're "working on exactly the projects that we want to work on"—rather than pursuing every possible architectural variation. However, they remain open to adding "other accelerators" if the workload characteristics change dramatically, though this depends on fundamental shifts in market shape rather than just algorithmic evolution.
**答：** Nvidia通过严格的模拟器测试来评估硬件替代方案，许多提议的方法在模拟器中被证明"provably worse"（可证明更差）。公司保持专注，只投入那些展现出明确优势的项目，避免没有可衡量改进的投机性硬件开发。他们的策略是有意识地精选——"working on exactly the projects that we want to work on"——而不是追逐每一个可能的架构变体。不过，如果工作负载特征发生根本性变化，他们仍然愿意添加"other accelerators"（其他加速器），但这取决于市场形态的根本转变，而非仅仅是算法层面的演进。

### Topic 76: Adding Groq accelerators based on workload changes
根据工作负载变化引入Groq加速器
_[01:37:32]_

**Q:** Under what conditions would Nvidia add other accelerators to their ecosystem?
**问：** Nvidia在什么情况下会在生态系统中引入其他加速器？

**A:** The speaker explains that Nvidia's decision to incorporate additional accelerators hinges on dramatic shifts in workload characteristics rather than algorithmic changes, with the distinction being driven by "the shape of the market." The recent addition of Groq to the CUDA ecosystem serves as a concrete example of this strategic flexibility, motivated by a fundamental economic shift where "the value of tokens has gone up so high." This increased token value creates opportunities for differentiated pricing models that justify integrating specialized accelerators alongside Nvidia's core offerings. The decision reflects a pragmatic approach where market economics and workload patterns—not just technical capabilities—determine ecosystem expansion.
**答：** Nvidia决定引入其他加速器的关键在于工作负载特征的显著变化，而非算法本身的改变，这种决策取决于"市场的形态"。最近将Groq整合进CUDA生态系统就是一个具体案例，其背后驱动力是token价值的大幅提升。当"token的价值变得如此之高"时，就为差异化定价创造了空间，这使得引入专用加速器变得合理。这一决策体现了务实的策略：市场经济因素和工作负载模式——而不仅仅是技术能力——决定了生态系统的扩展方向。

### Topic 77: Token economics and market segmentation
Token经济学与市场细分
_[01:38:00]_

**Q:** How has the value of tokens changed and what new market segments has this created?
**问：** Token的价值如何变化，这创造了哪些新的市场细分？

**A:** The speaker explains that token economics have fundamentally shifted as tokens have become significantly more expensive compared to "just a couple years ago" when they were "either free or barely expensive." This price increase has enabled differentiated pricing strategies where different customer segments are willing to pay premium prices for enhanced performance characteristics. The speaker uses software engineers as a concrete example, noting that because "customers make so much money," they would pay more for "much more responsive tokens" that increase productivity. This emerging market dynamic has led to a strategic decision to "expand the Pareto frontier" by creating inference segments optimized for faster response times rather than maximum throughput, even though this represents a tradeoff. The key insight is that higher token values have unlocked the ability to segment "the same model, based on the response time" into different market tiers that didn't exist before.
**答：** 讲者指出，Token经济学发生了根本性转变，因为Token的价格相比几年前已经大幅上涨，那时Token "要么免费，要么几乎不值钱"。价格上涨使得差异化定价策略成为可能，不同客户群体愿意为更好的性能特征支付溢价。讲者以软件工程师为例，指出由于这些客户"赚很多钱"，他们愿意为"响应更快的Token"付费以提高生产力。这种新兴的市场动态促使他们决定"扩展Pareto前沿"，创建针对更快响应时间而非最大吞吐量优化的推理服务细分市场，尽管这涉及权衡取舍。核心洞察是，更高的Token价值使得基于响应时间对"同一个模型"进行市场细分成为可能，而这种细分市场此前并不存在。

### Topic 78: Response time vs throughput tradeoffs in inference
推理中的响应时间与吞吐量权衡
_[01:38:46]_

**Q:** Why did Nvidia expand the Pareto frontier to include faster response time with lower throughput?
**问：** 为什么Nvidia扩展帕累托前沿以包含更快响应时间但更低吞吐量的选项？

**A:** Nvidia deliberately expanded the Pareto frontier to create "a segment of inference that is faster response time, even though it's lower throughput," breaking from the traditional assumption that "higher throughput is always better." The strategic rationale centers on the emergence of "very high ASP tokens" where the average selling price per token is sufficiently premium that lower factory throughput becomes economically viable. This represents a fundamental shift in inference economics: when tokens command high enough prices, optimizing for response speed rather than raw throughput can still deliver strong unit economics because "the ASPs make up for it." The speaker views this as a confident architectural bet, noting that "if I had more money, I would put more behind Nvidia's architecture," suggesting this tradeoff opens a defensible market segment focused on premium, latency-sensitive inference workloads.
**答：** Nvidia有意扩展了帕累托前沿，创造了一个"响应时间更快但吞吐量更低"的推理细分市场，打破了"更高吞吐量总是更好"的传统假设。这一战略的核心逻辑在于"very high ASP tokens"的出现——当单个token的平均售价足够高时，即使工厂吞吐量较低也能在经济上可行，因为"ASPs make up for it"（高价格弥补了吞吐量损失）。这代表了推理经济学的根本转变：当token价格足够高时，优化响应速度而非原始吞吐量仍能带来良好的单位经济效益。说话者对这一架构选择充满信心，表示"如果有更多资金，会投入更多到Nvidia的架构上"，暗示这种权衡开辟了一个专注于高端、延迟敏感推理工作负载的可防御市场细分。

### Topic 79: Premium token pricing and inference market disaggregation
高端Token定价与推理市场细分
_[01:39:21]_

**Q:** What is the significance of premium tokens and market segmentation in the AI inference market?
**问：** 高端Token和市场细分对AI推理市场有什么意义？

**A:** Speaker B identifies two emerging trends in the AI inference market that they find particularly compelling. The first is the concept of "extremely premium tokens," suggesting a pricing model where certain inference requests or token generations command significantly higher prices, likely due to quality, capability, or computational intensity. The second trend is "the disaggregation of the inference market," which Speaker A clarifies as "segmentation," indicating a shift away from one-size-fits-all inference services toward specialized market segments. This segmentation likely reflects differentiation based on use cases, performance tiers, latency requirements, or model capabilities. The brief exchange suggests both speakers see this market evolution as a significant development worth noting, though the conversation pivots to a different topic before deeper exploration.
**答：** Speaker B指出了AI推理市场中两个值得关注的新趋势。第一个是"extremely premium tokens"的概念，暗示某些推理请求或token生成会有显著更高的定价，可能是因为质量、能力或计算强度的差异。第二个趋势是推理市场的"disaggregation"，Speaker A将其明确为"segmentation"（细分），表明市场正在从统一的推理服务转向专业化的细分市场。这种细分可能基于使用场景、性能层级、延迟要求或模型能力等维度进行区分。这段简短的交流显示两位speaker都认为这种市场演变很有意义，但话题很快转向了其他方向。

### Topic 80: Nvidia's mission without deep learning
没有深度学习的Nvidia使命
_[01:39:34]_

**Q:** What would Nvidia be doing if the deep learning revolution hadn't happen?
**问：** 如果深度学习革命没有发生，Nvidia会做什么？

**A:** The speaker argues that Nvidia would still be pursuing "accelerated computing, the same thing we've been doing all along," suggesting deep learning was an application rather than the core mission. The company's foundational premise is that "general purpose computing" has limitations and that combining GPU architecture with CUDA and CPU enables offloading "different kernels of code or algorithms" to achieve dramatic speedups of "100x, 200x." This acceleration model applies broadly across "engineering and science and physics, data processing, computer graphics, image generation, all kinds of things," indicating Nvidia's strategy was always about specialized compute acceleration rather than being dependent on any single breakthrough like deep learning.
**答：** 发言人认为Nvidia仍会专注于"accelerated computing"，也就是"我们一直在做的事情"，这表明深度学习只是应用场景而非核心使命。公司的基本理念是通用计算存在局限性，通过将GPU架构、CUDA与CPU结合，可以将"不同的代码内核或算法"卸载到GPU上，实现"100倍、200倍"的惊人加速。这种加速模型广泛适用于"工程、科学、物理、数据处理、计算机图形、图像生成等各种领域"，说明Nvidia的战略始终围绕专用计算加速，而不依赖于深度学习这样的单一突破。

### Topic 81: Accelerated computing beyond AI applications
AI之外的加速计算应用
_[01:40:24]_

**Q:** What non-AI domains benefit from Nvidia's accelerated computing?
**问：** 哪些非AI领域受益于Nvidia的加速计算？

**A:** The speaker argues that Nvidia would remain "very, very large" even without AI because general purpose computing has "largely run its course" in terms of scalability, making domain-specific acceleration essential. Starting with computer graphics, Nvidia's CUDA platform now serves diverse fields including particle physics, fluids simulation, and structured data processing. The company's core mission has been to "advance the type of applications that general purpose computing can't do" and enable breakthroughs in scientific fields where traditional computing is "simply too inefficient." Early accelerated computing applications included molecular dynamics, seismic processing for energy discovery, and image processing—all domains where the performance gap between general purpose and specialized computing creates transformative value.
**答：** 演讲者认为即使没有AI，Nvidia依然会是一家非常大的公司，因为通用计算的扩展能力已经"基本走到尽头"，领域专用加速成为必然选择。从计算机图形学起步，Nvidia的CUDA平台现在服务于粒子物理、流体模拟、结构化数据处理等多个领域。公司的核心使命是推进"通用计算无法完成的应用类型"，在传统计算"效率过低"的科学领域实现突破。早期的加速计算应用包括分子动力学、能源勘探中的地震处理、图像处理等——这些都是专用计算相比通用计算能创造巨大价值差异的领域。

### Topic 82: Democratizing deep learning and broader impact
深度学习民主化与更广泛影响
_[01:41:50]_

**Q:** How did Nvidia democratize deep learning and what other important work continues beyond AI?
**问：** Nvidia如何使深度学习民主化，AI之外还有哪些重要工作在继续？

**A:** The speaker emphasizes that Nvidia's core mission to democratize computing remains unchanged, having made deep learning accessible to "any researcher, any scientist, anywhere, any student" through affordable hardware like GeForce cards. While acknowledging AI's excitement, the speaker deliberately redirects attention to Nvidia's substantial non-AI work, pointing to "the whole beginning part" of GTC that covers computational lithography, quantum chemistry, and data processing. The speaker stresses that "tensors are not the only way that you compute" and that many people are doing "very important work that's not AI related." This reflects a broader vision where AI is one application among many in Nvidia's mission to "help everybody" advance computational science.
**答：** 演讲者强调Nvidia让计算民主化的核心使命从未改变，通过GeForce显卡等平价硬件，让"任何研究者、科学家、学生"都能接触深度学习。虽然承认AI很令人兴奋，但演讲者特意将注意力引向Nvidia大量的非AI工作，指出GTC大会"整个开头部分"都在讲计算光刻、量子化学、数据处理等内容。他强调"tensor不是唯一的计算方式"，很多人在做"非常重要的非AI相关工作"。这反映了一个更宏大的愿景：AI只是Nvidia帮助所有人推进计算科学这一使命中的一个应用领域。

---

## Vocabulary (CEFR B2+)

### commoditize  /kəˈmɑːdɪtaɪz/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 4

**EN:** to turn a product or service into a commodity with little differentiation, reducing its value and competitive advantage  
**CN:** 使商品化，使成为无差异的普通商品（从而降低其价值和竞争优势）

**Original examples:**
- [00:00] We've seen the valuations of a bunch of software companies crash because people are expecting AI to **commoditize** software.  
  我们看到许多软件公司的估值暴跌，因为人们预期 AI 会让软件**商品化**。
- [00:23] Nvidia is fundamentally making software that other people are manufacturing, and if software gets **commoditized**, does Nvidia get **commoditized**?  
  Nvidia 本质上是在制造由其他人生产的软件，如果软件被**商品化**了，Nvidia 会不会也被**商品化**？
- [00:42] The transformation of electrons to tokens and making those tokens more valuable over time is hard to completely **commoditize**.  
  将电子转化为 token 并随着时间推移让这些 token 变得更有价值，这很难被完全**商品化**。
- [02:46] I don't think that gets **commoditized**.  
  我认为那不会被**商品化**。

**Extra example:**
- As the technology matures, there's a risk that premium features will be **commoditized** by competitors.  
  随着技术成熟，高端功能有可能被竞争对手**商品化**。

### valuation  /ˌvæljuˈeɪʃən/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** an estimation of the worth or market value of a company or asset  
**CN:** 估值，对公司或资产价值的评估

**Original examples:**
- [00:00] We've seen the **valuations** of a bunch of software companies crash because people are expecting AI to commoditize software.  
  我们看到许多软件公司的**估值**暴跌，因为人们预期 AI 会让软件商品化。

**Extra example:**
- The startup's **valuation** reached $2 billion after the latest funding round.  
  这家初创公司在最新一轮融资后**估值**达到了 20 亿美元。

### capacity  /kəˈpæsəti/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 1

**EN:** the maximum amount that something can produce, contain, or handle  
**CN:** 产能，容量，能力

**Original examples:**
- [05:51] The reason for that is because they know that I have the **capacity** to buy their supply and sell it through my downstream. The fact is that Nvidia's downstream supply chain and our downstream demand is so large, they're willing to make the investment upstream.  
  原因是他们知道我有能力购买他们的供应并通过我的下游渠道销售出去。事实是 Nvidia 的下游供应链和下游需求规模如此之大,他们才愿意在上游做投资。

**Extra example:**
- The factory is operating at full **capacity** to meet demand.  
  工厂正在满负荷运转以满足需求。

### transformation  /ˌtrænsfərˈmeɪʃən/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 4

**EN:** a thorough or dramatic change in form, appearance, or character  
**CN:** 转变，转化，彻底改变

**Original examples:**
- [00:42] The **transformation** of electrons to tokens and making those tokens more valuable over time is hard to completely commoditize.  
  把电子转换成 token,并且让这些 token 随着时间推移变得更有价值,这件事很难被完全商品化。
- [00:59] The **transformation** from electrons to tokens is such an incredible journey.  
  从电子到 token 的**转化**是一段令人难以置信的旅程。
- [01:21] The **transformation**, the manufacturing, all of the science that goes in there is far from deeply understood and the journey is far from over.  
  这种**转化**、制造以及其中涉及的所有科学原理远未被深入理解，这段旅程还远未结束。
- [01:50] Our job is to do as much as necessary and as little as possible to enable that **transformation** to be done at incredible capabilities.  
  我们的工作是尽可能多地做必要的事，尽可能少地做不必要的事，以使这种**转化**能够以惊人的能力完成。

**Extra example:**
- The company underwent a digital **transformation** to stay competitive in the market.  
  这家公司经历了数字化**转型**以保持市场竞争力。

### ecosystem  /ˈiːkoʊˌsɪstəm/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 16

**EN:** a complex network of interconnected organizations, partners, and technologies that work together  
**CN:** 生态系统，由相互关联的组织、合作伙伴和技术组成的复杂网络

**Original examples:**
- [02:05] Whatever I don't need to do, I partner with somebody and make it part of my **ecosystem**.  
  凡是我不需要做的事，我就和别人合作，让它成为我**生态系统**的一部分。
- [02:16] If you look at Nvidia today, we probably have the largest **ecosystem** of partners, both in the supply chain upstream and downstream, all of the computer companies, application developers, and model makers.  
  如果你看今天的 Nvidia，我们可能拥有最大的合作伙伴**生态系统**，包括上下游供应链、所有计算机公司、应用开发者和模型制造商。
- [06:48] I spend a lot of my time informing, directly or indirectly, our supply chain, partners, and **ecosystem** about the opportunity in front of us.  
  我花很多时间直接或间接地向我们的供应链、合作伙伴和**生态系统**介绍我们面前的机会。
- [12:06] For example, the investments that we've done with Lumentum, Coherent, and the silicon photonics **ecosystem** over the last several years really reshaped the supply chain.  
  比如说,过去几年我们对 Lumentum、Coherent 以及硅光子生态系统的投资,真正重塑了整个供应链。
- [17:47] We have a gigantic **ecosystem**.  
  我们有一个庞大的**生态系统**。
- [26:11] Because the **ecosystem** is so rich, we support every framework.  
  因为**生态系统**如此丰富，我们支持每一个框架。
- [26:50] So if you want to build on an architecture, building on CUDA makes the most sense because you know the **ecosystem** is great.  
  所以如果你想在某个架构上构建,在 CUDA 上构建最合理,因为你知道生态系统很棒。
- [27:31] That's number one: the richness, programmability, and capability of the **ecosystem**.  
  这是第一点：**生态系统**的丰富性、可编程性和能力。
- [27:57] Nvidia's CUDA **ecosystem** is ultimately its great treasure.  
  Nvidia 的 CUDA 生态系统最终是它最宝贵的财富。
- [01:17:00] The chip industry is part of the American **ecosystem**.  
  芯片产业是美国**生态系统**的一部分。
- [01:17:03] It's part of American technology leadership. It's part of the AI **ecosystem**.  
  它是美国技术领导地位的一部分，是 AI **生态系统**的一部分。
- [01:19:40] Most important thing to our company is the richness of our **ecosystem**, which is about developers. 50% of the AI developers are in China. The United States should not give that up.  
  对我们公司来说最重要的是我们生态系统的丰富性,也就是开发者。全球 50% 的 AI 开发者在中国。美国不应该放弃这一点。
- [01:19:40] Most important thing to our company is the richness of our **ecosystem**, which is about developers.  
  对我们公司来说最重要的是我们**生态系统**的丰富性，这关乎开发者。
- [01:20:55] These **ecosystems** are hard to replace.  
  这些**生态系统**很难被替代。
- [01:24:43] No, it's just better. It's better because it's easier to program. We have a better **ecosystem**. But whatever the better is, whatever the better is... And of course we're going to send them compute.  
  不,就是更好。它更好是因为更容易编程。我们有更好的生态系统。但无论更好是什么,无论更好是什么……当然我们会给他们提供算力。
- [01:27:38] You have all the software. It's just hard to imagine that there's a long-term lock-in to the Chinese **ecosystem**, even if they have a slightly better open source model for a while.  
  你拥有所有的软件。很难想象会长期锁定在中国的生态系统里,即使他们在一段时间内有一个稍微好一点的开源模型。

**Extra example:**
- Building a thriving developer **ecosystem** is crucial for any platform's long-term success.  
  建立一个繁荣的开发者**生态系统**对任何平台的长期成功都至关重要。

### commitment  /kəˈmɪtmənt/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 3

**EN:** a pledge or obligation to do something, especially a financial or contractual obligation  
**CN:** 承诺，义务（尤指财务或合同义务）

**Original examples:**
- [04:26] I think in your latest filings, you had almost $100 billion in purchase **commitments** with foundries, memory, and packaging.  
  我记得在你们最新的文件中，你们对晶圆厂、内存和封装有近 1000 亿美元的采购**承诺**。
- [04:38] SemiAnalysis has reported that you will have $250 billion of these kinds of purchase **commitments**.  
  SemiAnalysis 报道称你们将有 2500 亿美元的此类采购**承诺**。
- [05:01] We've made enormous **commitments** upstream.  
  我们在上游做出了巨大的**承诺**。

**Extra example:**
- The company's **commitment** to sustainability includes reducing carbon emissions by 50% by 2030.  
  该公司对可持续发展的**承诺**包括到 2030 年将碳排放减少 50%。

### moat  /moʊt/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 6

**EN:** a competitive advantage that protects a company from competitors, like a moat protects a castle  
**CN:** 护城河，竞争壁垒（保护公司免受竞争对手威胁的竞争优势）

**Original examples:**
- [04:44] One interpretation is that Nvidia's **moat** is really that you've locked up many years of these scarce components.  
  一种解读是 Nvidia 的**护城河**实际上是你们锁定了这些稀缺组件多年的供应。
- [04:58] Is this really Nvidia's big **moat** for the next few years?  
  这真的是 Nvidia 未来几年的主要**护城河**吗？
- [07:35] Regarding the **moat** as you describe it, we're able to build for a future.  
  关于你所描述的**护城河**，我们能够为未来而建。
- [30:13] Whereas historically Nvidia has just had, and still has, the best margins in all of AI across hardware and software, over 70%, because of this CUDA **moat**.  
  而从历史上看，Nvidia 一直拥有，现在仍然拥有 AI 领域所有硬件和软件中最好的利润率，超过 70%，这是因为 CUDA 这个**护城河**。
- [30:21] And the question is, can you sustain those margins if for most of your customers, they can actually afford to build instead of the CUDA **moat**?  
  问题是，如果你的大多数客户实际上有能力自己构建而不依赖 CUDA **护城河**，你还能维持这些利润率吗？
- [01:19:28] When we started the conversation today, you acknowledged that Nvidia's position is very different. You used words like **moat**.  
  当我们今天开始对话时，你承认 Nvidia 的地位非常不同。你用了像**护城河**这样的词。

**Extra example:**
- The company's strong brand loyalty serves as a powerful **moat** against new entrants.  
  该公司强大的品牌忠诚度成为抵御新进入者的有力**护城河**。

### align  /əˈlaɪn/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to arrange in a straight line or bring into agreement with a particular standard or goal  
**CN:** 使一致，使协调，使结盟

**Original examples:**
- [05:33] As a result of that process of informing, inspiring, and **aligning** with CEOs of all different industries upstream, they're willing to make the investments.  
  通过这个告知、激励和与上游各行各业 CEO **协调一致**的过程，他们愿意进行投资。

**Extra example:**
- We need to **align** our marketing strategy with the company's overall business objectives.  
  我们需要让营销策略与公司的整体业务目标**保持一致**。

### scale  /skeɪl/
**CEFR:** B2 | **Part of speech:** n./v. | **Occurrences:** 13

**EN:** the size or extent of something; to grow or expand in size, number, or capacity  
**CN:** 规模，范围；扩大规模，按比例增长

**Original examples:**
- [06:11] If you look at GTC, people are marveled by the **scale** of it and the people that go.  
  如果你看 GTC，人们会惊叹于它的**规模**和参与的人数。
- [07:56] If our next several years are a trillion dollars in **scale**, we have the supply chain to do it.  
  如果我们未来几年的**规模**达到一万亿美元，我们有供应链来实现它。
- [07:56] If our next several years are a trillion dollars in **scale**, we have the supply chain to do it. Without our reach, the velocity of our business...  
  如果我们未来几年的规模是万亿美元级别,我们有供应链来实现它。没有我们的影响力,没有我们业务的速度...
- [08:27] That allows us to do the things we're able to do at the **scale** we do them.  
  这让我们能够以我们现在的**规模**做我们能做的事情。
- [08:44] And 2x-ing at this **scale** now is really incredible.  
  确实如此。
- [10:30] TSMC now knows that CoWoS supply has to keep up with the rest of the logic demand and the memory demand. They're **scaling** CoWoS and future packaging technologies at the same level as they **scale** logic.  
  TSMC 现在知道 CoWoS 供应必须跟上其他逻辑需求和内存需求。他们正在以与**扩展**逻辑相同的水平**扩展** CoWoS 和未来的封装技术。
- [10:30] TSMC now knows that CoWoS supply has to keep up with the rest of the logic demand and the memory demand. They're **scaling** CoWoS and future packaging technologies at the same level as they scale logic.  
  TSMC 现在明白了,CoWoS 的供应必须跟上逻辑芯片需求和内存需求的步伐。他们正在以与逻辑芯片相同的规模扩展 CoWoS 和未来的封装技术。
- [12:42] New testing equipment like double-sided probing, investing in companies, and helping them **scale** up their capacity. You can see that we're trying to shape the ecosystem so that the supply chain is ready to support the scale.  
  比如双面探针这样的新测试设备,投资相关公司,帮助它们扩大产能。你可以看到我们在努力塑造整个生态系统,让供应链能够支撑这样的规模。
- [12:42] You can see that we're trying to shape the ecosystem so that the supply chain is ready to support the **scale**.  
  你可以看到我们正在努力塑造生态系统，以便供应链准备好支持这个**规模**。
- [13:00] **Scaling** up CoWoS versus scaling up—I went to the hardest one, by the way.  
  扩大 CoWoS 产能和扩大——顺便说一句,我说的是最难的那个。
- [13:00] **Scaling** up CoWoS versus **scaling** up—I went to the hardest one, by the way.  
  **扩大** CoWoS 规模与**扩大**——顺便说一句，我说的是最难的那个。
- [14:10] None of that is impossible to **scale** quickly.  
  这些都不是不可能快速**扩大规模**的。
- [14:10] None of that is impossible to **scale** quickly. All of that is easy to do within two or three years. You just need a demand signal.  
  这些都不是不可能快速扩大规模的。这些在两三年内都很容易做到。你只需要一个需求信号。

**Extra example:**
- The startup needs to **scale** its operations rapidly to meet growing customer demand.  
  这家初创公司需要快速**扩大**运营规模以满足不断增长的客户需求。

### churn  /tʃɜːrn/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** the rate of movement, turnover, or change in a system or process  
**CN:** 流转率，周转速度

**Original examples:**
- [08:05] Just as there's cash flow, there's supply chain flow, there's **churn**.  
  就像有现金流一样，也有供应链流，也有流转速度。

**Extra example:**
- High inventory **churn** indicates efficient supply chain management.  
  高库存周转率表明供应链管理高效。

### bottleneck  /ˈbɑːtlnek/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 11

**EN:** a point of congestion or blockage that slows down a process or system  
**CN:** 瓶颈，阻碍进程或系统的拥堵点

**Original examples:**
- [11:56] Each one of these **bottlenecks** gets a great deal of attention.  
  这些**瓶颈**中的每一个都得到了极大的关注。
- [12:02] Now we're prefetching the **bottlenecks** years in advance.  
  现在我们提前数年预判这些**瓶颈**。
- [14:03] Ultimately, memory and logic are **bottlenecked** by EUV.  
  最终，内存和逻辑都受到EUV的**瓶颈制约**。
- [15:04] My point is that none of the **bottlenecks** last longer than a couple of years, two, three years, none of them.  
  我的观点是，这些**瓶颈**没有一个会持续超过几年，两三年，一个都没有。
- [01:02:50] We're limited by energy, but we've got a lot of people working on that. We've got to not make energy a **bottleneck** for our country.  
  我们受到能源的限制，但有很多人在解决这个问题。我们不能让能源成为我们国家的瓶颈。
- [01:04:44] If you talk to any AI lab in America, they say the thing that's **bottlenecking** them is compute.  
  如果你和美国任何一个AI实验室交流，他们都会说**制约**他们的是算力。
- [01:04:44] If you talk to any AI lab in America, they say the thing that's **bottlenecking** them is compute. There are quotes from the DeepSeek founder, or Qwen leadership or whatever. They say the thing they're bottlenecked on is compute.  
  如果你跟美国任何一家 AI 实验室聊,他们都会说瓶颈在算力。DeepSeek 创始人或者 Qwen 领导层也有类似的引述,他们说瓶颈就是算力。
- [01:08:42] Right. But as you know, the **bottleneck** often in training and doing inference on these models is the amount of bandwidth.  
  对。但你知道，在训练和推理这些模型时，**瓶颈**往往是带宽量。
- [01:16:22] But how is shipping chips to China keeping the US ahead if they're **bottlenecked** on compute?  
  但如果他们在算力上受到**瓶颈制约**，向中国出口芯片如何能让美国保持领先？
- [01:23:28] But there's a reason they're buying it from you. We have quotes from the founders of Chinese companies that say that they're **bottlenecked** on compute. Because our chips are better. On balance, our chips are better.  
  但他们从你这里购买是有原因的。我们有中国公司创始人的引述,说他们在算力上遇到了瓶颈。因为我们的芯片更好。总体而言,我们的芯片更好。
- [01:35:06] We were discussing earlier these **bottlenecks** at TSMC and memory and so forth.  
  我们之前讨论了TSMC和内存等方面的这些**瓶颈**。

**Extra example:**
- Traffic congestion at the bridge creates a major **bottleneck** during rush hour.  
  桥梁处的交通拥堵在高峰时段造成了严重的**瓶颈**。

### diverse  /daɪˈvɜːrs/
**CEFR:** B2 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** showing a great deal of variety; very different from each other  
**CN:** 多样化的，各种各样的

**Original examples:**
- [17:14] In addition, we use it for AI. Accelerated computing is much more **diverse**.  
  此外，我们将其用于AI。加速计算要**多样化**得多。

**Extra example:**
- The company has a **diverse** portfolio of products across multiple industries.  
  该公司在多个行业拥有**多样化**的产品组合。

### offtaker  /ˈɒfˌteɪkər/
**CEFR:** C2 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a customer or entity that purchases and consumes the output or product  
**CN:** 购买方，承购方

**Original examples:**
- [18:31] If you want to operate it to rent, you better have a large ecosystem of customers in many industries to be the **offtakers**.  
  如果你想通过租赁来运营,那你最好拥有一个庞大的客户生态系统,覆盖多个行业,这些客户可以成为承租方。

**Extra example:**
- The power plant secured a long-term **offtaker** agreement with the utility company.  
  这家电厂与公用事业公司签订了长期承购协议。

### programmable  /ˈproʊɡræməbl/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 3

**EN:** able to be programmed or configured to perform different tasks or functions  
**CN:** 可编程的，能够被编程以执行不同任务或功能的

**Original examples:**
- [21:07] If you want to come up with a new attention mechanism, disaggregate in a different way, or invent a whole new type of architecture altogether—like a hybrid SSM—you want an architecture that's generally **programmable**.  
  如果你想提出新的注意力机制，以不同方式解耦，或者发明全新类型的架构——比如混合SSM——你需要一个通用**可编程**的架构。
- [21:23] If you want to create a model that fuses diffusion and autoregressive techniques, you want an architecture that's just generally **programmable**.  
  如果你想创建一个融合扩散和自回归技术的模型，你需要一个通用**可编程**的架构。
- [21:38] We run everything you can imagine. That's the advantage. It allows for the invention of new algorithms a lot more easily, because it's a **programmable** system.  
  我们可以运行你能想到的一切。这就是优势所在。它让新算法的发明变得容易得多,因为这是一个可编程的系统。

**Extra example:**
- Modern thermostats are **programmable**, allowing users to set different temperatures for different times of day.  
  现代恒温器是**可编程的**，允许用户为一天中的不同时段设置不同温度。

### sustain  /səˈsteɪn/
**CEFR:** B2 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to maintain or keep something going over time; to support or uphold  
**CN:** 维持，保持；支撑

**Original examples:**
- [30:21] And the question is, can you **sustain** those margins if for most of your customers, they can actually afford to build instead of the CUDA moat?  
  问题是，如果你的大多数客户实际上有能力自己构建而不是依赖CUDA护城河，你能**维持**这些利润率吗？

**Extra example:**
- The company needs to **sustain** its growth rate to meet investor expectations.  
  公司需要**维持**其增长率以满足投资者的期望。

### optimize  /ˈɑːptɪmaɪz/
**CEFR:** B2 | **Part of speech:** v. | **Occurrences:** 3

**EN:** to make something as effective, efficient, or functional as possible  
**CN:** 优化，使最优化

**Original examples:**
- [31:44] It's not unusual that by the time we're done **optimizing** their stack or **optimizing** a particular kernel, their model sped up by 3x, 2x, 50%.  
  当我们完成对他们的技术栈或特定内核的**优化**后，他们的模型速度提升3倍、2倍、50%，这并不罕见。
- [01:32:04] Their models have to be so far **optimized** for  
  他们的模型必须被充分**优化**以适应
- [01:33:29] So the fact of the matter is, AI is about the stack above as much as it is about the architecture below. To the extent that we have architectures and software stacks that are **optimized** for our stack, for our ecosystem, it is obviously good.  
  事实上，AI既关乎上层技术栈，也关乎底层架构。在我们拥有为我们的技术栈、为我们的生态系统**优化**的架构和软件栈的程度上，这显然是好事。

**Extra example:**
- We need to **optimize** our website's loading speed to improve user experience.  
  我们需要**优化**网站的加载速度以改善用户体验。

### install base  /ɪnˈstɔːl beɪs/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 3

**EN:** the total number of units of a product or system currently in use by customers  
**CN:** 装机量，现有用户群的设备总数

**Original examples:**
- [32:01] That's a huge number, especially when you're talking about the **install base** of the fleet that they have, of all the Hoppers and Blackwells that they have.  
  这是一个巨大的数字，尤其是当你谈论他们拥有的所有Hopper和Blackwell的**装机量**时。
- [34:01] So I think the flywheel is really **install base**, the programmability of our architecture, the richness of our ecosystem, and the fact that there's so many AI companies in the world.  
  所以我认为飞轮效应实际上是**装机量**、我们架构的可编程性、我们生态系统的丰富性，以及世界上有如此多AI公司这一事实。
- [34:29] We're the most abundant in the world. You'd choose the one that has the largest **installed base**. We're the largest **install base**. And you'd choose the one that has a rich ecosystem.  
  我们在全球最为充足。你会选择拥有最大**装机量**的那个。我们是最大的**装机量**。而且你会选择拥有丰富生态系统的那个。

**Extra example:**
- Apple's large **install base** of iPhone users gives them a significant advantage in launching new services.  
  Apple庞大的iPhone用户**装机量**为他们推出新服务提供了显著优势。

### TCO  /ˌtiː siː ˈoʊ/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 3

**EN:** Total Cost of Ownership; the complete cost of acquiring, operating, and maintaining a product or system over its entire lifecycle  
**CN:** 总拥有成本（Total Cost of Ownership的缩写），指在整个生命周期内获取、运营和维护产品或系统的全部成本

**Original examples:**
- [32:16] Nvidia's computing stack is the best performance per **TCO** in the world, bar none.  
  Nvidia的计算技术栈在全球拥有最佳的性能**总拥有成本比**，无一例外。
- [32:24] Nobody can demonstrate to me that any single platform in the world today has a better performance-**TCO** ratio.  
  没有人能向我证明当今世界上任何单一平台拥有更好的性能**总拥有成本比**。
- [33:18] So I think the reason why we're so successful is simply because our **TCO** is so great.  
  所以我认为我们如此成功的原因就是我们的**总拥有成本**非常出色。

**Extra example:**
- When evaluating cloud providers, companies should consider **TCO** rather than just upfront costs.  
  在评估云服务提供商时，公司应该考虑**总拥有成本**而不仅仅是前期成本。

### flywheel  /ˈflaɪwiːl/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 2

**EN:** a business concept describing a self-reinforcing cycle where success in one area drives success in others, creating momentum  
**CN:** 飞轮效应，指一个自我强化的循环，其中一个领域的成功推动其他领域的成功，从而产生动力

**Original examples:**
- [34:01] So I think the **flywheel** is really install base, the programmability of our architecture, the richness of our ecosystem, and the fact that there's so many AI companies in the world.  
  所以我认为**飞轮效应**实际上是装机量、我们架构的可编程性、我们生态系统的丰富性，以及世界上有如此多AI公司这一事实。
- [35:17] Lastly, if your goal is to rent the infrastructure, we have the most customers in the world. So that's the reason why the **flywheel** works.  
  最后，如果你的目标是租用基础设施，我们在全球拥有最多的客户。这就是**飞轮效应**起作用的原因。

**Extra example:**
- Amazon's **flywheel** effect connects lower prices, more customers, more sellers, and greater selection in a virtuous cycle.  
  Amazon的**飞轮效应**将更低的价格、更多的客户、更多的卖家和更丰富的选择连接在一个良性循环中。

### premise  /ˈpremɪs/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 9

**EN:** a statement or idea that forms the basis for a theory or argument  
**CN:** 前提，假设（构成理论或论证基础的陈述或观点）

**Original examples:**
- [35:51] No, I think your **premise** is wrong.  
  不，我认为你的**前提**是错的。
- [36:08] But still make sure to make me come back and fix because it's just too important to AI. It's too important to the future of science. It's too important to the future of the industry. That **premise**...  
  但仍然要确保让我回来修复，因为这对AI太重要了。对科学的未来太重要了。对行业的未来太重要了。那个**前提**……
- [51:49] No. No. Your **premise** is just wrong. We're sufficiently mindful about these things.  
  不同意。你的前提就是错的。我们对这些事情考虑得很周全。
- [01:09:14] They've already demonstrated silicon photonics, connecting all of this compute together into one giant supercomputer. Your **premise** is just wrong.  
  他们已经展示了硅光子技术，将所有这些算力连接成一台巨型超级计算机。你的**前提**就是错的。
- [01:11:15] You set it up as a **premise** that it was bad news. I'm going to give you the bad news, that AI models around the world are developed and they run best on non-American hardware.  
  但你把它设定为一个坏消息的前提。我要告诉你真正的坏消息:全球各地开发的 AI 模型在非美国硬件上运行得最好。
- [01:20:12] The **premise** that even if we competed in China, that we're going to lose that market anyways... You're not talking to somebody who woke up a loser.  
  那种即使我们在中国竞争，我们也会失去那个市场的**前提**……你面对的不是一个天生的失败者。
- [01:21:10] Conceding a marketplace based on the **premise** you described, I simply can't acknowledge that.  
  基于你描述的**前提**而放弃一个市场，我根本无法认同。
- [01:29:29] So I'm making the case that when you make a **premise** that is so extreme, everything goes—  
  所以我要说的是，当你提出一个如此极端的**前提**时，一切都会——
- [01:39:55] The **premise** of our company is that Moore's law is going to… General purpose computing  
  我们公司的**前提**是Moore定律将会……通用计算

**Extra example:**
- The entire argument rests on the **premise** that economic growth will continue.  
  整个论证建立在经济增长将持续这一**前提**之上。

### internalize  /ɪnˈtɜːrnəlaɪz/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 2

**EN:** to accept or absorb an idea, belief, or value so that it becomes part of your character or way of thinking  
**CN:** 内化，使成为思想的一部分（接受或吸收某种观点、信念或价值观，使其成为性格或思维方式的一部分）

**Original examples:**
- [39:12] A long time ago, we just didn't have the ability to do it. At the time, I didn't deeply **internalize** how difficult it would be to build a foundation  
  很久以前,我们确实没有能力做这件事。当时我没有深刻意识到,建立一个像 OpenAI 和 Anthropic 这样的基础
- [40:00] I would say my mistake is I didn't deeply **internalize** that they really had no other options, that a VC would never put in $5-10 billion of investment into an AI lab with the hopes of it turning out to be Anthropic.  
  我想说我的错误是我没有深刻**内化**这一点：他们真的别无选择，风投永远不会向一个AI实验室投入50-100亿美元，指望它成为Anthropic。

**Extra example:**
- It takes time for children to **internalize** social norms and values.  
  孩子们需要时间来**内化**社会规范和价值观。

### philosophy  /fəˈlɑːsəfi/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a set of beliefs or an approach to life that guides someone's behavior or decisions  
**CN:** 理念，哲学（指导某人行为或决策的一套信念或生活方式）

**Original examples:**
- [44:15] This is a **philosophy** of the company, and I think it's wise.  
  这是公司的一种**理念**，我认为这很明智。

**Extra example:**
- Our **philosophy** is to put customer satisfaction first in everything we do.  
  我们的**理念**是在所有事情上都把客户满意度放在首位。

### thrive  /θraɪv/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 2

**EN:** to grow, develop, or be successful  
**CN:** 繁荣，兴旺（成长、发展或取得成功）

**Original examples:**
- [46:34] Because I want our ecosystem to **thrive**. I want the architecture and AI to be able to connect with as many industries as possible, as many countries as possible, and make it possible for the planet to be built on AI and to be built on the American tech stack.  
  因为我希望我们的生态系统能够**繁荣发展**。我希望这个架构和AI能够与尽可能多的行业、尽可能多的国家连接，让整个地球都能建立在AI之上，建立在American技术栈之上。
- [01:01:58] We know very well that this ecosystem needs to **thrive**.  
  我们非常清楚这个生态系统需要**繁荣发展**。

**Extra example:**
- Small businesses can **thrive** in the right economic environment.  
  小企业在合适的经济环境中能够**蓬勃发展**。

### imperative  /ɪmˈperətɪv/
**CEFR:** C1 | **Part of speech:** adj./n. | **Occurrences:** 1

**EN:** extremely important or urgent; something that is essential  
**CN:** 极其重要的，紧迫的；必要的事情

**Original examples:**
- [47:08] This is another thing that we do. We don't pick winners. We need to support everyone. It's part of our joy of doing so. It's **imperative** to our business. But we also go out of our way not to pick winners. So when I invest in one of them, I invest in all of them.  
  这是我们做的另一件事。我们不挑选赢家。我们需要支持所有人。这是我们乐于做的事情的一部分。这对我们的业务来说是**至关重要的**。但我们也会特意不去挑选赢家。所以当我投资其中一家时，我会投资所有人。

**Extra example:**
- It is **imperative** that we act quickly to address climate change.  
  我们必须迅速采取行动应对气候变化，这是**极其重要的**。

### humility  /hjuːˈmɪləti/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** the quality of being humble; not thinking you are better than other people  
**CN:** 谦逊，谦卑（不认为自己比别人优越的品质）

**Original examples:**
- [48:07] It was never going to make it. We reasoned about it from good first principles, but we ended up with the wrong solution. Everybody would have counted us out. And here we are. So I have enough **humility** to recognize that. Don't pick winners. Either let them all take care of themselves, or take care of all of them.  
  它永远不会成功。我们从良好的第一性原理出发进行推理，但最终得出了错误的解决方案。所有人都会认为我们出局了。但我们还在这里。所以我有足够的**谦逊**来认识到这一点。不要挑选赢家。要么让他们都自己照顾自己，要么照顾所有人。

**Extra example:**
- Despite his success, he showed great **humility** and never boasted about his achievements.  
  尽管取得了成功，他表现出极大的**谦逊**，从不夸耀自己的成就。

### dependable  /dɪˈpendəbl/
**CEFR:** B2 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** reliable and trustworthy; able to be counted on  
**CN:** 可靠的，值得信赖的（能够被依靠的）

**Original examples:**
- [54:39] That's just never been a practice of ours. You can count on us. I prefer to be **dependable**, to be the foundation of the industry. You don't need to second-guess. If I quoted you a price, we quoted you a price. That's it. If demand goes through the roof, so be it.  
  那从来不是我们的做法。你可以信赖我们。我更愿意做到**可靠**，成为行业的基石。你不需要猜疑。如果我给你报了价，我们就报了价。就是这样。如果需求暴增，那就这样吧。

**Extra example:**
- She's a **dependable** employee who always meets her deadlines.  
  她是一位**可靠的**员工，总是按时完成任务。

### adversary  /ˈædvərseri/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 2

**EN:** an opponent or rival, especially in a competition or conflict  
**CN:** 对手，敌手（尤指竞争或冲突中的）

**Original examples:**
- [59:55] Victimizing them, turning them into an enemy, likely isn't the best answer. They are an **adversary**. We want the United States to win.  
  把他们当作受害者，把他们变成敌人，可能不是最好的答案。他们是一个**对手**。我们希望United States获胜。
- [01:00:23] This is an area that is glaringly missing because of our current attitude about China as an **adversary**. It is essential that our AI researchers and their AI researchers are actually talking.  
  由于我们目前将中国视为**对手**的态度，这是一个明显缺失的领域。我们的AI研究人员和他们的AI研究人员进行实际对话是至关重要的。

**Extra example:**
- In business, your biggest **adversary** today could become your partner tomorrow.  
  在商业中，你今天最大的**对手**明天可能成为你的合作伙伴。

### victimize  /ˈvɪktɪmaɪz/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to treat someone unfairly or make them suffer, especially by making them seem like a victim  
**CN:** 使受害,使成为牺牲品;不公正地对待

**Original examples:**
- [59:55] **Victimizing** them, turning them into an enemy, likely isn't the best answer.  
  把他们当作受害者,把他们变成敌人,这可能不是最好的答案。

**Extra example:**
- The policy should not **victimize** innocent people who are caught in the middle.  
  这项政策不应该让夹在中间的无辜民众成为牺牲品。

### vulnerability  /ˌvʌlnərəˈbɪləti/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 2

**EN:** a weakness in a system, design, or security that can be exploited to cause harm  
**CN:** 漏洞,弱点;易受攻击性

**Original examples:**
- [01:04:12] They're going to patch up all their **vulnerabilities**, and now we release it.  
  他们会修补所有的漏洞,然后我们再发布它。
- [01:12:18] They find all the security **vulnerabilities** in American software first, but they can do it on Nvidia hardware and they ship it to the global south.  
  他们首先发现American软件中的所有安全漏洞,但他们可以在Nvidia硬件上做到这一点,然后把它运送到全球南方国家。

**Extra example:**
- The security team discovered a critical **vulnerability** in the payment system.  
  安全团队在支付系统中发现了一个严重漏洞。

### deploy  /dɪˈplɔɪ/
**CEFR:** B2 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to put something into use or action, especially on a large scale  
**CN:** 部署,调动;投入使用

**Original examples:**
- [01:04:22] Furthermore, even if they train a model like this, the ability to **deploy** it at scale… If you had a cyber hacker, it's much more dangerous if they have a million of them versus a thousand of them.  
  此外,即使他们训练出这样的模型,大规模部署它的能力……如果你有一个网络黑客,拥有一百万个比拥有一千个要危险得多。

**Extra example:**
- The company plans to **deploy** the new software across all branches next month.  
  公司计划下个月在所有分支机构部署新软件。

### inference  /ˈɪnfərəns/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 2

**EN:** in AI, the process of using a trained model to make predictions or generate outputs  
**CN:** 推理;(AI领域)推断运算

**Original examples:**
- [01:04:22] So that **inference** compute really matters a lot.  
  所以推理计算真的非常重要。
- [01:08:42] But as you know, the bottleneck often in training and doing **inference** on these models is the amount of bandwidth.  
  但你知道,在训练和对这些模型进行推理时,瓶颈往往是带宽的大小。

**Extra example:**
- Running **inference** on large language models requires significant computational resources.  
  在大型语言模型上运行推理需要大量的计算资源。

### threshold  /ˈθreʃhoʊld/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 2

**EN:** the minimum level or point at which something starts to happen or change  
**CN:** 门槛,临界点;起点

**Original examples:**
- [01:06:55] The amount of **threshold** they need for the concern you're worried about, they've already reached that **threshold** and beyond.  
  对于你担心的问题,他们所需的门槛,他们已经达到甚至超过了那个门槛。
- [01:21:56] Those extremes, they're childish. Let me just make my argument for myself. The idea is not that there is some key **threshold** of compute. It's that any marginal compute is helpful. So if you have more compute, you can train a better model. And I just want you to acknowledge that any marginal sales for the American technology industry is beneficial.  
  那些极端说法,很幼稚。让我自己来阐述我的论点。我的意思不是说存在某个算力的关键阈值。而是任何边际算力都是有帮助的。如果你有更多算力,你就能训练出更好的模型。我只是希望你承认,对美国科技行业来说,任何边际销售都是有益的。

**Extra example:**
- The company must meet a minimum revenue **threshold** to qualify for the program.  
  公司必须达到最低收入门槛才有资格参加该项目。

### abundance  /əˈbʌndəns/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 3

**EN:** a very large quantity of something; more than enough  
**CN:** 丰富,充裕;大量

**Original examples:**
- [01:07:04] When you have an **abundance** of energy, it makes up for chips.  
  当你拥有充裕的能源时,它可以弥补芯片的不足。
- [01:07:10] If you have an **abundance** of chips, it makes up for energy.  
  如果你拥有充足的芯片,就可以弥补能源的不足。
- [01:08:07] So 7nm chips are plenty good. The **abundance** of energy is their advantage.  
  所以 7nm 芯片已经足够好了。充足的能源供应就是他们的优势。

**Extra example:**
- The region is blessed with an **abundance** of natural resources.  
  该地区拥有丰富的自然资源。

### lever  /ˈlevər/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a means of achieving something; a factor that provides the most power or influence  
**CN:** 杠杆;手段;关键因素

**Original examples:**
- [01:09:52] What I'm saying is that great computer science is where the **lever** is.  
  我想说的是,优秀的计算机科学才是关键杠杆所在。

**Extra example:**
- Education is the most powerful **lever** for social change.  
  教育是推动社会变革最有力的杠杆。

### analogy  /əˈnælədʒi/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 2

**EN:** a comparison between two things to show how they are similar, used to explain or clarify something  
**CN:** 类比,比拟;相似之处

**Original examples:**
- [01:17:48] The **analogy** is that enriched uranium is like compute.  
  这个类比是说浓缩铀就像算力。
- [01:17:51] It's a lousy **analogy**. It's an illogical **analogy**.  
  这是个糟糕的类比。这是个不合逻辑的类比。

**Extra example:**
- She used the **analogy** of a garden to explain how a business ecosystem works.  
  她用花园的类比来解释商业生态系统是如何运作的。

### concede  /kənˈsiːd/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 6

**EN:** to admit that something is true or valid after first denying or resisting it; to surrender or yield (a possession, right, or advantage)  
**CN:** 承认（某事属实）；让与，放弃（权利、优势或市场）

**Original examples:**
- [01:18:59] **Conceding** the entire market is not going to allow the United States to win the technology race long-term in the chip layer, in the computing stack.  
  **放弃**整个市场不会让美国在芯片层面、在计算技术栈上赢得长期的技术竞赛。
- [01:21:10] **Conceding** a marketplace based on the premise you described, I simply can't acknowledge that.  
  基于你描述的前提**放弃**一个市场，我完全无法认同。
- [01:24:01] To **concede** that market for the United States technology industry is a disservice to our country.  
  为了美国科技产业而**放弃**那个市场是对我们国家的伤害。
- [01:26:45] I'll also tell you the potential cost is we allow one of the most important layers of the AI stack, the chip layer, to **concede** an entire market—the second largest market in the world—so that they could develop scale, so that they could develop their own ecosystem, so that future AI models are optimized in a very different way than the American tech stack.  
  我也告诉你潜在代价是什么:我们让 AI 技术栈中最重要的一层——芯片层——拱手让出一个完整的市场,世界第二大市场,让他们发展规模,让他们建立自己的生态系统,让未来的 AI 模型以一种与美国技术栈非常不同的方式进行优化。
- [01:30:38] I will tell you exactly about today's conversation, about how your policy and what you imagined literally caused the United States to **concede** the second largest market in the world for no good reason at all.  
  我会明确告诉你今天的对话，关于你的政策和你的设想如何导致美国毫无理由地**放弃**了世界第二大市场。
- [01:30:48] We shouldn't **concede** it. If we lose it, we lose it. But why do we **concede** it?  
  我们不应该**放弃**它。如果我们输了，那就输了。但我们为什么要**放弃**它？

**Extra example:**
- After hours of debate, he finally **conceded** that his opponent had a valid point.  
  经过数小时的辩论，他最终**承认**对手的观点有道理。

### crux  /krʌks/
**CEFR:** C2 | **Part of speech:** n. | **Occurrences:** 4

**EN:** the most important or difficult part of a problem or issue; the decisive or critical point  
**CN:** 关键，症结；核心问题

**Original examples:**
- [01:19:10] I guess then the **crux** comes down to, how does selling them chips now help us win in the long term?  
  我想**关键**就在于，现在卖芯片给他们如何帮助我们赢得长期竞争？
- [01:21:37] Okay, great. Then I won't. I appreciate that. But I think maybe the **crux**... and thanks for walking around the circles with me, because I think it helps bring out what the **crux** here is.  
  好的，很好。那我就不说了。我很感激。但我认为**关键**可能是...感谢你和我绕了这么多圈，因为我觉得这有助于揭示这里的**关键**是什么。
- [01:21:42] The **crux** is you're going to extremes.  
  **关键**在于你走向了极端。
- [01:25:48] Okay, let's just step back. It seems like the **crux** here is there's a potential benefit and there's a potential cost.  
  好的，让我们退一步。看起来这里的**关键**是存在潜在收益和潜在成本。

**Extra example:**
- The **crux** of the matter is whether we can afford the investment.  
  问题的**关键**在于我们是否负担得起这项投资。

### marginal  /ˈmɑːrdʒɪnəl/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** relating to or situated at the edge or margin; of secondary importance; small in amount or degree  
**CN:** 边缘的；次要的；微小的，少量的

**Original examples:**
- [01:21:56] Those extremes, they're childish. Let me just make my argument for myself. The idea is not that there is some key threshold of compute. It's that any **marginal** compute is helpful. So if you have more compute, you can train a better model. And I just want you to acknowledge that any marginal sales for the American technology industry is beneficial.  
  那些极端说法,很幼稚。让我自己来阐述我的论点。我的意思不是说存在某个算力的关键阈值。而是任何边际算力都是有帮助的。如果你有更多算力,你就能训练出更好的模型。我只是希望你承认,对美国科技行业来说,任何边际销售都是有益的。

**Extra example:**
- The **marginal** cost of producing one more unit is relatively low.  
  生产多一个单位的**边际**成本相对较低。

### disservice  /dɪsˈsɜːrvɪs/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 4

**EN:** a harmful action; something that causes harm or damage rather than help  
**CN:** 伤害，损害；帮倒忙

**Original examples:**
- [01:24:01] To concede that market for the United States technology industry is a **disservice** to our country. It is a **disservice** to our national security.  
  为了美国科技产业而放弃那个市场是对我们国家的**伤害**。这是对我们国家安全的**伤害**。
- [01:24:10] It is a **disservice** to our technology leadership, all for the benefit of one company.  
  这是对我们技术领导地位的**伤害**，而这一切都是为了一家公司的利益。
- [01:28:27] But my point is that every layer has to succeed. If we scare this country into thinking that AI is somehow a nuclear bomb, so that everybody hates AI and everybody's afraid of AI, I don't know how you're helping the United States. You're doing it a **disservice**.  
  但我的观点是每一层都必须成功。如果我们吓唬这个国家，让他们认为AI某种程度上是核弹，以至于每个人都讨厌AI、每个人都害怕AI，我不知道你如何帮助美国。你是在**伤害**它。
- [01:28:45] If we scare everybody out of doing software engineering jobs because it's going to kill every software engineering job—and we don't have any software engineers as a result of that—we're doing a **disservice** to the United States.  
  如果我们吓得每个人都不做软件工程工作，因为它会消灭所有软件工程工作——结果我们没有任何软件工程师——我们就是在对美国**帮倒忙**。

**Extra example:**
- Spreading misinformation does a **disservice** to public understanding of the issue.  
  传播错误信息对公众理解这个问题是一种**伤害**。

### diffuse  /dɪˈfjuːz/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 4

**EN:** to spread or cause to spread over a wide area or among a large number of people  
**CN:** 扩散，传播；散布

**Original examples:**
- [01:25:16] **Diffuse** American technology. That, I believe, is a positive.  
  **传播**美国技术。我认为这是积极的。
- [01:27:21] As AI **diffuses** out into the rest of the world, their standards, their tech stack, will become...  
  随着AI**扩散**到世界其他地区，他们的标准、他们的技术栈将会变得...
- [01:28:14] The layer that **diffuses** into society, the one that uses it most will benefit from this industrial revolution most.  
  **扩散**到社会中的那一层，使用它最多的人将从这场工业革命中获益最多。
- [01:30:07] But in a few years' time, I'm making you the prediction that when we want the American tech stack, when we want American technology to be **diffused** around the world—out to India, out to the Middle East, out to Africa, out to Southeast Asia—when our country would like to export, because we would like to export our technology, we would like to export our standards, on that day, I want you and I to have that same conversation again.  
  但在几年后，我向你预测，当我们想要American技术栈，当我们想要American技术**扩散**到全世界——到India、到Middle East、到Africa、到Southeast Asia——当我们国家想要出口，因为我们想要出口我们的技术，我们想要出口我们的标准，在那一天，我希望你和我再进行同样的对话。

**Extra example:**
- New ideas **diffuse** rapidly through social media networks.  
  新想法通过社交媒体网络迅速**传播**。

### nuance  /ˈnuːɑːns/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a subtle difference in or shade of meaning, expression, or sound; a quality of complexity and careful distinction  
**CN:** 细微差别；微妙之处

**Original examples:**
- [01:31:22] Both of those things can simultaneously happen. It requires some amount of **nuance**, some amount of maturity instead of absolutes.  
  这两件事可以同时发生。这需要一定程度的**细致考量**，一定程度的成熟，而不是绝对化。

**Extra example:**
- Understanding the **nuances** of cultural differences is essential for effective communication.  
  理解文化差异的**细微之处**对有效沟通至关重要。

### backlash  /ˈbæklæʃ/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a strong negative reaction by a large number of people, especially to a social or political development  
**CN:** 强烈反对，激烈反应；反弹

**Original examples:**
- [01:34:06] It's a policy mistake. Obviously it has **backlash**. It has turned out badly for the United States.  
  这是一个政策错误。显然它引发了**反弹**。结果对美国很不利。

**Extra example:**
- The company faced severe **backlash** from consumers after the controversial advertisement.  
  这家公司在发布有争议的广告后遭到了消费者的强烈**反对**。

### accelerate  /əkˈseləreɪt/
**CEFR:** B2 | **Part of speech:** v. | **Occurrences:** 3

**EN:** to make something happen faster or earlier; to increase speed  
**CN:** 加速，促进

**Original examples:**
- [01:34:19] It enabled, it **accelerated** their chip industry.  
  它促成并**加速**了他们的芯片产业发展。
- [01:39:48] **Accelerated** computing, the same thing we've been doing all along.  
  **加速**计算，这正是我们一直在做的事情。
- [01:40:00] It is good for a lot of things, but for a lot of computation it's not ideal. So we combined an architecture called a GPU, CUDA, to a CPU, so that we can **accelerate** the workload of the CPU.  
  在很多方面都很好用,但对于大量计算任务来说并不理想。所以我们把一种叫 GPU 的架构,结合 CUDA,与 CPU 结合起来,这样就能加速 CPU 的工作负载。

**Extra example:**
- The new policy will **accelerate** economic growth in the region.  
  新政策将**加速**该地区的经济增长。

### democratize  /dɪˈmɑːkrətaɪz/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to make something accessible to everyone, not just a privileged few  
**CN:** 使民主化，使大众化，使普及

**Original examples:**
- [01:41:57] But because of the advances that we made in computing, we **democratized** deep learning.  
  但由于我们在计算领域取得的进步，我们**普及**了深度学习。

**Extra example:**
- The internet has **democratized** access to information worldwide.  
  互联网已经**普及**了全球范围内的信息获取。

### upstream  /ˌʌpˈstriːm/
**CEFR:** C1 | **Part of speech:** adj./adv. | **Occurrences:** 7

**EN:** in the earlier stages of a production process or supply chain; toward the source  
**CN:** 上游的；在供应链前端的

**Original examples:**
- [02:16] We probably have the largest ecosystem of partners, both in the supply chain **upstream** and downstream.  
  我们可能拥有最大的合作伙伴生态系统，包括供应链的**上游**和下游。
- [05:01] We've made enormous commitments **upstream**.  
  我们在**上游**做出了巨大的承诺。
- [05:33] As a result of that process of informing, inspiring, and aligning with CEOs of all different industries **upstream**, they're willing to make the investments.  
  通过与各行业**上游**的CEO们进行沟通、激励和协调，他们愿意进行投资。
- [06:27] I bring them together so that the downstream can see the **upstream**, the **upstream** can see the downstream.  
  我把他们聚集在一起，让下游能看到**上游**，**上游**能看到下游。
- [08:32] I do want to understand more concretely whether the **upstream** can keep up.  
  我确实想更具体地了解**上游**能否跟上。
- [09:09] Are we in a regime now where the growth rate in AI compute has to slow because of **upstream**?  
  我们现在是否处于一种状态，即AI算力的增长速度必须因为**上游**而放缓？
- [09:25] At some level, the instantaneous demand is greater than the supply **upstream** and downstream in the world.  
  在某种程度上，瞬时需求大于全球**上游**和下游的供应。

**Extra example:**
- The company is investing heavily in **upstream** suppliers to secure raw materials.  
  该公司正在大力投资**上游**供应商以确保原材料供应。

### downstream  /ˌdaʊnˈstriːm/
**CEFR:** C1 | **Part of speech:** adj./adv. | **Occurrences:** 6

**EN:** in the later stages of a production process or supply chain; toward the end user  
**CN:** 下游的；在供应链末端的

**Original examples:**
- [02:16] We probably have the largest ecosystem of partners, both in the supply chain upstream and **downstream**.  
  我们可能拥有最大的合作伙伴生态系统，包括供应链的上游和**下游**。
- [05:51] The fact is that Nvidia's **downstream** supply chain and our **downstream** demand is so large, they're willing to make the investment upstream.  
  事实是Nvidia的**下游**供应链和我们的**下游**需求如此庞大，他们愿意在上游进行投资。
- [06:27] I bring them together so that the **downstream** can see the upstream, the upstream can see the **downstream**.  
  我把他们聚集在一起，让**下游**能看到上游，上游能看到**下游**。
- [08:17] Our ability to sustain the scale is only because our **downstream** demand is so great.  
  我们能够维持这个规模,只是因为我们的下游需求如此巨大。
- [09:25] At some level, the instantaneous demand is greater than the supply upstream and **downstream** in the world.  
  在某种程度上，瞬时需求大于全球上游和**下游**的供应。
- [15:36] It's the stuff that's **downstream** from us.  
  这是我们**下游**的东西。

**Extra example:**
- The oil refinery focuses on **downstream** operations like distribution and retail.  
  这家炼油厂专注于**下游**业务，如分销和零售。

### instantaneous  /ˌɪnstənˈteɪniəs/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** happening immediately, without any delay  
**CN:** 瞬间的，即时的

**Original examples:**
- [09:25] At some level, the **instantaneous** demand is greater than the supply upstream and downstream in the world.  
  在某种程度上，**瞬时**需求大于全球上游和下游的供应。

**Extra example:**
- The system provides **instantaneous** feedback to users.  
  该系统为用户提供**即时**反馈。

### swarm  /swɔːrm/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 2

**EN:** to move somewhere in large numbers; to attack or deal with something collectively and intensively  
**CN:** 蜂拥而至；集体攻克

**Original examples:**
- [10:05] If one particular component is too far away, the industry **swarms** it.  
  如果某个特定组件太落后，整个行业就会**集体攻克**它。
- [10:20] The reason for that is because for two years we **swarmed** the living daylights out of it.  
  原因是我们用了两年时间**全力攻克**它。

**Extra example:**
- Engineers **swarmed** the problem until they found a solution.  
  工程师们**集中力量攻克**这个问题，直到找到解决方案。

### prefetch  /ˌpriːˈfetʃ/
**CEFR:** C2 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to retrieve or prepare something in advance before it is needed  
**CN:** 预取，提前获取

**Original examples:**
- [12:02] Now we're **prefetching** the bottlenecks years in advance.  
  现在我们提前数年**预判**瓶颈。

**Extra example:**
- The browser can **prefetch** resources to speed up page loading.  
  浏览器可以**预取**资源以加快页面加载速度。

### demand signal  /dɪˈmænd ˈsɪɡnəl/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** an indicator from the market that shows the level of customer need or desire for a product or service  
**CN:** 需求信号，市场需求指标

**Original examples:**
- [14:10] None of that is impossible to scale quickly. All of that is easy to do within two or three years. You just need a **demand signal**.  
  这些都不是不可能快速扩大规模的。这些在两三年内都很容易做到。你只需要一个需求信号。

**Extra example:**
- Strong **demand signals** from customers prompted the company to increase production capacity.  
  来自客户的强烈**需求信号**促使公司增加产能。

### adjudicate  /əˈdʒuːdɪkeɪt/
**CEFR:** C2 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to make an official decision about who is right in a dispute or competition  
**CN:** 裁决，判定（争议或竞赛的结果）

**Original examples:**
- [16:17] In this case, I just don't have the technical knowledge to **adjudicate**.  
  在这种情况下，我没有足够的技术知识来做出判断。

**Extra example:**
- An independent panel was appointed to **adjudicate** the contract dispute.  
  一个独立小组被任命来裁决这起合同纠纷。

### unprecedented  /ʌnˈpresɪdentɪd/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** never having happened or existed before  
**CN:** 前所未有的，空前的

**Original examples:**
- [20:10] You're making it because AI is an unprecedented technology that is growing **unprecedentedly** fast.  
  你这样做是因为AI是一项前所未有的技术，而且正在以空前的速度增长。

**Extra example:**
- The company achieved **unprecedented** success in its first year.  
  这家公司在第一年就取得了前所未有的成功。

### disaggregate  /ˌdɪsˈæɡrɪɡeɪt/
**CEFR:** C2 | **Part of speech:** v. | **Occurrences:** 2

**EN:** to separate something into its component parts or elements  
**CN:** 分解，拆分（成组成部分）

**Original examples:**
- [21:07] If you want to come up with a new attention mechanism, **disaggregate** in a different way, or invent a whole new type of architecture altogether—like a hybrid SSM—you want an architecture that's generally programmable.  
  如果你想提出一种新的注意力机制，以不同的方式进行拆分，或者发明一种全新的架构——比如混合SSM——你需要一个通用可编程的架构。
- [22:53] The way we solve that problem is with new models, like MoEs, that are parallelized, **disaggregated**, and distributed across a computing system. Without the ability to really get down and come up with new kernels with CUDA, it's really hard to do.  
  我们解决这个问题的方法是使用新模型,比如 MoE,它们是并行化的、解耦的,并分布在整个计算系统中。如果没有能力真正深入并用 CUDA 开发新的内核,这真的很难做到。

**Extra example:**
- The analyst decided to **disaggregate** the data by region to identify specific trends.  
  分析师决定按地区拆分数据以识别具体趋势。

### sandbag  /ˈsændbæɡ/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to deliberately understate one's ability or performance, especially to gain an advantage later  
**CN:** 故意低估（能力或表现），保守陈述

**Original examples:**
- [22:42] Then Dylan wrote an article saying I **sandbagged**, and it's actually fifty times.  
  然后Dylan写了一篇文章说我保守了，实际上是五十倍。

**Extra example:**
- The team was accused of **sandbagging** their quarterly forecast to make the actual results look more impressive.  
  该团队被指控故意低估季度预测，以使实际结果看起来更令人印象深刻。

### offload  /ˌɔːfˈloʊd/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 2

**EN:** to transfer work, data, or responsibility from one system or person to another  
**CN:** 转移，分流（工作、数据或责任）

**Original examples:**
- [23:15] We can even **offload** some of the computation into the fabric itself, like NVLink, or into the network with Spectrum-X.  
  我们甚至可以将一些计算任务分流到结构本身，比如NVLink，或者通过Spectrum-X分流到网络中。
- [01:40:12] Different kernels of code or algorithms could be **offloaded** onto our GPU.  
  不同的代码内核或算法可以被分流到我们的GPU上。

**Extra example:**
- The manager decided to **offload** some administrative tasks to the assistant.  
  经理决定将一些行政任务转移给助理。

### clientele  /ˌklaɪənˈtel/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** the customers or clients of a business, shop, or professional person collectively  
**CN:** 客户群，顾客群

**Original examples:**
- [24:43] This gets at an interesting question about Nvidia's **clientele**.  
  这涉及到一个关于Nvidia客户群的有趣问题。

**Extra example:**
- The restaurant attracts an upscale **clientele** with its refined menu and elegant atmosphere.  
  这家餐厅以其精致的菜单和优雅的氛围吸引了高端客户群。

### versatility  /ˌvɜːrsəˈtɪləti/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 2

**EN:** the ability to adapt or be adapted to many different functions or activities  
**CN:** 多功能性，通用性，适应性

**Original examples:**
- [29:01] The combination of the richness of the ecosystem, the expansiveness of the install base, and the **versatility** of where we are makes CUDA invaluable.  
  生态系统的丰富性、安装基数的广泛性，以及我们所在位置的通用性，这些因素的结合使CUDA变得无价。
- [33:54] And the reason why all these companies are built on Nvidia is because our reach and our **versatility** is so great.  
  所有这些公司都基于Nvidia构建的原因是我们的覆盖范围和通用性非常强大。

**Extra example:**
- The software's **versatility** allows it to be used across multiple industries.  
  该软件的多功能性使其能够应用于多个行业。

### benchmark  /ˈbentʃmɑːrk/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a standard or point of reference against which things may be compared or assessed  
**CN:** 基准，标准（用于比较或评估）

**Original examples:**
- [32:24] In fact, the **benchmarks** that are out there.  
  事实上，现有的那些基准测试。

**Extra example:**
- The company's performance exceeded industry **benchmarks** for customer satisfaction.  
  该公司的表现超过了客户满意度的行业基准。

### demonstrate  /ˈdemənstreɪt/
**CEFR:** B2 | **Part of speech:** v. | **Occurrences:** 2

**EN:** to show or prove something clearly by giving evidence or examples  
**CN:** 证明，展示（通过提供证据或实例）

**Original examples:**
- [32:24] Nobody can **demonstrate** to me that any single platform in the world today has a better performance-TCO ratio.  
  没有人能向我证明当今世界上任何一个平台有更好的性能总拥有成本比。
- [32:55] I would love to hear them **demonstrate** the cost advantage of TPUs.  
  我很想听他们证明 TPU 的成本优势。

**Extra example:**
- The study **demonstrates** a clear link between exercise and mental health.  
  这项研究清楚地证明了运动与心理健康之间的联系。

### abundant  /əˈbʌndənt/
**CEFR:** B2 | **Part of speech:** adj. | **Occurrences:** 2

**EN:** existing or available in large quantities; plentiful  
**CN:** 丰富的，充裕的

**Original examples:**
- [34:26] You would choose an architecture that's most **abundant**.  
  你会选择最丰富的架构。
- [34:29] We're the most **abundant** in the world.  
  我们是世界上最丰富的。

**Extra example:**
- The region has **abundant** natural resources including oil and minerals.  
  该地区拥有丰富的自然资源，包括石油和矿产。

### sensibility  /ˌsensəˈbɪləti/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** the ability to understand and respond to something; a particular way of thinking or awareness  
**CN:** 感知力，意识；（特定的）思维方式

**Original examples:**
- [42:17] Was it like a cash thing or a **sensibility** at the time?  
  是因为现金问题还是当时的意识问题？

**Extra example:**
- The company needs to develop a stronger environmental **sensibility**.  
  公司需要培养更强的环保意识。

### regret  /rɪˈɡret/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a feeling of sadness or disappointment about something you did or did not do  
**CN:** 遗憾，后悔

**Original examples:**
- [43:21] It's still okay to have **regrets**.  
  有遗憾也没关系。

**Extra example:**
- My only **regret** is not traveling more when I was younger.  
  我唯一的遗憾是年轻时没有多去旅行。

### backstop  /ˈbækstɑːp/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a support or safeguard that prevents failure or provides protection in case something goes wrong  
**CN:** 支持，保障措施（防止失败或提供保护）

**Original examples:**
- [44:08] Why doesn't Nvidia become a cloud themselves and provide a **backstop** with all this cash?  
  为什么 Nvidia 不自己成为云服务商，用这些现金提供保障？

**Extra example:**
- The central bank serves as a **backstop** for the financial system during crises.  
  中央银行在危机期间为金融系统提供保障。

### dedicate  /ˈdedɪkeɪt/
**CEFR:** B2 | **Part of speech:** v. | **Occurrences:** 2

**EN:** to give all your time, energy, or effort to something  
**CN:** 致力于，投入（全部时间、精力）

**Original examples:**
- [44:31] If we didn't **dedicate** ourselves to 20 years of CUDA while losing money most of that time—if we didn't do it, nobody else would have done it.  
  如果我们没有在 20 年里致力于 CUDA 开发——在大部分时间里都在亏损——如果我们不做，就没有其他人会做。
- [45:19] We **dedicated** resources to create a library for computational lithography called cuLitho.  
  我们投入资源创建了一个名为 cuLitho 的计算光刻库。

**Extra example:**
- She **dedicated** her entire career to cancer research.  
  她将整个职业生涯都致力于癌症研究。

### divvy up  /ˈdɪvi ʌp/
**CEFR:** C1 | **Part of speech:** phrasal v. | **Occurrences:** 1

**EN:** to divide and share something among several people or groups  
**CN:** 分配，分摊

**Original examples:**
- [51:25] Nvidia is known for **divvying up** the scarce allocation, not just based on high bidder, but rather on making sure that these neoclouds exist.  
  Nvidia 以分配稀缺资源而闻名，不仅仅基于出价高低，而是确保这些新兴云服务商能够存在。

**Extra example:**
- The team **divvied up** the project tasks according to each member's expertise.  
  团队根据每个成员的专长分配了项目任务。

### characterization  /ˌkærəktəraɪˈzeɪʃn/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** the way in which something or someone is described or represented  
**CN:** 描述，刻画（某事物或某人的方式）

**Original examples:**
- [51:45] First of all, would you agree with this **characterization** of fracturing the market?  
  首先，你同意这种关于分割市场的描述吗？

**Extra example:**
- The media's **characterization** of the event was quite different from what actually happened.  
  媒体对这一事件的描述与实际发生的情况大相径庭。

### throughput  /ˈθruːpʊt/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 3

**EN:** the amount of material, data, or work processed or produced in a given period of time  
**CN:** 吞吐量,产出量(指在特定时间内处理或生产的材料、数据或工作量)

**Original examples:**
- [53:04] That's just maximizing the **throughput** of our own factory.  
  这只是在最大化我们自己工厂的产出量。
- [01:39:00] Until now, higher **throughput** is always better.  
  到目前为止,更高的吞吐量总是更好。
- [01:39:05] We think there could be a world where there could be very high ASP tokens, and even though the **throughput** is lower in the factory, the ASPs make up for it.  
  我们认为可能存在这样一个世界,即使工厂的吞吐量较低,但由于ASP token非常高,ASP可以弥补这一点。

**Extra example:**
- The new server increased network **throughput** by 40%.  
  新服务器将网络吞吐量提高了40%。

### formidable  /ˈfɔːrmɪdəbl/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** inspiring fear or respect through being impressively large, powerful, or capable  
**CN:** 强大的,令人敬畏的(因规模大、力量强或能力强而令人印象深刻)

**Original examples:**
- [01:02:02] It turns out this ecosystem needs open source. This ecosystem needs open models. They need open stacks so that all of these AI researchers and all these great computer scientists can go build AI systems that are as **formidable** and can keep AI safe.  
  事实证明，这个生态系统需要开源。这个生态系统需要开放模型。他们需要开放技术栈，这样所有这些 AI 研究人员和所有这些优秀的计算机科学家才能去构建同样强大的 AI 系统，来保证 AI 的安全。

**Extra example:**
- China has become a **formidable** competitor in the AI industry.  
  中国已经成为AI行业中一个强大的竞争对手。

### vibrant  /ˈvaɪbrənt/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** full of energy, activity, and life  
**CN:** 充满活力的,生机勃勃的

**Original examples:**
- [01:02:22] So one of the things that we need to make sure that we do is we keep the open source ecosystem **vibrant**.  
  所以我们需要确保做的一件事就是保持开源生态系统的活力。

**Extra example:**
- The city has a **vibrant** tech startup scene.  
  这座城市拥有充满活力的科技创业场景。

### suffocate  /ˈsʌfəkeɪt/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to prevent something from developing freely or successfully; to restrict severely  
**CN:** 扼杀,压制(阻止某事自由或成功发展;严格限制)

**Original examples:**
- [01:02:37] We ought to not **suffocate** that.  
  我们不应该扼杀它。

**Extra example:**
- Excessive regulations can **suffocate** innovation in small businesses.  
  过度的监管可能会扼杀小企业的创新。

### triage  /ˈtriːɑːʒ/
**CEFR:** C1 | **Part of speech:** v./n. | **Occurrences:** 1

**EN:** to prioritize tasks, problems, or patients according to urgency or importance  
**CN:** 分类处理,优先排序(根据紧急程度或重要性对任务、问题或患者进行优先排序)

**Original examples:**
- [01:03:34] Since there are a lot of things, let me just **triage** the response.  
  由于有很多事情,让我先对回应进行分类处理。

**Extra example:**
- The team had to **triage** the bug reports to fix critical issues first.  
  团队必须对错误报告进行优先排序,首先修复关键问题。

### aggregate  /ˈæɡrɪɡət/
**CEFR:** C1 | **Part of speech:** v./n./adj. | **Occurrences:** 1

**EN:** to combine or collect separate items into a single whole or total  
**CN:** 汇总,聚合(将分散的项目组合或收集成一个整体或总数)

**Original examples:**
- [01:05:34] If they want to **aggregate** their compute, they've got plenty of compute to **aggregate**.  
  如果他们想汇总他们的算力,他们有大量的算力可以汇总。

**Extra example:**
- The platform can **aggregate** data from multiple sources in real time.  
  该平台可以实时汇总来自多个来源的数据。

### inconsequential  /ˌɪnkɒnsɪˈkwenʃl/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** not important or significant; trivial  
**CN:** 不重要的,无关紧要的,微不足道的

**Original examples:**
- [01:10:25] DeepSeek is not an **inconsequential** advance.  
  DeepSeek不是一个无关紧要的进步。

**Extra example:**
- The cost difference was **inconsequential** compared to the quality improvement.  
  与质量改进相比,成本差异微不足道。

### disparity  /dɪˈspærəti/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a great difference or inequality between things  
**CN:** 差异,不平等(事物之间的巨大差别或不平等)

**Original examples:**
- [01:11:27] I guess I just don't see the evidence that there's these huge **disparities** that would prevent you from switching accelerators.  
  我想我只是没有看到证据表明存在这些巨大的差异会阻止你切换加速器。

**Extra example:**
- There is a significant **disparity** in computing power between the two systems.  
  这两个系统之间在算力上存在显著差异。

### fungible  /ˈfʌndʒəbl/
**CEFR:** C2 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** interchangeable with another identical item; mutually exchangeable  
**CN:** 可互换的，可替代的

**Original examples:**
- [01:12:39] Why do you think it's perfectly **fungible**, that if you didn't ship them compute it would exactly be replaced by Huawei?  
  你为什么认为它是完全可替代的，如果你不给他们提供算力，Huawei就能完全替代？

**Extra example:**
- Money is **fungible**—one dollar bill can be exchanged for any other dollar bill.  
  货币是可互换的——一张一美元钞票可以和任何其他一美元钞票互换。

### fixated  /fɪkˈseɪtɪd/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** obsessively focused on something; unable to stop thinking about something  
**CN:** 过分关注的，执着于……的

**Original examples:**
- [01:14:05] Why are you so **fixated** on that AI model? That one company? For what reason?  
  你为什么这么执着于那个 AI 模型?那一家公司?到底为什么?

**Extra example:**
- He became **fixated** on the idea that someone was following him.  
  他变得执着于有人在跟踪他这个想法。

### extrapolate  /ɪkˈstræpəleɪt/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to estimate or conclude something by extending known information beyond its original range  
**CN:** 推断，外推

**Original examples:**
- [01:14:31] If you think the base model was here and the backdoor model was here, you can kind of linearly interpolate the weights to adjust the strength of the backdoor, but you can also **extrapolate** it to make the backdoor even stronger.  
  如果你认为基础模型在这里，后门模型在这里，你可以线性插值权重来调整后门的强度，但你也可以外推它来让后门变得更强。

**Extra example:**
- Scientists **extrapolated** future climate trends from current temperature data.  
  科学家们从当前的温度数据外推出未来的气候趋势。

### shortfall  /ˈʃɔːrtfɔːl/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a deficit or lack in the amount needed or expected  
**CN:** 缺口，不足

**Original examples:**
- [01:15:24] So while you're on 1.6nm, they're still going to be on 7nm, and they have to produce enough of it to make up for the **shortfall**.  
  所以当你们用上 1.6nm 时,他们还会停留在 7nm,而且他们必须生产足够多的 7nm 芯片来弥补差距。

**Extra example:**
- The company faced a budget **shortfall** of $2 million this quarter.  
  公司本季度面临200万美元的预算缺口。

### enriched uranium  /ɪnˈrɪtʃt jʊˈreɪniəm/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 4

**EN:** uranium in which the percentage of uranium-235 has been increased, used for nuclear fuel or weapons  
**CN:** 浓缩铀

**Original examples:**
- [01:17:37] But AI is similar to **enriched uranium**, right? It can have positive uses, it can have negative uses.  
  但AI类似于浓缩铀，对吧？它可以有积极的用途，也可以有消极的用途。
- [01:17:44] We still don't want to send **enriched uranium** to other countries.  
  我们仍然不想把浓缩铀送到其他国家。
- [01:23:11] Again, we have more nuclear weapons than anybody else, but we don't want to send **enriched uranium** anywhere.  
  再说一次，我们拥有比任何人都多的核武器，但我们不想把浓缩铀送到任何地方。
- [01:23:20] We're not **enriched uranium**. It's a chip, and it's a chip that they can make themselves.  
  我们不是浓缩铀。这是芯片，而且是他们自己能制造的芯片。

**Extra example:**
- The production of **enriched uranium** is closely monitored by international agencies.  
  浓缩铀的生产受到国际机构的密切监控。

### sticky  /ˈstɪki/
**CEFR:** B2 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** (in business context) tending to retain customers or users; creating strong loyalty or lock-in effects  
**CN:** （商业语境）有黏性的，能留住客户的

**Original examples:**
- [01:20:49] There's a reason why the x86 deal exists. There's a reason why ARM is so **sticky**.  
  x86 协议之所以存在是有原因的。ARM 之所以如此有粘性也是有原因的。

**Extra example:**
- The app's social features make it very **sticky**—users keep coming back daily.  
  这个应用的社交功能让它很有黏性——用户每天都会回来使用。

### hinge  /hɪndʒ/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to depend entirely on something; to be contingent upon  
**CN:** 取决于，依赖于

**Original examples:**
- [01:31:34] Okay. The argument **hinges** on this. They've built models that are specified for the best chips that they make in a few years.  
  好的。这个论点取决于这一点。他们已经构建了专门针对他们几年内制造的最好芯片的模型。

**Extra example:**
- The success of the project **hinges** on securing adequate funding.  
  这个项目的成功取决于获得充足的资金。

### disaggregation  /ˌdɪsæɡrɪˈɡeɪʃən/
**CEFR:** C2 | **Part of speech:** n. | **Occurrences:** 1

**EN:** the process of separating something into its component parts or breaking down a unified system into smaller, independent elements  
**CN:** 分解，拆分(将整体系统分解为独立组成部分的过程)

**Original examples:**
- [01:39:21] I think this idea of extremely premium tokens and just the **disaggregation** of the inference market is very interesting.  
  我觉得这种极高价值 token 的概念,以及推理市场的细分化,都非常有意思。

**Extra example:**
- The **disaggregation** of cloud services allows companies to choose best-of-breed solutions for each component.  
  云服务的**拆分**使公司能够为每个组件选择最佳解决方案。

### distill  /dɪˈstɪl/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to reduce something to its most essential or simplified form, often oversimplifying complex ideas  
**CN:** 提炼，简化(常指过度简化复杂概念)

**Original examples:**
- [01:35:02] It's not simplistic, like the way you're trying to **distill** it.  
  这不是简单化的问题，不像你试图**简化**的那样。

**Extra example:**
- It's difficult to **distill** years of research into a single presentation.  
  很难把多年的研究**浓缩**成一场演讲。

### segment  /ˈseɡmənt/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a distinct part or section of a market, industry, or customer base with specific characteristics  
**CN:** 细分市场，部分(具有特定特征的市场、行业或客户群体的独特部分)

**Original examples:**
- [01:38:46] That's the reason why we decided to expand the Pareto frontier and create a **segment** of inference that is faster response time, even though it's lower throughput.  
  这就是为什么我们决定扩展Pareto前沿，创建一个响应时间更快的推理**细分市场**，尽管吞吐量较低。

**Extra example:**
- The company targets the premium **segment** of the smartphone market.  
  该公司瞄准智能手机市场的高端**细分市场**。

### systolic  /sɪˈstɒlɪk/
**CEFR:** C2 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** relating to a type of array architecture in computing where data flows rhythmically through processing elements in a synchronized manner  
**CN:** 脉动式的(计算机架构中数据以同步方式有节奏地流经处理单元)

**Original examples:**
- [20:19] I'm not in the details, but I talk to my AI researcher friends and they say, "Look, when I use a TPU, it's this big **systolic** array that's perfect for doing matrix multiplies, whereas a GPU is very flexible."  
  我不了解细节，但我和做AI研究的朋友聊过，他们说："你看，当我使用TPU时，它是一个大型**脉动**阵列，非常适合做矩阵乘法，而GPU则非常灵活。"

**Extra example:**
- The **systolic** architecture enables efficient parallel processing of large-scale computations.  
  **脉动式**架构能够高效地并行处理大规模计算。

---

## Useful Phrases

### crash
**Type:** collocation

**EN:** (of prices, values, or markets) fall suddenly and dramatically  
**CN:** (价格、价值或市场)暴跌，崩盘

**Original examples:**
- [00:00] We've seen the valuations of a bunch of software companies **crash** because people are expecting AI to commoditize software.  
  我们看到一大批软件公司的估值暴跌，因为人们预期AI会让软件商品化。

**Extra example:**
- The stock market **crashed** in 2008, wiping out trillions in value.  
  股市在2008年崩盘，蒸发了数万亿的价值。

### in real time
**Type:** collocation

**EN:** as things happen, without delay  
**CN:** 实时地，即时地

**Original examples:**
- [01:05] The amount of artistry, engineering, science, and invention that goes into making that token valuable, obviously we're watching it happen **in real time**.  
  让这个token变得有价值所需的艺术性、工程、科学和发明，显然我们正在实时见证这一切的发生。

**Extra example:**
- The dashboard updates **in real time** so you can monitor performance instantly.  
  仪表盘实时更新，所以你可以即时监控性能。

### turn out
**Type:** phrasal_verb

**EN:** to be discovered to be; to prove to be  
**CN:** 结果是，原来是

**Original examples:**
- [02:34] We try to do as little as possible, but the part that we have to do, as it **turns out**, is insanely hard.  
  我们尽量做得越少越好，但结果是我们必须做的那部分极其困难。

**Extra example:**
- The solution **turned out** to be much simpler than we expected.  
  结果解决方案比我们预期的简单得多。

### skyrocket
**Type:** collocation

**EN:** to increase rapidly and dramatically  
**CN:** 飙升，猛涨

**Original examples:**
- [03:34] It's very likely that the number of instances of all these tools is going to **skyrocket**.  
  这些工具的实例数量很可能会飙升。
- [04:10] I think tool use is going to cause the software companies to **skyrocket**.  
  我认为工具使用会让软件公司飙升。

**Extra example:**
- Housing prices **skyrocketed** after the new tech campus opened.  
  新科技园区开业后，房价飙升。

### lock up
**Type:** phrasal_verb

**EN:** to secure exclusive access to something, making it unavailable to others  
**CN:** 锁定，独占

**Original examples:**
- [04:44] One interpretation is that Nvidia's moat is really that you've **locked up** many years of these scarce components.  
  一种解读是，英伟达的护城河实际上是你们锁定了这些稀缺组件多年的供应。

**Extra example:**
- The company **locked up** all the prime office space in the district.  
  这家公司锁定了该区域所有的优质办公空间。

### swarm
**Type:** collocation

**EN:** to attack or address a problem with overwhelming force or numbers  
**CN:** 集中力量攻克，蜂拥而至解决

**Original examples:**
- [10:05] If we're too far apart, if one particular component is too far away, the industry **swarms** it.  
  如果差距太大，如果某个特定组件供应不足，整个行业就会集中力量攻克它。
- [10:20] The reason for that is because for two years we **swarmed** the living daylights out of it.  
  原因是我们用了两年时间全力攻克它。

**Extra example:**
- When the bug was discovered, the entire team **swarmed** the issue until it was fixed.  
  当发现这个bug时，整个团队集中力量解决这个问题，直到修复为止。

### double down on
**Type:** idiom

**EN:** to strengthen one's commitment to a strategy or course of action  
**CN:** 加倍投入，加大力度

**Literal:** 在赌博中加倍下注  
**Figurative EN:** to increase one's commitment or investment in something, especially when facing risk or uncertainty  
**Figurative CN:** 加倍投入或加大对某事的承诺，尤其是在面临风险或不确定性时

**Original examples:**
- [11:09] I still remember the meeting really well where I was clear about exactly what was going to happen, why it was going to happen, and the predictions of today. They really **doubled down on** it.  
  我还清楚地记得那次会议，我明确说明了将会发生什么、为什么会发生，以及对今天的预测。他们真的加倍投入了。

**Extra example:**
- Despite the criticism, the CEO **doubled down on** the company's AI strategy.  
  尽管受到批评，CEO还是加大了对公司AI战略的投入。

### pinch point
**Type:** collocation

**EN:** a critical bottleneck or constraint in a process or system  
**CN:** 瓶颈，关键制约点

**Original examples:**
- [14:42] We have to think about the critical **pinch points**.  
  我们必须考虑关键的瓶颈点。

**Extra example:**
- Supply chain managers identified three major **pinch points** that were slowing production.  
  供应链管理者识别出三个拖慢生产的主要瓶颈点。

### come up with
**Type:** phrasal_verb

**EN:** to think of or produce (an idea, plan, solution, etc.)  
**CN:** 想出，提出（主意、计划、解决方案等）

**Original examples:**
- [15:19] We're **coming up with** new algorithms because CUDA is so flexible.  
  我们正在想出新的算法，因为CUDA非常灵活。
- [21:07] If you want to **come up with** a new attention mechanism, disaggregate in a different way, or invent a whole new type of architecture altogether—like a hybrid SSM—you want an architecture that's generally programmable.  
  如果你想提出一种新的注意力机制，以不同的方式解耦，或者发明一种全新的架构——比如混合SSM——你需要一个通用可编程的架构。
- [23:15] Without the ability to really get down and **come up with** new kernels with CUDA, it's really hard to do.  
  如果没有能力深入研究并用CUDA提出新的内核，这真的很难做到。
- [01:16:40] Number one, why is it that we don't **come up with** a regulation that's more balanced so that Nvidia can win around the world instead of giving up the world?  
  第一，为什么我们不能提出一个更平衡的监管方案，让英伟达能在全球范围内获胜，而不是放弃全球市场？

**Extra example:**
- The team needs to **come up with** a solution before the deadline.  
  团队需要在截止日期前想出一个解决方案。

### bring back
**Type:** phrasal_verb

**EN:** to restore or reintroduce something that existed before  
**CN:** 恢复，重新引入（以前存在的事物）

**Original examples:**
- [15:52] We want to reindustrialize the United States. We want to **bring back** chip manufacturing, computer manufacturing, and packaging.  
  我们想让美国重新工业化。我们想让芯片制造、计算机制造和封装回归。

**Extra example:**
- The company is trying to **bring back** jobs that were outsourced overseas.  
  公司正试图让那些外包到海外的工作岗位回归。

### sandbag
**Type:** idiom

**EN:** to deliberately understate performance or expectations  
**CN:** 故意低估性能或预期

**Literal:** 用沙袋  
**Figurative EN:** to deliberately set low expectations so that actual results appear more impressive  
**Figurative CN:** 故意设定较低的预期，以便实际结果显得更加令人印象深刻

**Original examples:**
- [22:42] Then Dylan wrote an article saying I **sandbagged**, and it's actually fifty times.  
  然后Dylan写了一篇文章说我故意低估了，实际上是五十倍。

**Extra example:**
- The CEO **sandbagged** the quarterly forecast to make the actual results look better.  
  CEO故意低估了季度预测，以使实际结果看起来更好。

### get down
**Type:** phrasal_verb

**EN:** to focus seriously on something; to work on something in detail  
**CN:** 认真专注于某事；深入研究某事

**Original examples:**
- [23:15] Without the ability to really **get down** and come up with new kernels with CUDA, it's really hard to do.  
  如果没有能力真正深入研究并用CUDA提出新的内核，这真的很难做到。

**Extra example:**
- Let's **get down** to business and discuss the budget.  
  让我们言归正传，讨论预算吧。

### wrung out
**Type:** collocation

**EN:** thoroughly tested and debugged; refined through extensive use  
**CN:** 经过彻底测试和调试；通过广泛使用而得到完善

**Original examples:**
- [27:17] Obviously, we still have lots of bugs ourselves, but our system is so well **wrung out** that you can at least build on top of the foundation.  
  显然，我们自己仍然有很多bug，但我们的系统经过了如此充分的测试，你至少可以在这个基础上进行构建。

**Extra example:**
- After years of production use, the codebase is well **wrung out** and stable.  
  经过多年的生产使用，代码库已经非常成熟稳定。

### install base
**Type:** collocation

**EN:** the total number of units of a product or system currently in use  
**CN:** 当前正在使用的产品或系统的总数量

**Original examples:**
- [27:34] The second thing is, if you're a developer building anything at all, the single most important thing you want is an **install base**.  
  第二件事是，如果你是一个开发者，无论构建什么，你最想要的就是安装基数。
- [29:01] The combination of the richness of the ecosystem, the expansiveness of the **install base**, and the versatility of where we are makes CUDA invaluable.  
  生态系统的丰富性、安装基数的广泛性以及我们所在位置的多样性，这些结合起来使CUDA变得无价。

**Extra example:**
- Apple's large **install base** gives developers confidence to build iOS apps.  
  苹果庞大的安装基数让开发者有信心开发iOS应用。

### cruise control
**Type:** collocation

**EN:** a state of easy, automatic operation requiring minimal effort  
**CN:** 轻松、自动运行的状态，只需最少的努力

**Original examples:**
- [30:54] A CPU is kind of like a Cadillac. It's a nice cruiser. It never goes too fast. Everybody drives it pretty well. It's got **cruise control**, and everything's easy.  
  CPU有点像凯迪拉克。它是一辆不错的巡航车。它从不开得太快。每个人都能很好地驾驶它。它有定速巡航，一切都很容易。

**Extra example:**
- After the initial setup, the project was on **cruise control** and required little management.  
  初始设置完成后，项目进入了自动巡航状态，几乎不需要管理。

### push to the limit
**Type:** collocation

**EN:** to use something to its maximum capacity or performance  
**CN:** 将某物用到极限或最大性能

**Original examples:**
- [31:18] I could imagine everybody's able to drive it at a hundred miles an hour, but it takes quite a bit of expertise to be able to **push it to the limit**.  
  我可以想象每个人都能以一百英里的时速驾驶它，但要把它推到极限需要相当多的专业知识。

**Extra example:**
- Professional athletes train daily to **push their bodies to the limit**.  
  职业运动员每天训练以将身体推向极限。

### bar none
**Type:** idiom

**EN:** with no exceptions; better than all others  
**CN:** 毫无例外；无可匹敌

**Literal:** 没有任何阻拦  
**Figurative EN:** without any exception; the absolute best  
**Figurative CN:** 毫无例外；绝对最好的

**Original examples:**
- [32:16] Nvidia's computing stack is the best performance per TCO in the world, **bar none**.  
  英伟达的计算堆栈是世界上性能与总拥有成本比最好的，毫无例外。

**Extra example:**
- She is the best programmer on the team, **bar none**.  
  她是团队里最好的程序员，毫无例外。

### show up
**Type:** phrasal_verb

**EN:** to appear or arrive; to participate or compete  
**CN:** 出现；参加；应战

**Original examples:**
- [32:55] Nobody wants to **show up**.  
  没人愿意来应战。
- [45:41] If I didn't do it, somebody would **show up**.  
  如果我不做，总会有人出现来做。

**Extra example:**
- If you want to challenge us, **show up** with your best product.  
  如果你想挑战我们，就拿出你最好的产品来应战。

### on first principles
**Type:** collocation

**EN:** based on fundamental truths or basic reasoning  
**CN:** 基于第一性原理；从基本原理出发

**Original examples:**
- [33:10] **On first principles**, it makes no sense.  
  从第一性原理来看，这完全说不通。

**Extra example:**
- Let's think about this problem **on first principles** rather than following convention.  
  让我们从第一性原理来思考这个问题，而不是遵循惯例。

### square (something)
**Type:** phrasal_verb

**EN:** to reconcile or make consistent; to resolve a contradiction  
**CN:** 使一致；调和矛盾

**Original examples:**
- [36:57] So I'm curious how to **square**, if these things are true on paper, why are they going with other accelerators?  
  所以我很好奇如何调和这个矛盾，如果这些在纸面上都是真的，为什么他们还要选择其他加速器？

**Extra example:**
- I can't **square** his claims with the actual data we're seeing.  
  我无法让他的说法与我们看到的实际数据相一致。

### crank out
**Type:** phrasal_verb

**EN:** to produce something quickly and in large quantities  
**CN:** 快速大量生产；源源不断地产出

**Original examples:**
- [38:32] We're the only company in the world that's **cranking it out** every single year.  
  我们是世界上唯一一家每年都在源源不断产出的公司。

**Extra example:**
- The factory **cranks out** thousands of units per day.  
  这家工厂每天生产数千件产品。

### turn out to be
**Type:** phrasal_verb

**EN:** to become or prove to be in the end  
**CN:** 结果是；最终成为

**Original examples:**
- [40:00] A VC would never put in $5-10 billion of investment into an AI lab with the hopes of it **turning out to be** Anthropic.  
  风投永远不会向一个AI实验室投入50-100亿美元，希望它最终成为Anthropic。

**Extra example:**
- The project **turned out to be** much more complex than we anticipated.  
  这个项目结果比我们预期的要复杂得多。

### pop up
**Type:** phrasal_verb

**EN:** to appear or emerge suddenly  
**CN:** 突然出现；冒出来

**Original examples:**
- [43:39] There's one answer which is that there's this whole middleman ecosystem that has **popped up** for converting CapEx into OpEx for these labs.  
  有一个答案是，已经冒出了一整个中间商生态系统，为这些实验室将资本支出转换为运营支出。

**Extra example:**
- New startups keep **popping up** in the AI space every month.  
  AI领域每个月都有新的创业公司冒出来。

### as much as needed, as little as possible
**Type:** collocation

**EN:** doing only what is necessary, nothing more  
**CN:** 做必要的事，尽可能少做；恰到好处

**Original examples:**
- [44:18] We should do **as much as needed, as little as possible**.  
  我们应该做必要的事，尽可能少做。

**Extra example:**
- Our philosophy is to intervene **as much as needed, as little as possible**.  
  我们的理念是尽可能少干预，但该干预时必须干预。

### get off the ground
**Type:** idiom

**EN:** to start successfully; to begin to function or operate  
**CN:** 成功启动；开始运作

**Literal:** 离开地面  
**Figurative EN:** to start a project or business successfully and begin operating  
**Figurative CN:** 使项目或企业成功启动并开始运营

**Original examples:**
- [49:01] But if, at the end of the day, they need some investment in order to **get it off the ground**, we would be there for them.  
  但归根结底，如果他们需要一些投资才能启动，我们会支持他们。

**Extra example:**
- The startup struggled to **get off the ground** without proper funding.  
  这家初创公司没有适当的资金很难启动。

### have the wind at one's back
**Type:** idiom

**EN:** to have favorable conditions or momentum for success  
**CN:** 处于有利形势；顺风顺水

**Literal:** 背后有风  
**Figurative EN:** to have favorable circumstances or momentum that helps one succeed  
**Figurative CN:** 拥有有利的环境或势头来帮助成功

**Original examples:**
- [50:02] They **have the wind at their back**.  
  他们现在势头正盛。

**Extra example:**
- With strong investor support, the company **has the wind at its back**.  
  有了强大的投资者支持，这家公司势头正盛。

### divvy up
**Type:** phrasal_verb

**EN:** to divide and distribute something among people  
**CN:** 分配，分摊

**Original examples:**
- [51:25] Nvidia is known for **divvying up** the scarce allocation, not just based on high bidder.  
  英伟达以分配稀缺资源而闻名，不仅仅基于出价最高者。

**Extra example:**
- Let's **divvy up** the work equally among the team members.  
  让我们在团队成员之间平均分配工作。

### stand up
**Type:** phrasal_verb

**EN:** to set up, establish, or make operational (especially infrastructure or systems)  
**CN:** 建立，搭建（尤指基础设施或系统）

**Original examples:**
- [52:52] if you're not ready because your data center's not ready, or certain components aren't ready to enable you to **stand up** a data center, we might decide to serve another customer first.  
  如果你还没准备好，因为你的数据中心还没准备好，或者某些组件还不足以让你搭建数据中心，我们可能会决定先服务另一个客户。

**Extra example:**
- It took us three months to **stand up** the new cloud infrastructure.  
  我们花了三个月时间搭建新的云基础设施。

### bet the farm
**Type:** idiom

**EN:** to risk everything on something; to make a very large commitment  
**CN:** 孤注一掷；押上全部

**Literal:** 赌上农场  
**Figurative EN:** to risk everything you have on a single venture or decision  
**Figurative CN:** 把所有的一切都押在一个项目或决定上

**Original examples:**
- [55:57] I can **bet the farm**, I can bet my entire business that you will be here for me every single year.  
  我可以孤注一掷，可以押上我的整个生意，相信你每年都会在这里支持我。

**Extra example:**
- I wouldn't **bet the farm** on this investment - it's too risky.  
  我不会在这项投资上孤注一掷——风险太大了。

### play devil's advocate
**Type:** idiom

**EN:** to argue against an idea or position for the sake of debate, even if you don't necessarily disagree  
**CN:** 故意唱反调；为辩论而反驳

**Literal:** 扮演魔鬼的代言人  
**Figurative EN:** to present a counterargument or opposing view to test the strength of an idea, not because you oppose it but to stimulate discussion  
**Figurative CN:** 提出反对意见或相反观点来测试想法的强度，不是因为反对而是为了激发讨论

**Original examples:**
- [57:38] I actually don't know what I think about whether it's good to sell chips to China or not, but I like to **play devil's advocate** against my guests.  
  我其实不知道向中国出售芯片是好是坏，但我喜欢对我的嘉宾唱反调。

**Extra example:**
- Let me **play devil's advocate** - what if this strategy backfires?  
  让我唱个反调——如果这个策略适得其反怎么办？

### keep after
**Type:** phrasal_verb

**EN:** to monitor, supervise, or watch over something/someone continuously  
**CN:** 持续监督，看管

**Original examples:**
- [01:01:46] you're going to have an AI agent that's incredible, surrounded by thousands of AI agents, **keeping it** safe, **keeping it** secure.  
  你会有一个令人难以置信的AI代理，被成千上万的AI代理包围着，保护它的安全。

**Extra example:**
- We need security systems **keeping after** our network 24/7.  
  我们需要安全系统全天候监控我们的网络。

### hold onto
**Type:** phrasal_verb

**EN:** to keep something and not release or share it  
**CN:** 保留，持有（不释放或分享）

**Original examples:**
- [01:04:12] Okay, we're going to **hold onto** it for a month while all these American companies, we'll give them access to it.  
  好的，我们会保留一个月，同时让所有这些美国公司都能访问它。

**Extra example:**
- The company decided to **hold onto** the technology until they could patent it.  
  公司决定保留这项技术，直到他们能够申请专利。

### patch up
**Type:** phrasal_verb

**EN:** to repair or fix something, especially temporarily  
**CN:** 修补，修复（尤指临时性的）

**Original examples:**
- [01:04:12] They're going to **patch up** all their vulnerabilities, and now we release it.  
  他们会修补所有的漏洞，然后我们再发布它。

**Extra example:**
- We need to **patch up** these security holes before the product launch.  
  我们需要在产品发布前修补这些安全漏洞。

### gang up
**Type:** phrasal_verb

**EN:** to combine or join together (in this context, to connect multiple units)  
**CN:** 联合起来，组合在一起（在此语境中指连接多个单元）

**Original examples:**
- [01:06:14] They just **gang up** more chips, even if they're 7nm.  
  他们只需把更多芯片组合起来，即使是7纳米的。
- [01:09:04] You could **gang them** together, just like we **gang them** together with NVL72.  
  你可以把它们组合在一起，就像我们用NVL72把它们组合在一起一样。

**Extra example:**
- They **ganged up** multiple servers to handle the increased traffic.  
  他们把多台服务器组合起来以应对增加的流量。

### make up for
**Type:** phrasal_verb

**EN:** to compensate for something, to provide an equivalent substitute  
**CN:** 弥补，补偿

**Original examples:**
- [01:07:04] When you have an abundance of energy, it **makes up for** chips.  
  当你有充足的能源时，它可以弥补芯片的不足。
- [01:07:10] If you have an abundance of chips, it **makes up for** energy.  
  如果你有充足的芯片，它可以弥补能源的不足。
- [01:15:34] They have to produce enough of it to **make up for** the shortfall.  
  他们必须生产足够的量来弥补短缺。
- [01:39:05] even though the throughput is lower in the factory, the ASPs **make up for** it.  
  尽管工厂的吞吐量较低，但平均售价弥补了这一点。

**Extra example:**
- His enthusiasm **makes up for** his lack of experience.  
  他的热情弥补了经验的不足。

### off the charts
**Type:** idiom

**EN:** extremely high or exceptional, beyond normal measurement  
**CN:** 极高的，超出常规的

**Literal:** 超出图表范围  
**Figurative EN:** extremely high, exceptional, or beyond normal levels  
**Figurative CN:** 极高的，异常的，超出正常水平

**Original examples:**
- [01:07:35] Our throughput per watt is **off the charts**.  
  我们每瓦的吞吐量高得惊人。

**Extra example:**
- The demand for their new product was **off the charts**.  
  他们新产品的需求量高得惊人。

### out of the box
**Type:** idiom

**EN:** immediately available without modification or setup  
**CN:** 开箱即用，无需修改或设置

**Literal:** 从盒子里拿出来  
**Figurative EN:** ready to use immediately without any additional setup or modification  
**Figurative CN:** 立即可用，无需任何额外设置或修改

**Original examples:**
- [01:12:02] Coming **out of the box**, if all of the AI models run best on somebody else's tech stack, you've got to be arguing some ridiculous claim right now that that's a good thing for the United States.  
  开箱即用的情况下，如果所有AI模型在别人的技术栈上运行得最好，你现在一定是在争辩说这对美国是件好事，这太荒谬了。

**Extra example:**
- This software works perfectly **out of the box** - no configuration needed.  
  这个软件开箱即用——不需要任何配置。

### step back
**Type:** phrasal_verb

**EN:** to pause and reconsider from a broader perspective  
**CN:** 退一步（重新审视），从更宏观的角度考虑

**Original examples:**
- [01:15:15] Okay, **stepping back**, it has to be the case that China is able to build enough 7nm capacity.  
  好的，退一步说，中国必须能够建立足够的7纳米产能。

**Extra example:**
- Let's **step back** and look at the bigger picture before making a decision.  
  让我们退一步，在做决定之前看看更大的格局。

### speak in absolutes
**Type:** collocation

**EN:** to make statements that are completely certain without acknowledging nuance or exceptions  
**CN:** 说话绝对化，不承认细微差别或例外情况

**Original examples:**
- [01:15:48] Listen, I just think you **speak in absolutes**.  
  听着，我只是觉得你说话太绝对了。

**Extra example:**
- Be careful not to **speak in absolutes** when discussing complex policy issues.  
  在讨论复杂的政策问题时，注意不要说话太绝对。

### give up
**Type:** phrasal_verb

**EN:** to stop trying or to surrender something  
**CN:** 放弃，停止尝试或交出某物

**Original examples:**
- [01:16:56] Why would you want the United States to **give up** the world?  
  你为什么希望美国放弃全球市场？
- [01:17:08] Why is it that your policy, your philosophy, leads to the United States **giving up** a vast part of the world's market?  
  为什么你的政策、你的理念会导致美国放弃全球市场的很大一部分？
- [01:19:40] The United States should not **give that up**.  
  美国不应该放弃这一点。

**Extra example:**
- Don't **give up** on your dreams just because of one setback.  
  不要因为一次挫折就放弃你的梦想。

### the crux comes down to
**Type:** collocation

**EN:** the most important or decisive point of something  
**CN:** 关键在于，核心问题归结为

**Original examples:**
- [01:19:10] I guess then **the crux comes down to**, how does selling them chips now help us win in the long term?  
  我想关键问题在于，现在向他们出售芯片如何帮助我们长期获胜？

**Extra example:**
- **The crux comes down to** whether we can secure funding before the deadline.  
  关键问题在于我们能否在截止日期前获得资金。

### wake up a loser
**Type:** idiom

**EN:** to have a defeatist mindset or attitude from the start  
**CN:** 一开始就抱有失败者心态

**Literal:** 醒来时是个失败者  
**Figurative EN:** to begin with a defeatist attitude or mentality  
**Figurative CN:** 从一开始就抱有失败者的心态或态度

**Original examples:**
- [01:20:12] You're not talking to somebody who **woke up a loser**.  
  你面对的不是一个一开始就抱有失败者心态的人。

**Extra example:**
- I didn't **wake up a loser** today - I'm going to fight for this deal.  
  我今天不是带着失败者心态醒来的——我会为这笔交易而战。

### move on
**Type:** phrasal_verb

**EN:** to stop discussing something and start talking about something else  
**CN:** 继续，转到下一个话题

**Original examples:**
- [01:21:25] Okay. I'll **move on**. I just want to make sure that—  
  好的。我会继续下一个话题。我只是想确保——
- [01:21:30] You don't have to **move on**. I'm enjoying it.  
  你不必转到下一个话题。我很享受这个讨论。

**Extra example:**
- Let's **move on** to the next agenda item.  
  让我们继续讨论下一个议程项目。

### go to extremes
**Type:** collocation

**EN:** to take an extreme position or use extreme arguments  
**CN:** 走极端，采取极端立场

**Original examples:**
- [01:21:42] The crux is you're **going to extremes**. Your argument starts from extremes.  
  关键是你在走极端。你的论点从极端开始。

**Extra example:**
- There's no need to **go to extremes** - we can find a balanced solution.  
  没必要走极端——我们可以找到一个平衡的解决方案。

### on balance
**Type:** collocation

**EN:** considering everything; overall  
**CN:** 总的来说，权衡之下

**Original examples:**
- [01:23:28] Because our chips are better. **On balance**, our chips are better.  
  因为我们的芯片更好。总的来说，我们的芯片更好。

**Extra example:**
- **On balance**, the benefits of the new policy outweigh the costs.  
  总的来说，新政策的好处大于成本。

### diffuse out
**Type:** phrasal_verb

**EN:** to spread gradually into a wider area or population  
**CN:** 扩散，传播到更广泛的区域

**Original examples:**
- [01:25:07] We get the benefit, as those AI models **diffuse out** into the rest of the world, that the American tech stack is therefore the best for it.  
  当这些AI模型扩散到世界其他地方时，我们会获得好处，因为美国的技术栈是最适合它的。
- [01:28:14] The layer that **diffuses into** society, the one that uses it most will benefit from this industrial revolution most.  
  扩散到社会中的那一层，使用最多的人将从这场工业革命中获益最多。

**Extra example:**
- New technologies tend to **diffuse out** from urban centers to rural areas.  
  新技术往往从城市中心扩散到农村地区。

### hinge on
**Type:** phrasal_verb

**EN:** to depend on something completely  
**CN:** 取决于，依赖于

**Original examples:**
- [01:31:34] The argument **hinges on** this.  
  这个论点取决于这一点。

**Extra example:**
- The success of the project **hinges on** securing adequate funding.  
  项目的成功取决于能否获得充足的资金。

### lean forward
**Type:** phrasal_verb

**EN:** to be proactive, to take initiative or advance aggressively  
**CN:** 积极主动，向前推进

**Original examples:**
- [01:36:23] We could afford to **lean forward**.  
  我们有能力积极推进。

**Extra example:**
- In this competitive market, we need to **lean forward** and innovate constantly.  
  在这个竞争激烈的市场中，我们需要积极主动，不断创新。

### in a heartbeat
**Type:** idiom

**EN:** immediately, without hesitation  
**CN:** 立刻，毫不犹豫地

**Literal:** 在一次心跳的时间内  
**Figurative EN:** instantly, without any delay or second thought  
**Figurative CN:** 立即，毫不迟疑地

**Original examples:**
- [01:36:26] Would I go back and use 7nm? **In a heartbeat**, of course I would.  
  我会回去使用7纳米吗？当然会，毫不犹豫。

**Extra example:**
- If they offered me that job, I'd accept **in a heartbeat**.  
  如果他们给我那份工作，我会立刻接受。

### put all the eggs in one basket
**Type:** idiom

**EN:** to risk everything on a single venture  
**CN:** 孤注一掷，把所有希望寄托在一件事上

**Literal:** 把所有鸡蛋放在一个篮子里  
**Figurative EN:** to invest all resources or efforts in a single option, risking total loss  
**Figurative CN:** 把全部资源或精力投入到单一选项上，冒着全盘皆输的风险

**Original examples:**
- [01:36:42] So why **put all the eggs in one basket**, given who knows where AI might go and architectures might go?  
  既然谁也不知道AI和架构会走向何方，为什么要孤注一掷呢？

**Extra example:**
- Don't **put all your eggs in one basket** - diversify your investments.  
  不要孤注一掷——分散你的投资。

### offload onto
**Type:** phrasal_verb

**EN:** to transfer work or responsibility to another system or person  
**CN:** 转移（工作或责任）到

**Original examples:**
- [01:40:12] Different kernels of code or algorithms could be **offloaded onto** our GPU.  
  不同的代码内核或算法可以转移到我们的GPU上。

**Extra example:**
- We can **offload** routine tasks **onto** automated systems.  
  我们可以把日常任务转移到自动化系统上。

### run its course
**Type:** idiom

**EN:** to continue until finished or exhausted naturally  
**CN:** 走完全程，自然结束

**Literal:** 跑完它的路程  
**Figurative EN:** to proceed to its natural conclusion or end without interference  
**Figurative CN:** 自然发展到结束，走到尽头

**Original examples:**
- [01:40:42] the ability for general purpose computing to continue to scale has largely **run its course**.  
  通用计算继续扩展的能力基本上已经走到尽头了。

**Extra example:**
- The pandemic will eventually **run its course** and life will return to normal.  
  疫情最终会自然结束，生活会恢复正常。

### break through
**Type:** phrasal_verb

**EN:** to make an important discovery or achieve success after difficulty  
**CN:** 突破，取得突破性进展

**Original examples:**
- [01:41:14] scale to the level of capability that helps **break through** certain fields of science.  
  扩展到有助于在某些科学领域取得突破的能力水平。

**Extra example:**
- After years of research, they finally **broke through** with a new treatment.  
  经过多年研究，他们终于在新疗法上取得了突破。

---

## Complex Sentences

### [01:05]
**Original:** Making that token is like making one molecule more valuable than another molecule, making one token more valuable than another.

**Translation:** 制造那个token就像让一个分子比另一个分子更有价值，让一个token比另一个token更有价值。

**Core structure:**
- Making that token is like making one molecule more valuable.  
  制造token就像让一个分子更有价值。

**Structure tree:**
```
main clause: Making that token is like...
gerund subject: Making that token
comparative structure: like making X more valuable than Y
parallel structure: making one molecule... + making one token...
```

**Grammar points:**
- **动名词作主语** - Making that token 整体作主语
- **比较级 + than 结构** - more valuable than 表示比较
- **平行结构** - 两个making短语并列，强化类比

### [01:50]
**Original:** Our job is to do as much as necessary and as little as possible to enable that transformation to be done at incredible capabilities.

**Translation:** 我们的工作是做必要的尽可能多，同时做尽可能少的事情，以使那个转换能够以令人难以置信的能力完成。

**Core structure:**
- Our job is to do as much as necessary and as little as possible.  
  我们的工作是做必要的尽可能多，同时做尽可能少。

**Structure tree:**
```
main clause: Our job is to do...
parallel infinitives: to do as much as necessary AND as little as possible
purpose clause: to enable that transformation
passive infinitive: to be done at incredible capabilities
```

**Grammar points:**
- **as...as 比较结构的对比** - as much as necessary 和 as little as possible 形成对比
- **不定式表目的** - to enable 说明做这些事的目的
- **被动不定式** - to be done 表示被动的动作

### [05:14]
**Original:** Some of it is implicit. For example, a lot of the investments that are upstream are made by our supply chain because I said to the CEOs, "Let me tell you how big this industry is going to be, let me explain to you why, let me reason through it with you, and let me show you what I see."

**Translation:** 其中一些是隐性的。例如，很多上游的投资是由我们的供应链做出的，因为我对CEO们说：'让我告诉你这个行业将会有多大，让我向你解释原因，让我和你一起推理，让我向你展示我所看到的。'

**Core structure:**
- Investments are made by our supply chain because I said to the CEOs...  
  投资是由我们的供应链做出的，因为我对CEO们说...

**Structure tree:**
```
main clause: investments are made by supply chain
reason clause: because I said to the CEOs
direct speech: four parallel imperative clauses (Let me...)
relative clause: that are upstream (modifying investments)
```

**Grammar points:**
- **被动语态** - are made by 强调动作的执行者
- **直接引语中的祈使句并列** - 四个Let me...结构并列，表达说服过程

### [05:51]
**Original:** The fact is that Nvidia's downstream supply chain and our downstream demand is so large, they're willing to make the investment upstream.

**Translation:** 事实是，英伟达的下游供应链和我们的下游需求是如此之大，以至于他们愿意在上游进行投资。

**Core structure:**
- The fact is that demand is so large, they're willing to invest.  
  事实是需求如此之大，他们愿意投资。

**Structure tree:**
```
main clause: The fact is that...
predicative clause: that supply chain and demand is so large
result clause (implicit): they're willing to make the investment
so...that structure (split by comma)
```

**Grammar points:**
- **表语从句** - that从句作表语，说明fact的内容
- **so...that 因果结构** - so large 导致后面的结果，但that被省略，用逗号连接

### [07:22]
**Original:** I need to make sure the entire supply chain, upstream and downstream, the ecosystem, understands what is coming at us, why it's coming,

**Translation:** 我需要确保整个供应链——上游和下游、整个生态系统——理解即将发生什么、为什么会发生，

**Core structure:**
- I need to make sure the supply chain understands what is coming.  
  我需要确保供应链理解即将发生什么。

**Structure tree:**
```
main clause: I need to make sure...
object clause: the supply chain understands...
appositive insertions: upstream and downstream, the ecosystem
embedded wh-clauses: what is coming, why it's coming
```

**Grammar points:**
- **make sure + that从句** - 确保某事发生，that常省略
- **同位语插入** - upstream and downstream, the ecosystem 解释 supply chain
- **多重宾语从句** - understands后接多个wh-从句作宾语

### [07:35]
**Original:** When it's coming, how big it's going to be, and is able to reason about it systematically, just like I reason about it.

**Translation:** 什么时候会发生、规模会有多大，并且能够像我一样系统地推理这件事。

**Core structure:**
- When it's coming, how big it's going to be, and is able to reason about it.  
  什么时候发生、规模多大，并能推理它。

**Structure tree:**
```
parallel wh-clauses: When it's coming, how big it's going to be
coordinated predicate: and is able to reason...
comparison: just like I reason about it
```

**Grammar points:**
- **并列宾语从句** - 多个wh-从句并列作understands的宾语
- **省略主语的并列谓语** - is able to与前面understands共享主语

### [09:25]
**Original:** At some level, the instantaneous demand is greater than the supply upstream and downstream in the world.

**Translation:** 在某种程度上，瞬时需求大于世界上游和下游的供应。

**Core structure:**
- The demand is greater than the supply.  
  需求大于供应。

**Structure tree:**
```
prepositional phrase: At some level
main clause: demand is greater than supply
modifiers: instantaneous (demand), upstream and downstream in the world (supply)
```

**Grammar points:**
- **比较结构** - greater than连接比较对象
- **后置修饰** - upstream and downstream修饰supply

### [10:30]
**Original:** TSMC now knows that CoWoS supply has to keep up with the rest of the logic demand and the memory demand.

**Translation:** 台积电现在知道CoWoS供应必须跟上其余的逻辑需求和内存需求。

**Core structure:**
- TSMC knows that supply has to keep up with demand.  
  台积电知道供应必须跟上需求。

**Structure tree:**
```
main clause: TSMC knows that...
object clause: CoWoS supply has to keep up with...
parallel objects: the logic demand and the memory demand
```

**Grammar points:**
- **have to + 动词原形** - 表示必须、不得不
- **keep up with** - 跟上、赶上（速度或水平）

### [11:09]
**Original:** At the beginning of the AI revolution, all the things that I say now, I was saying five years ago.

**Translation:** 在人工智能革命开始时，我现在说的所有这些事情，我五年前就在说了。

**Core structure:**
- I was saying all the things five years ago.  
  我五年前就在说所有这些事情。

**Structure tree:**
```
time phrase: At the beginning of the AI revolution
fronted object: all the things that I say now
relative clause: that I say now
main clause: I was saying [them] five years ago
```

**Grammar points:**
- **宾语前置** - all the things提前强调，主句中省略宾语
- **定语从句** - that I say now修饰things
- **过去进行时** - was saying表示过去持续的动作

### [12:23]
**Original:** We built up an entire supply chain around TSMC. We partnered with them on COUPE, invented a whole bunch of technology, and licensed those patents to the supply chain to keep it nice and open.

**Translation:** 我们围绕台积电建立了整个供应链。我们与他们在COUPE上合作,发明了一大堆技术,并将这些专利授权给供应链以保持其良好和开放。

**Core structure:**
- We partnered, invented, and licensed patents to keep it open.  
  我们合作、发明并授权专利以保持开放。

**Structure tree:**
```
main clause: We partnered... invented... and licensed...
parallel verbs: partnered / invented / licensed
purpose clause: to keep it nice and open
```

**Grammar points:**
- **并列动词结构** - 三个过去式动词并列,共享主语We
- **不定式表目的** - to keep表示授权专利的目的

### [12:42]
**Original:** New testing equipment like double-sided probing, investing in companies, and helping them scale up their capacity. You can see that we're trying to shape the ecosystem so that the supply chain is ready to support the scale.

**Translation:** 新的测试设备如双面探测、投资公司并帮助它们扩大产能。你可以看到我们正在努力塑造生态系统,以便供应链准备好支持这种规模。

**Core structure:**
- We're trying to shape the ecosystem so that the supply chain is ready.  
  我们正在努力塑造生态系统,以便供应链准备就绪。

**Structure tree:**
```
main clause: You can see that...
object clause: we're trying to shape the ecosystem
purpose clause: so that the supply chain is ready
infinitive phrase: to support the scale
```

**Grammar points:**
- **so that引导目的状语从句** - 表示塑造生态系统的目的
- **be ready to do** - 准备好做某事的固定搭配

### [13:43]
**Original:** You might hear some of those videos still on the web saying radiology is going to be the first career to go and the world is not going to need any more radiologists.

**Translation:** 你可能仍然会在网上听到一些视频说,放射科将是第一个消失的职业,世界将不再需要更多的放射科医生。

**Core structure:**
- You might hear videos saying radiology is going to go.  
  你可能会听到视频说放射科将会消失。

**Structure tree:**
```
main clause: You might hear videos
participle phrase: saying...
indirect speech: radiology is going to be... and the world is not going to need...
```

**Grammar points:**
- **现在分词作定语** - saying修饰videos,表示视频的内容
- **间接引语中的并列句** - saying后接两个并列的陈述句

### [14:42]
**Original:** Do you go to ASML and say, 'Hey, if I look out three years from now, for Nvidia to be generating two trillion a year in revenue, we need way more EUV machines'?

**Translation:** 你会去找ASML说,'嘿,如果我展望三年后,为了让英伟达每年产生两万亿的收入,我们需要更多的EUV机器'吗?

**Core structure:**
- Do you go to ASML and say we need more machines?  
  你会去找ASML说我们需要更多机器吗?

**Structure tree:**
```
main clause: Do you go... and say...
direct speech: Hey, if I look out... we need...
conditional clause: if I look out three years
purpose phrase: for Nvidia to be generating...
```

**Grammar points:**
- **for sb to do结构** - 表示目的或条件,这里说明需要机器的原因
- **条件句嵌套在直接引语中** - if从句作为说话内容的一部分

### [15:04]
**Original:** My point is that none of the bottlenecks last longer than a couple of years, two, three years, none of them. Meanwhile, we're improving computing efficiency by 10x, 20x, and in the case of Hopper to Blackwell, 30x to 50x.

**Translation:** 我的观点是,没有一个瓶颈会持续超过几年,两三年,一个都没有。与此同时,我们正在将计算效率提高10倍、20倍,而在Hopper到Blackwell的情况下,提高30到50倍。

**Core structure:**
- None of the bottlenecks last longer than a couple of years. We're improving efficiency by 10x, 20x.  
  没有瓶颈会持续超过几年。我们正在将效率提高10倍、20倍。

**Structure tree:**
```
main clause 1: My point is that...
predicative clause: none of the bottlenecks last...
main clause 2: Meanwhile, we're improving...
prepositional phrase: in the case of Hopper to Blackwell
```

**Grammar points:**
- **none of + 复数名词** - 表示全部否定,谓语动词用单数
- **in the case of** - 就...而言,用于引出特定例子
- **倍数表达法** - by + 倍数表示增长幅度

### [18:08]
**Original:** With most of these home-built systems, you have to be your own operator because they were never designed to be flexible enough for others to operate.

**Translation:** 对于大多数这些自建系统,你必须自己做运营商,因为它们从未被设计得足够灵活以供他人操作。

**Core structure:**
- You have to be your own operator because they were never designed to be flexible enough.  
  你必须自己做运营商,因为它们从未被设计得足够灵活。

**Structure tree:**
```
main clause: you have to be your own operator
reason clause: because they were never designed...
adjective phrase: flexible enough for others to operate
prepositional phrase: With most of these home-built systems
```

**Grammar points:**
- **enough + for sb to do** - 表示程度足以让某人做某事
- **被动语态 + 否定副词** - were never designed 强调从未被设计用于某目的

### [18:55]
**Original:** And because we can enable operators in any company and any industry, you could use it to build a supercomputer for scientific research and drug discovery at Lilly.

**Translation:** 因为我们可以为任何公司和任何行业的运营商提供支持,你可以用它来为礼来公司建造一台用于科学研究和药物发现的超级计算机。

**Core structure:**
- Because we can enable operators, you could use it to build a supercomputer.  
  因为我们可以支持运营商,你可以用它来建造超级计算机。

**Structure tree:**
```
reason clause: because we can enable operators...
main clause: you could use it to build...
purpose phrase: to build a supercomputer
modifier: for scientific research and drug discovery
location: at Lilly
```

**Grammar points:**
- **use sth to do** - 使用某物来做某事
- **并列介词短语** - for scientific research and drug discovery 表示用途

### [19:10]
**Original:** We can help them operate their own supercomputer and use it for the entire diversity of drug discovery and biological sciences that we accelerate.

**Translation:** 我们可以帮助他们运营自己的超级计算机,并将其用于我们加速的整个药物发现和生物科学的多样性领域。

**Core structure:**
- We can help them operate their supercomputer and use it for drug discovery and biological sciences.  
  我们可以帮助他们运营超级计算机并将其用于药物发现和生物科学。

**Structure tree:**
```
main clause: We can help them...
parallel verbs: operate...and use...
object phrase: the entire diversity of...
relative clause: that we accelerate
```

**Grammar points:**
- **help sb do sth** - 帮助某人做某事,省略 to
- **限制性定语从句** - that we accelerate 修饰前面的领域

### [21:07]
**Original:** If you want to come up with a new attention mechanism, disaggregate in a different way, or invent a whole new type of architecture altogether—like a hybrid SSM—you want an architecture that's generally programmable.

**Translation:** 如果你想提出一种新的注意力机制、以不同方式解耦,或者完全发明一种全新类型的架构——比如混合SSM——你需要一个通用可编程的架构。

**Core structure:**
- If you want to invent a new architecture, you want an architecture that's programmable.  
  如果你想发明新架构,你需要一个可编程的架构。

**Structure tree:**
```
condition clause: If you want to...
parallel verbs: come up with...disaggregate...or invent...
parenthetical: —like a hybrid SSM—
main clause: you want an architecture
relative clause: that's generally programmable
```

**Grammar points:**
- **并列动词短语** - 三个动词短语用逗号和 or 连接
- **破折号插入语** - —like a hybrid SSM— 补充举例,可跳过不影响主句理解

### [23:15]
**Original:** It's the combination of the programmability of our architecture and the fact that Nvidia is an extreme co-design company.

**Translation:** 这是我们架构的可编程性与英伟达是一家极致协同设计公司这一事实的结合。

**Core structure:**
- It's the combination of the programmability and the fact that Nvidia is a co-design company.  
  这是可编程性与英伟达是协同设计公司这一事实的结合。

**Structure tree:**
```
main clause: It's the combination of...
parallel noun phrases: the programmability...and the fact...
appositive clause: that Nvidia is an extreme co-design company
modifier: of our architecture
```

**Grammar points:**
- **the fact that 同位语从句** - that 从句解释 fact 的具体内容
- **of 介词短语嵌套** - 多层 of 短语修饰,需从后往前理解

### [25:38]
**Original:** Down to CUDA C++, instead of using cuBLAS and NCCL, they've got their own stack which compiles to other accelerators as well.

**Translation:** 深入到CUDA C++层面,他们不使用cuBLAS和NCCL,而是有自己的技术栈,这个技术栈也能编译到其他加速器上。

**Core structure:**
- They've got their own stack which compiles to other accelerators.  
  他们有自己的技术栈,能编译到其他加速器。

**Structure tree:**
```
prepositional phrase: Down to CUDA C++
instead of phrase: instead of using cuBLAS and NCCL
main clause: they've got their own stack
relative clause: which compiles to other accelerators as well
```

**Grammar points:**
- **介词短语前置** - Down to CUDA C++作状语前置,强调技术深度
- **instead of 替代结构** - 表示选择自己的方案而非使用现成工具

### [27:00]
**Original:** You know that if something happens, it's more likely in your code and not in the mountain of code underneath.

**Translation:** 你知道如果出了问题,更可能是在你的代码里,而不是在底层那堆积如山的代码中。

**Core structure:**
- You know that it's more likely in your code.  
  你知道问题更可能在你的代码里。

**Structure tree:**
```
main clause: You know that...
object clause: it's more likely in your code
conditional clause: if something happens
comparison: more likely in X and not in Y
metaphor: mountain of code underneath
```

**Grammar points:**
- **宾语从句嵌套条件从句** - know后接that从句,从句内又包含if条件句
- **比较结构 more...and not** - 表示可能性对比,强调前者而非后者

### [27:34]
**Original:** The second thing is, if you're a developer building anything at all, the single most important thing you want is an install base.

**Translation:** 第二点是,如果你是一个开发者,无论在构建什么,你想要的最重要的东西就是安装基数。

**Core structure:**
- The thing is you want an install base.  
  关键是你想要安装基数。

**Structure tree:**
```
main clause: The second thing is
conditional clause: if you're a developer building anything
predicate clause: the single most important thing you want is an install base
relative clause (reduced): (that) you want
```

**Grammar points:**
- **表语从句的复杂嵌套** - is后的表语从句内又包含条件句和定语从句
- **最高级强调 the single most important** - single加强most的唯一性和重要性

### [29:01]
**Original:** The combination of the richness of the ecosystem, the expansiveness of the install base, and the versatility of where we are makes CUDA invaluable.

**Translation:** 生态系统的丰富性、安装基数的广泛性,以及我们所在位置的多样性,这三者的结合使得CUDA变得无价。

**Core structure:**
- The combination makes CUDA invaluable.  
  这种结合使CUDA无价。

**Structure tree:**
```
subject: The combination of A, B, and C
A: the richness of the ecosystem
B: the expansiveness of the install base
C: the versatility of where we are
predicate: makes CUDA invaluable
```

**Grammar points:**
- **三重并列的of短语** - 三个the...of结构并列作combination的定语
- **名词性从句 where we are** - where引导的从句作of的宾语,表示位置/存在范围

### [31:44]
**Original:** It's not unusual that by the time we're done optimizing their stack or optimizing a particular kernel, their model sped up by 3x, 2x, 50%.

**Translation:** 当我们完成对他们的技术栈或特定内核的优化时,他们的模型速度提升3倍、2倍或50%,这并不罕见。

**Core structure:**
- It's not unusual that their model sped up.  
  他们的模型加速并不罕见。

**Structure tree:**
```
main: It's not unusual that...
subject: It (形式主语)
that-clause: their model sped up
time clause: by the time we're done optimizing...
parallel structure: optimizing their stack or optimizing a particular kernel
```

**Grammar points:**
- **It 作形式主语** - 真正主语是 that 从句,It 占位使句子平衡
- **by the time 时间状语从句** - 表示'到...时候为止',从句用完成时
- **并列结构** - optimizing A or optimizing B 表示选择关系

### [32:01]
**Original:** That's a huge number, especially when you're talking about the install base of the fleet that they have, of all the Hoppers and Blackwells that they have.

**Translation:** 这是一个巨大的数字,特别是当你谈论他们拥有的整个设备群的安装基数时,即他们拥有的所有Hopper和Blackwell芯片。

**Core structure:**
- That's a huge number, especially when you're talking about the install base.  
  这是个巨大的数字,尤其是当你谈论安装基数时。

**Structure tree:**
```
main: That's a huge number
adverbial: especially when you're talking about...
of-phrase 1: of the fleet that they have
of-phrase 2: of all the Hoppers and Blackwells that they have
relative clauses: that they have (修饰 fleet 和 Hoppers/Blackwells)
```

**Grammar points:**
- **especially when 引导时间/条件状语** - 强调特定情况下的重要性
- **多层 of 短语嵌套** - 第二个 of 短语进一步解释第一个,造成理解难度
- **重复的定语从句** - 两个 that they have 分别修饰不同名词

### [33:54]
**Original:** We can bring them all of the great customers in the world. They're all built on Nvidia. And the reason why all these companies are built on Nvidia is because our reach and our versatility is so great.

**Translation:** 我们可以为他们带来世界上所有优秀的客户。这些客户都是基于英伟达构建的。所有这些公司都基于英伟达构建的原因是我们的覆盖范围和多功能性非常强大。

**Core structure:**
- The reason why companies are built on Nvidia is because our reach and versatility is great.  
  公司基于英伟达构建的原因是我们的覆盖范围和多功能性很强。

**Structure tree:**
```
main: The reason is because...
reason clause: why all these companies are built on Nvidia
because clause: our reach and versatility is so great
compound subject: our reach and our versatility
```

**Grammar points:**
- **The reason why... is because...** - 虽然语法上有争议(重复表原因),但口语中常用
- **被动语态** - are built on 表示'基于...构建'

### [34:36]
**Original:** That's the reason why, between the combination of: one, our perf per dollar is so great that they have the lowest cost tokens.

**Translation:** 这就是原因,在以下因素的组合之间:第一,我们的性价比如此之高,以至于他们拥有最低成本的token。

**Core structure:**
- That's the reason why our perf per dollar is so great that they have the lowest cost tokens.  
  这就是为什么我们的性价比如此之高,以至于他们拥有最低成本token的原因。

**Structure tree:**
```
main: That's the reason why...
prepositional phrase: between the combination of
enumeration: one, our perf per dollar...
so...that structure: so great that they have the lowest cost tokens
```

**Grammar points:**
- **between the combination of** - 口语化表达,引出多个并列因素
- **so...that 结果状语从句** - 表示'如此...以至于',强调因果关系
- **冒号后列举** - 用 one 开始列举,后续应有 two, three 等

### [34:53]
**Original:** So if one of these companies, if our partners, built a one gigawatt data center, that one gigawatt data center better deliver the maximum amount of revenues and number of tokens, which directly translates to revenues.

**Translation:** 所以如果这些公司中的一家,如果我们的合作伙伴,建造了一个1吉瓦的数据中心,那个1吉瓦的数据中心最好能提供最大量的收入和token数量,这直接转化为收入。

**Core structure:**
- If our partners built a data center, that data center better deliver maximum revenues.  
  如果我们的合作伙伴建造数据中心,那个数据中心最好能提供最大收入。

**Structure tree:**
```
conditional: if one of these companies... built a data center
appositive: if our partners (重述主语)
main: that data center better deliver...
object: the maximum amount of revenues and number of tokens
relative clause: which directly translates to revenues
```

**Grammar points:**
- **双重 if 从句** - 第二个 if 是同位语,重新表述第一个条件的主语
- **had better 的变体** - better 在此表示'最好',省略 had,口语化用法
- **非限制性定语从句** - which 指代前面整个概念,补充说明

### [37:01]
**Original:** Without Anthropic, why would there be any TPU growth at all?

**Translation:** 如果没有Anthropic,为什么会有任何TPU增长呢?

**Core structure:**
- Why would there be TPU growth?  
  为什么会有TPU增长?

**Structure tree:**
```
conditional phrase: Without Anthropic
main clause: why would there be growth
virtual mood: would + base verb
emphasis: at all
```

**Grammar points:**
- **虚拟语气(would)** - 表示假设情况下的推测,Without引导含蓄条件
- **there be结构的疑问形式** - why would there be表示对存在性的质疑

### [38:32]
**Original:** I guess their logic is, "Hey, it doesn't need to be better. It just needs to be not more than 70% worse," because they're paying you 70% margins.

**Translation:** 我猜他们的逻辑是,'嘿,它不需要更好。它只需要不比你的差超过70%',因为他们支付给你70%的利润率。

**Core structure:**
- Their logic is that it needs to be not more than 70% worse.  
  他们的逻辑是它不需要差超过70%。

**Structure tree:**
```
main clause: their logic is...
predicative clause: it doesn't need to be better
parallel structure: It just needs to be...
double negative: not more than...worse
causal clause: because they're paying...
```

**Grammar points:**
- **双重否定结构** - not more than...worse创造复杂的比较关系
- **并列表语从句** - 两个needs to be从句通过转折关系连接

### [40:00]
**Original:** I would say my mistake is I didn't deeply internalize that they really had no other options, that a VC would never put in $5-10 billion of investment into an AI lab with the hopes of it turning out to be Anthropic.

**Translation:** 我想说我的错误是,我没有深刻理解他们真的别无选择,风投永远不会向一个AI实验室投入50-100亿美元,希望它能成为Anthropic。

**Core structure:**
- My mistake is I didn't internalize that they had no options.  
  我的错误是我没有理解他们别无选择。

**Structure tree:**
```
main clause: my mistake is...
predicative clause 1: I didn't internalize that...
object clause 1: they had no options
appositive clause: that a VC would never...
purpose phrase: with the hopes of...
```

**Grammar points:**
- **多层嵌套从句** - 表语从句内含宾语从句,再接同位语从句解释
- **with the hopes of结构** - 表示伴随的目的或期望

### [40:54]
**Original:** If I could rewind everything—and Nvidia could have been as big back then as we are now—I would've been more than happy to do it.

**Translation:** 如果我能让一切倒回去——而且英伟达当时就能像现在这么大——我会非常乐意这么做。

**Core structure:**
- If I could rewind everything, I would've been happy to do it.  
  如果我能让一切倒回去,我会乐意这么做。

**Structure tree:**
```
conditional clause: If I could rewind...
parenthetical insertion: and Nvidia could have been...
main clause: I would've been happy
virtual mood: past unreal conditional
```

**Grammar points:**
- **虚拟语气(与过去事实相反)** - If...could/would've结构表示对过去的假设
- **插入语** - 破折号内的并列条件打断主句,增加理解难度

### [41:28]
**Original:** So if over these many years you were giving them the compute, you saw where it was headed, and they were worth like one tenth what they're worth now a couple years ago—or even a year ago in some cases and you had all this cash—there's a world where either Nvidia themselves becomes a foundation lab, does a huge investment to make that possible, or has made the deals you've made now at current valuations much earlier on.

**Translation:** 所以如果这么多年来你一直给他们提供算力,你看到了发展方向,而且几年前——或者在某些情况下甚至一年前——他们的估值只有现在的十分之一,而你又有这么多现金,那么就存在这样一种可能:要么英伟达自己成为基础模型实验室,进行巨额投资来实现这一点,要么在更早的时候就以当前估值达成你现在达成的交易。

**Core structure:**
- If you were giving them compute and had cash, there's a world where Nvidia becomes a lab or makes deals earlier.  
  如果你提供算力且有现金,就可能英伟达成为实验室或更早达成交易。

**Structure tree:**
```
long conditional: if you were giving...saw...were worth...had cash
parenthetical: or even a year ago in some cases
main clause: there's a world where...
alternative outcomes: either...or...
complex comparison: worth one tenth what they're worth now
```

**Grammar points:**
- **超长条件从句** - 多个并列条件+插入语+复杂比较结构
- **there's a world where结构** - 表示假设情境,相当于'有一种可能性是...'
- **either...or并列结构** - 连接三个复杂的选项,每个都有自己的修饰成分

### [44:31]
**Original:** If we didn't take the risk that we take—if we didn't build NVLink the way we built it, if we didn't build the whole stack, if we didn't create the ecosystem the way we did, if we didn't dedicate ourselves to 20 years of CUDA while losing money most of that time—if we didn't do it, nobody else would have done it.

**Translation:** 如果我们不承担我们所承担的风险——如果我们不按照我们的方式构建NVLink,如果我们不构建整个技术栈,如果我们不按照我们的方式创建生态系统,如果我们不在大部分时间亏损的情况下坚持20年投入CUDA——如果我们不做这些,没有其他人会做。

**Core structure:**
- If we didn't do it, nobody else would have done it.  
  如果我们不做,没有其他人会做。

**Structure tree:**
```
main clause: nobody else would have done it
condition: If we didn't do it
parenthetical expansion: multiple parallel if-clauses
  - if we didn't take the risk
  - if we didn't build NVLink
  - if we didn't build the stack
  - if we didn't create the ecosystem
  - if we didn't dedicate ourselves to CUDA
```

**Grammar points:**
- **虚拟语气(与过去事实相反)** - If + didn't do / would have done 表示对过去的假设
- **破折号插入多重条件从句** - 主条件句中插入多个并列的if从句,增加理解难度
- **while引导时间状语从句** - while losing money 表示伴随状态

### [44:52]
**Original:** A decade and a half ago, we pushed into domain-specific libraries because we realized that if we didn't create these domain-specific libraries, whether it's for ray tracing or image generation or even the early works of AI, these models, if we didn't create them, for data processing, structured data processing, or vector data processing, if we didn't create them, nobody would.

**Translation:** 十五年前,我们推进领域专用库,因为我们意识到如果我们不创建这些领域专用库——无论是用于光线追踪、图像生成还是AI的早期工作,这些模型——如果我们不创建它们,用于数据处理、结构化数据处理或向量数据处理,如果我们不创建它们,没有人会做。

**Core structure:**
- We pushed into libraries because we realized nobody would create them.  
  我们推进这些库,因为我们意识到没人会创建它们。

**Structure tree:**
```
main clause: we pushed into libraries
reason clause: because we realized that...
nested condition: if we didn't create them, nobody would
parenthetical insertions:
  - whether it's for ray tracing or...
  - these models
  - for data processing...
```

**Grammar points:**
- **多层嵌套从句** - because从句内嵌套that从句,再嵌套if条件句
- **多重插入语** - whether从句和列举打断主句流畅性
- **省略结构** - nobody would后省略create them

### [46:34]
**Original:** Because I want our ecosystem to thrive. I want the architecture and AI to be able to connect with as many industries as possible, as many countries as possible, and make it possible for the planet to be built on AI and to be built on the American tech stack.

**Translation:** 因为我希望我们的生态系统蓬勃发展。我希望这个架构和AI能够与尽可能多的行业、尽可能多的国家连接,并使整个地球有可能建立在AI之上,建立在美国技术栈之上。

**Core structure:**
- I want AI to connect with industries and countries and make it possible for the planet to be built on AI.  
  我希望AI与行业和国家连接,并使地球能够建立在AI之上。

**Structure tree:**
```
main clause: I want the architecture and AI to...
parallel infinitives:
  - to be able to connect with...
  - (to) make it possible for...
parallel prepositional phrases:
  - as many industries as possible
  - as many countries as possible
  - to be built on AI / on the American tech stack
```

**Grammar points:**
- **want + object + to do结构** - 复杂宾语补足语,包含多个并列不定式
- **as...as possible重复结构** - 强调最大化范围
- **make it possible for sb/sth to do** - it作形式宾语,for引出逻辑主语

### [47:42]
**Original:** If you would have taken those 60 graphics companies and asked yourself which one was going to make it, Nvidia would be at the top of that list not to make it.

**Translation:** 如果你当时拿那60家图形公司来问自己哪一家会成功,英伟达会排在那个不会成功的名单的首位。

**Core structure:**
- If you had asked which one would make it, Nvidia would be at the top of the list not to make it.  
  如果你问哪家会成功,英伟达会排在不会成功名单的首位。

**Structure tree:**
```
main clause: Nvidia would be at the top
condition: If you would have taken... and asked...
embedded question: which one was going to make it
modifier: not to make it (modifying 'list')
```

**Grammar points:**
- **混合虚拟语气** - would have taken(非标准)表过去假设,主句would be表结果
- **间接疑问句** - which one was going to make it作asked的宾语
- **make it习语** - 表示成功、做到

### [48:07]
**Original:** Either let them all take care of themselves, or take care of all of them.

**Translation:** 要么让他们都自己照顾自己,要么照顾他们所有人。

**Core structure:**
- Either let them take care of themselves, or take care of them.  
  要么让他们自己照顾自己,要么照顾他们。

**Structure tree:**
```
parallel imperatives:
  - let them all take care of themselves
  - (you) take care of all of them
connector: either...or...
```

**Grammar points:**
- **either...or...并列结构** - 连接两个祈使句,表示二选一
- **let + object + do** - 使役动词结构
- **all/all of them强调** - 重复使用all强调全部包含,无一例外

### [49:41]
**Original:** When someone like OpenAI needs an investment of a $30 billion scale because it's still before their IPO, and we deeply believe in them and I deeply believe that they're going to be an... Well, they're an extraordinary company already today.

**Translation:** 当像OpenAI这样的公司需要300亿美元规模的投资，因为他们还没有上市，而我们深信他们，我也深信他们将会成为……嗯，他们今天已经是一家非凡的公司了。

**Core structure:**
- When someone needs an investment and we believe in them, they're an extraordinary company.  
  当有人需要投资而我们相信他们时，他们是一家非凡的公司。

**Structure tree:**
```
time clause: When someone needs investment...
reason clause: because it's before their IPO
coordinate clause: and we believe in them
coordinate clause: and I believe that...
main clause: they're an extraordinary company
```

**Grammar points:**
- **多重并列从句** - when从句内嵌套because从句，后接两个and并列从句，结构层次复杂
- **口语化中断** - 句中出现自我修正(Well)和未完成的表达，增加理解难度

### [52:12]
**Original:** So the first thing is, we work really hard with everybody to get a forecast done, because these things take a long time to build, and the data centers take a long time to build.

**Translation:** 所以第一件事是，我们与每个人都非常努力地合作来完成预测，因为这些东西需要很长时间来制造，而数据中心也需要很长时间来建设。

**Core structure:**
- We work hard to get a forecast done because these things take a long time to build.  
  我们努力完成预测，因为这些东西需要很长时间来制造。

**Structure tree:**
```
main clause: the first thing is...
predicative clause: we work hard to get forecast done
reason clause: because these things take time
coordinate clause: and data centers take time
```

**Grammar points:**
- **表语从句** - is后接完整句子作表语
- **不定式作目的状语** - to get a forecast done表示工作的目的

### [52:52]
**Original:** But beyond that, if you're not ready because your data center's not ready, or certain components aren't ready to enable you to stand up a data center, we might decide to serve another customer first.

**Translation:** 但除此之外，如果你还没准备好，因为你的数据中心还没准备好，或者某些组件还没准备好让你建立数据中心，我们可能会决定先服务另一个客户。

**Core structure:**
- If you're not ready, we might decide to serve another customer first.  
  如果你还没准备好，我们可能会决定先服务另一个客户。

**Structure tree:**
```
condition clause: if you're not ready
reason clause: because your data center's not ready
alternative clause: or components aren't ready
purpose clause: to enable you to stand up...
main clause: we might decide to serve...
```

**Grammar points:**
- **多层嵌套条件句** - if从句内嵌套because从句和or并列从句，再嵌套不定式目的状语
- **enable sb to do结构** - 表示使某人能够做某事

### [55:37]
**Original:** One of the things you can count on with Nvidia is that this year, Vera Rubin is going to be incredible. Next year, Vera Rubin Ultra will come.

**Translation:** 关于英伟达，你可以指望的一件事是，今年Vera Rubin将会非常出色。明年，Vera Rubin Ultra将会推出。

**Core structure:**
- One thing you can count on is that Vera Rubin is going to be incredible.  
  你可以指望的一件事是Vera Rubin将会非常出色。

**Structure tree:**
```
main clause: One of the things is that...
attributive clause: you can count on
predicative clause: that Vera Rubin is going to be incredible
time modifier: this year
```

**Grammar points:**
- **定语从句省略关系代词** - you can count on后省略了that/which
- **表语从句** - that从句说明things的具体内容

### [55:57]
**Original:** You're going to have to go find another ASIC team in the world—pick your ASIC team—where you can say, 'I can bet the farm, I can bet my entire business that you will be here for me every single year.'

**Translation:** 你得去找世界上另一个ASIC团队——随便挑一个ASIC团队——在那里你可以说，'我可以押上全部家当，我可以押上我的整个生意，赌你每一年都会在这里支持我。'

**Core structure:**
- You have to find another team where you can say 'I can bet that you will be here for me.'  
  你得找另一个团队，在那里你可以说'我敢打赌你会支持我。'

**Structure tree:**
```
main clause: You're going to have to find team
parenthetical: pick your ASIC team
relative clause: where you can say...
direct speech: I can bet... that you will be here
object clause: that you will be here for me
```

**Grammar points:**
- **插入语** - 破折号间的内容打断主句，增加口语真实感
- **bet the farm习语** - 押上全部家当，表示完全信任
- **多层引语嵌套** - say后接直接引语，引语内又有that宾语从句

### [57:12]
**Original:** So I think this ability for Nvidia to be the foundation of the world's AI industry, this is a position that has taken us a couple of decades to arrive at.

**Translation:** 所以我认为英伟达成为全球AI行业基础的这种能力,这是一个我们花了几十年才达到的位置。

**Core structure:**
- This ability is a position that has taken us decades to arrive at.  
  这种能力是一个我们花了几十年才达到的位置。

**Structure tree:**
```
main: I think [that]...
object clause: this ability... is a position
appositive: this is a position
relative clause: that has taken us decades to arrive at
```

**Grammar points:**
- **同位语结构** - this ability 和 this is a position 是同位关系,用逗号分隔重述同一概念
- **定语从句 + 不定式** - that 从句修饰 position, arrive at 后省略了先行词 which/that

### [58:05]
**Original:** It has such cyber-offensive capabilities that we don't think the world is ready until we make sure these zero-days are patched up.

**Translation:** 它具有如此强大的网络攻击能力,以至于我们认为在确保这些零日漏洞被修补之前,世界还没有准备好。

**Core structure:**
- It has such capabilities that we don't think the world is ready.  
  它具有如此强大的能力,以至于我们认为世界还没有准备好。

**Structure tree:**
```
main: It has such capabilities that...
result clause: that we don't think...
object clause: the world is ready
time clause: until we make sure...
object clause: these zero-days are patched up
```

**Grammar points:**
- **such...that 结果状语从句** - 表示'如此...以至于',强调程度和结果
- **否定转移** - we don't think 实际否定的是从句内容(世界没准备好),不是'我们不认为'
- **until 时间状语从句** - 修饰 ready,表示'直到...才'

### [59:39]
**Original:** So the question is, considering all the assets they already have—they have an abundance of energy, they have plenty of chips, they've got most of the AI researchers—if you're worried about them, what is the best way to create a safe world?

**Translation:** 所以问题是,考虑到他们已经拥有的所有资产——他们有充足的能源,有大量的芯片,拥有大多数AI研究人员——如果你担心他们,创造一个安全世界的最佳方式是什么?

**Core structure:**
- The question is, what is the best way to create a safe world?  
  问题是,创造一个安全世界的最佳方式是什么?

**Structure tree:**
```
main: the question is...
predicative clause: what is the best way
participle phrase: considering all the assets
parenthetical insertion: —they have... they have... they've got—
conditional clause: if you're worried about them
```

**Grammar points:**
- **插入语(破折号)** - 破折号内的三个并列句补充说明 assets,打断主句结构
- **现在分词短语作状语** - considering 引导条件/原因状语,修饰整个主句

### [01:01:08]
**Original:** One of the things that is underemphasized is the richness of the ecosystem around cybersecurity, AI cybersecurity and AI security and AI privacy and AI safety.

**Translation:** 被低估的事情之一是围绕网络安全、AI网络安全、AI安全、AI隐私和AI安全的生态系统的丰富性。

**Core structure:**
- One of the things is the richness of the ecosystem.  
  其中一件事是生态系统的丰富性。

**Structure tree:**
```
main: One of the things is the richness
relative clause: that is underemphasized
prepositional phrase: around cybersecurity...
parallel structure: AI cybersecurity and AI security and AI privacy and AI safety
```

**Grammar points:**
- **被动语态定语从句** - that is underemphasized 修饰 things,表示'被低估的'
- **多重并列结构** - 四个 AI 相关领域用 and 连接,形成复杂名词短语

### [01:03:00]
**Original:** But what we also want is to make sure that all the AI developers in the world are developing on the American tech stack, and making the contributions, the advancements of AI—especially when it's open source—available to the American ecosystem.

**Translation:** 但我们还想要的是确保全世界所有的AI开发者都在美国技术栈上开发,并使AI的贡献和进步——尤其是开源的——能够为美国生态系统所用。

**Core structure:**
- What we want is to make sure that developers are developing and making contributions available.  
  我们想要的是确保开发者在开发并使贡献可用。

**Structure tree:**
```
main: what we want is to make sure that...
subject clause: what we also want
object clause: all developers are developing... and making...
parallel gerunds: developing / making
parenthetical: —especially when it's open source—
```

**Grammar points:**
- **What 主语从句 + 表语** - What we want 作主语,is to make sure 作表语
- **make + 宾语 + 宾补(形容词)** - making the contributions available 表示'使贡献可用'
- **插入的时间状语从句** - 破折号内的 when 从句补充说明条件,打断句子流畅性

### [01:04:22]
**Original:** Furthermore, even if they train a model like this, the ability to deploy it at scale… If you had a cyber hacker, it's much more dangerous if they have a million of them versus a thousand of them.

**Translation:** 此外，即使他们训练出这样的模型，大规模部署的能力……如果你有一个网络黑客，如果他们有一百万个而不是一千个，那就危险得多。

**Core structure:**
- The ability to deploy it at scale matters. It's more dangerous if they have a million versus a thousand.  
  大规模部署的能力很重要。如果他们有一百万个而不是一千个，就更危险。

**Structure tree:**
```
main: the ability to deploy...
condition 1: even if they train a model
condition 2: If you had a cyber hacker
comparison: if they have a million vs a thousand
```

**Grammar points:**
- **even if 让步状语从句** - 表示即使在某种假设条件下
- **虚拟语气 (If you had...)** - 与现在事实相反的假设
- **versus 比较结构** - 对比两种数量级的差异

### [01:04:54]
**Original:** So then the question is, isn't it better that we get American companies, because they have more compute, to get to the Mythos-level capabilities first, prepare our society for it, before China can get to it because they have less compute?

**Translation:** 那么问题是，让美国公司（因为他们有更多算力）首先达到 Mythos 级别的能力，为我们的社会做好准备，在中国（因为他们算力较少）达到之前，这样不是更好吗？

**Core structure:**
- Isn't it better that American companies get to the capabilities first before China?  
  让美国公司先达到这些能力不是更好吗？

**Structure tree:**
```
main: isn't it better that...
subject clause: that we get American companies to...
reason 1: because they have more compute
purpose: prepare our society
time clause: before China can get to it
reason 2: because they have less compute
```

**Grammar points:**
- **反意疑问句 (isn't it better)** - 用否定疑问形式表达肯定建议
- **get sb to do 使役结构** - 让某人做某事
- **多重原因状语从句** - 两个 because 从句分别修饰不同部分

### [01:07:14]
**Original:** For example, the United States is scarce on energy, which is the reason why Nvidia has to keep advancing our architecture and do this extreme co-design so that with the few chips that we ship—with the few chips, Because the amount of energy is so limited, our throughput per watt is off the charts.

**Translation:** 例如，美国能源稀缺，这就是为什么英伟达必须不断推进我们的架构并进行这种极端的协同设计，以便用我们出货的少量芯片——用少量芯片，因为能源数量如此有限，我们每瓦特的吞吐量高得离谱。

**Core structure:**
- The US is scarce on energy, which is why Nvidia has to advance architecture so that our throughput per watt is off the charts.  
  美国能源稀缺，这就是为什么英伟达必须推进架构，使我们每瓦特的吞吐量极高。

**Structure tree:**
```
main: the US is scarce on energy
result clause: which is the reason why Nvidia has to...
purpose clause: so that... throughput is off the charts
reason clause: Because the amount of energy is limited
interruption: with the few chips (repeated)
```

**Grammar points:**
- **which 引导非限制性定语从句** - 指代前面整个句子的内容
- **so that 目的状语从句** - 表示采取行动的目的
- **句中自我打断和重复** - 口语特征，说话者中途重新组织思路

### [01:08:42]
**Original:** But as you know, the bottleneck often in training and doing inference on these models is the amount of bandwidth.

**Translation:** 但如你所知，在训练和对这些模型进行推理时，瓶颈往往是带宽的数量。

**Core structure:**
- The bottleneck is the amount of bandwidth.  
  瓶颈是带宽的数量。

**Structure tree:**
```
parenthetical: as you know
main: the bottleneck is the amount of bandwidth
adverbial: often
context: in training and doing inference on these models
```

**Grammar points:**
- **as you know 插入语** - 打断主句，增加口语自然感
- **倒装结构** - often 位于主语和谓语之间，强调频率

### [01:09:04]
**Original:** But that doesn't change the fact that you need EUV for the most advanced HBM.

**Translation:** 但这并不能改变一个事实，即你需要 EUV（极紫外光刻）来制造最先进的 HBM（高带宽内存）。

**Core structure:**
- That doesn't change the fact that you need EUV.  
  这不能改变你需要 EUV 的事实。

**Structure tree:**
```
main: that doesn't change the fact
appositive clause: that you need EUV
purpose: for the most advanced HBM
```

**Grammar points:**
- **同位语从句** - that 从句解释说明 the fact 的具体内容
- **最高级 (the most advanced)** - 表示技术的最高水平

### [01:09:58]
**Original:** There is no question, MoE is a great invention. There's no question, all the incredible attention mechanisms reduce the amount of compute. We have got to acknowledge that most of the advances in AI came out of algorithm advances, not just the raw hardware.

**Translation:** 毫无疑问，MoE是一项伟大的发明。毫无疑问，所有令人难以置信的注意力机制都减少了计算量。我们必须承认，AI的大部分进步来自算法进步，而不仅仅是原始硬件。

**Core structure:**
- We have got to acknowledge that most advances came from algorithm advances.  
  我们必须承认大部分进步来自算法进步。

**Structure tree:**
```
main clause: We have got to acknowledge that...
object clause: that most advances came from algorithm advances
contrast: not just the raw hardware
parallel structure: There is no question... (repeated twice)
```

**Grammar points:**
- **have got to** - 表示强烈义务，相当于must
- **that引导宾语从句**
- **not just... 否定对比** - 强调重点在algorithm而非hardware

### [01:10:19]
**Original:** Now, if most advances came from algorithms and computer science and programming, tell me that their army of AI researchers is not their fundamental advantage.

**Translation:** 现在，如果大部分进步来自算法、计算机科学和编程，那么告诉我，他们的AI研究人员大军难道不是他们的根本优势吗？

**Core structure:**
- Tell me that their army of researchers is not their advantage.  
  告诉我他们的研究人员大军不是他们的优势。

**Structure tree:**
```
conditional clause: if most advances came from...
main clause (imperative): tell me that...
object clause: that their army is not their advantage
rhetorical question structure (negative expectation)
```

**Grammar points:**
- **条件状语从句** - if引导，为主句提供逻辑前提
- **祈使句+宾语从句** - tell me that...结构表达反问语气
- **反问修辞** - 否定形式表达肯定意思，期待对方无法反驳

### [01:11:15]
**Original:** You set it up as a premise that it was bad news. I'm going to give you the bad news, that AI models around the world are developed and they run best on non-American hardware.

**Translation:** 你把它设定为一个前提，认为这是坏消息。我要告诉你真正的坏消息，那就是世界各地开发的AI模型在非美国硬件上运行得最好。

**Core structure:**
- I'm going to give you the bad news that AI models run best on non-American hardware.  
  我要告诉你坏消息，AI模型在非美国硬件上运行最好。

**Structure tree:**
```
main clause: I'm going to give you the bad news
appositive clause: that AI models are developed and run best...
compound predicate: are developed and run best
modifier: around the world
```

**Grammar points:**
- **be going to 表意图** - 表达说话者即将陈述的内容
- **同位语从句** - that从句解释说明the bad news的具体内容
- **并列谓语** - are developed and run连接两个被动/主动动作

### [01:12:07]
**Original:** Coming out of the box, if all of the AI models run best on somebody else's tech stack, you've got to be arguing some ridiculous claim right now that that's a good thing for the United States.

**Translation:** 开箱即用的情况下，如果所有AI模型在别人的技术栈上运行得最好，那你现在一定是在提出某种荒谬的主张，说这对美国是件好事。

**Core structure:**
- You've got to be arguing some ridiculous claim that that's a good thing.  
  你一定是在提出荒谬主张说这是好事。

**Structure tree:**
```
adverbial phrase: Coming out of the box
conditional clause: if all models run best on...
main clause: you've got to be arguing some claim
appositive clause: that that's a good thing
double 'that' structure
```

**Grammar points:**
- **must be doing 推测** - you've got to be arguing表示对当前行为的强烈推测
- **双that结构** - 第一个that引导同位语从句，第二个that指代前文情况
- **条件从句前置** - if从句插在分词短语和主句之间，增加理解难度

### [01:13:44]
**Original:** Listen, why are you causing one layer of the AI industry to lose an entire market so that you could benefit another layer of the AI industry?

**Translation:** 听着，你为什么要让AI行业的一个层级失去整个市场，以便让AI行业的另一个层级受益？

**Core structure:**
- Why are you causing one layer to lose a market so that you could benefit another layer?  
  你为什么让一个层级失去市场以便让另一个层级受益？

**Structure tree:**
```
main clause: why are you causing one layer to lose...
cause structure: cause sb/sth to do
purpose clause: so that you could benefit...
parallel structure: one layer... another layer
```

**Grammar points:**
- **cause sb to do** - 使役结构，表示导致某人/某物做某事
- **so that目的状语从句** - 表达行为的目的，could表示可能性
- **对比平行结构** - one layer... another layer形成对比

### [01:14:31]
**Original:** If you think the base model was here and the backdoor model was here, you can kind of linearly interpolate the weights to adjust the strength of the backdoor, but you can also extrapolate it to make the backdoor even stronger.

**Translation:** 如果你认为基础模型在这里，后门模型在这里，你可以对权重进行线性插值来调整后门的强度，但你也可以外推它来使后门更强。

**Core structure:**
- You can interpolate the weights, but you can also extrapolate it.  
  你可以插值权重，但也可以外推它。

**Structure tree:**
```
conditional clause: If you think...
parallel main clauses: you can interpolate... but you can also extrapolate...
purpose clause 1: to adjust the strength
purpose clause 2: to make the backdoor stronger
```

**Grammar points:**
- **条件从句 + 并列主句** - If 从句设定前提，but 连接两个对比的可能性
- **不定式表目的** - to adjust 和 to make 说明动作的目的

### [01:15:02]
**Original:** Being able to verify that a model only does what you think it does is one of the most important open questions in AI security.

**Translation:** 能够验证一个模型只做你认为它会做的事情，是人工智能安全领域最重要的开放性问题之一。

**Core structure:**
- Being able to verify is one of the most important questions.  
  能够验证是最重要的问题之一。

**Structure tree:**
```
subject: Being able to verify that...
  gerund phrase as subject
  object clause: that a model only does...
    embedded clause: what you think it does
predicate: is one of the most important questions
```

**Grammar points:**
- **动名词短语作主语** - Being able to... 整个短语作句子主语
- **嵌套从句** - that 从句中包含 what 从句，层层嵌套

### [01:15:21]
**Original:** And remember, they're still stuck on 7nm while you'll move on to 3nm and then 2nm or 1.6nm with Feynman.

**Translation:** 记住，他们仍然停留在7纳米，而你将转向3纳米，然后是2纳米或1.6纳米的费曼工艺。

**Core structure:**
- They're stuck on 7nm while you'll move on to 3nm.  
  他们停留在7纳米，而你将转向3纳米。

**Structure tree:**
```
main clause: they're stuck on 7nm
contrast clause: while you'll move on to 3nm
sequence: and then 2nm or 1.6nm
modifier: with Feynman
```

**Grammar points:**
- **while 表对比** - 连接两个对比的情况，强调差异
- **be stuck on** - 固定搭配，表示停滞不前

### [01:17:08]
**Original:** Why is it that your policy, your philosophy, leads to the United States giving up a vast part of the world's market?

**Translation:** 为什么你的政策、你的理念会导致美国放弃世界市场的很大一部分？

**Core structure:**
- Why does your policy lead to the US giving up the market?  
  为什么你的政策导致美国放弃市场？

**Structure tree:**
```
question: Why is it that...
  强调句型
subject: your policy, your philosophy
predicate: leads to
object: the United States giving up...
  gerund as object of preposition
```

**Grammar points:**
- **Why is it that 强调句** - 用于强调疑问，比简单 Why 更正式
- **lead to + 动名词** - 导致某事发生，to 是介词后接动名词

### [01:17:27]
**Original:** I guess the claim here is, Dario had this quote where he said that it's like Boeing bragging that we're selling North Korea nukes, but the missile casings are made by Boeing.

**Translation:** 我想这里的说法是，达里奥有一句话，他说这就像波音公司吹嘘说我们在向朝鲜出售核武器，但导弹外壳是波音制造的。

**Core structure:**
- The claim is that it's like Boeing bragging.  
  这个说法是，这就像波音在吹嘘。

**Structure tree:**
```
main: the claim is
predicative: Dario had this quote
relative clause: where he said that...
object clause: it's like Boeing bragging that...
  nested clause: we're selling North Korea nukes
contrast: but the missile casings are made by Boeing
```

**Grammar points:**
- **多层嵌套从句** - quote where... that... that... 三层从句嵌套
- **类比结构** - it's like... 用具体例子类比抽象概念

### [01:18:40]
**Original:** However, we also have to recognize that AI is not just a model. AI is a five-layer cake. The AI industry matters across every single layer, and we want the United States to win at every single layer, including the chip layer.

**Translation:** 然而,我们也必须认识到,人工智能不仅仅是一个模型。人工智能是一个五层蛋糕。人工智能产业在每一层都很重要,我们希望美国在每一层都能获胜,包括芯片层。

**Core structure:**
- We have to recognize that AI is not just a model, and we want the US to win at every layer.  
  我们必须认识到人工智能不仅仅是一个模型,我们希望美国在每一层都能获胜。

**Structure tree:**
```
main clause 1: we have to recognize that...
object clause: AI is not just a model
main clause 2: we want the US to win...
modifier: including the chip layer
```

**Grammar points:**
- **宾语从句** - recognize 后接 that 引导的宾语从句
- **并列句** - and 连接两个独立主句

### [01:19:40]
**Original:** The single most important thing to our company is the richness of our ecosystem, which is about developers. 50% of the AI developers are in China. The United States should not give that up.

**Translation:** 对我们公司来说最重要的事情是我们生态系统的丰富性,这关乎开发者。50%的人工智能开发者在中国。美国不应该放弃这一点。

**Core structure:**
- The most important thing is the richness of our ecosystem.  
  最重要的事情是我们生态系统的丰富性。

**Structure tree:**
```
main clause: The most important thing is the richness
modifier: to our company
non-restrictive clause: which is about developers
```

**Grammar points:**
- **非限制性定语从句** - which 补充说明 ecosystem
- **最高级修饰** - the single most important 强调程度

### [01:20:33]
**Original:** The fact that I can buy this car brand one day and use another car brand another day, easy. Computing is not like that.

**Translation:** 我可以今天买这个汽车品牌,明天用另一个汽车品牌,这很容易。但计算不是这样的。

**Core structure:**
- The fact is easy. Computing is not like that.  
  这个事实很容易。计算不是这样的。

**Structure tree:**
```
subject: The fact that...
appositive clause: that I can buy... and use...
predicate: easy
contrast: Computing is not like that
```

**Grammar points:**
- **同位语从句** - that 从句解释 fact 的具体内容
- **省略句** - easy 前省略了 is

### [01:21:56]
**Original:** The idea is not that there is some key threshold of compute. It's that any marginal compute is helpful. So if you have more compute, you can train a better model. And I just want you to acknowledge that any marginal sales for the American technology industry is beneficial.

**Translation:** 这个想法不是说存在某个关键的计算阈值。而是任何边际计算都是有帮助的。所以如果你有更多的计算能力,你就能训练出更好的模型。我只是想让你承认,对美国技术产业来说,任何边际销售都是有益的。

**Core structure:**
- The idea is that any marginal compute is helpful, and I want you to acknowledge that any marginal sales is beneficial.  
  这个想法是任何边际计算都有帮助,我想让你承认任何边际销售都有益。

**Structure tree:**
```
main clause 1: The idea is that...
predicative clause: any marginal compute is helpful
conditional: if you have more compute...
main clause 2: I want you to acknowledge that...
object clause: any marginal sales is beneficial
```

**Grammar points:**
- **表语从句** - that 从句作 is 的表语
- **条件状语从句** - if 引导条件,表示假设情况
- **宾语从句嵌套** - acknowledge 后接 that 从句作宾语

### [01:24:01]
**Original:** To concede that market for the United States technology industry is a disservice to our country. It is a disservice to our national security. It is a disservice to our technology leadership, all for the benefit of one company.

**Translation:** 对美国技术产业来说,放弃那个市场是对我们国家的伤害。这是对我们国家安全的伤害。这是对我们技术领导地位的伤害,而这一切都是为了一家公司的利益。

**Core structure:**
- To concede that market is a disservice to our country, security, and leadership.  
  放弃那个市场是对我们国家、安全和领导地位的伤害。

**Structure tree:**
```
subject: To concede that market
modifier: for the US technology industry
predicate: is a disservice to...
parallel structure: country / security / leadership
modifier: all for the benefit of one company
```

**Grammar points:**
- **不定式作主语** - To concede 整个短语作句子主语
- **排比结构** - 三个 disservice to 强调负面影响

### [01:25:40]
**Original:** It's a little narrow-minded, and it led to unintended consequences that I'm describing to you right now that you seem to have a very hard time understanding.

**Translation:** 这有点目光短浅,而且它导致了我现在正在向你描述的意想不到的后果,而你似乎很难理解这些后果。

**Core structure:**
- It's narrow-minded, and it led to consequences that you have a hard time understanding.  
  这很目光短浅,它导致了你难以理解的后果。

**Structure tree:**
```
compound sentence: It's narrow-minded, and it led to...
first clause: It's a little narrow-minded
second clause: it led to unintended consequences
first relative clause: that I'm describing to you right now
second relative clause: that you seem to have a very hard time understanding
```

**Grammar points:**
- **双重定语从句** - 两个 that 从句连续修饰 consequences,第二个 that 的先行词仍是 consequences 而非 right now
- **have a hard time doing** - 做某事有困难的固定搭配

### [01:26:09]
**Original:** It is a good thing that American companies got to Mythos-level capabilities first, and then now they're going to hold off on those capabilities so that the American companies and American government can make their software more protected before that level of capability was announced.

**Translation:** 美国公司首先达到 Mythos 级别的能力是一件好事,然后现在他们将推迟使用这些能力,以便美国公司和美国政府能够在宣布该级别能力之前使其软件得到更好的保护。

**Core structure:**
- It is a good thing that companies got to capabilities first and will hold off so that they can make software more protected.  
  公司首先获得能力并将推迟使用以便能够更好地保护软件,这是好事。

**Structure tree:**
```
main clause: It is a good thing that...
subject clause: that American companies got to... first
coordinate clause: and then now they're going to hold off...
purpose clause: so that... can make their software more protected
time clause: before that level... was announced
```

**Grammar points:**
- **It is... that 主语从句** - it 作形式主语,that 从句是真正主语
- **so that 目的状语从句** - 表示推迟使用的目的
- **before 时间状语从句** - 修饰整个 so that 从句,说明保护软件的时间节点

### [01:26:45]
**Original:** I'll also tell you the potential cost is we allow one of the most important layers of the AI stack, the chip layer, to concede an entire market—the second largest market in the world—so that they could develop scale, so that they could develop their own ecosystem, so that future AI models are optimized in a very different way than the American tech stack.

**Translation:** 我还要告诉你潜在的代价是,我们允许 AI 技术栈中最重要的层之一——芯片层——让出整个市场(世界第二大市场),以便他们能够发展规模,以便他们能够发展自己的生态系统,以便未来的 AI 模型以与美国技术栈非常不同的方式进行优化。

**Core structure:**
- The cost is we allow the chip layer to concede a market so that they could develop scale and ecosystem.  
  代价是我们允许芯片层让出市场,以便他们能够发展规模和生态系统。

**Structure tree:**
```
main clause: the potential cost is...
predicative clause: we allow... to concede...
appositive: the chip layer (explains 'one of the most important layers')
appositive: the second largest market (explains 'an entire market')
three parallel purpose clauses: so that they could develop...
```

**Grammar points:**
- **allow sb/sth to do** - 允许某人/某物做某事
- **同位语** - 两个破折号中的内容分别解释前面的名词
- **三个并列 so that 从句** - 表示让出市场的三个目的,结构平行

### [01:28:45]
**Original:** If we scare everybody out of doing software engineering jobs because it's going to kill every software engineering job—and we don't have any software engineers as a result of that—we're doing a disservice to the United States.

**Translation:** 如果我们因为 AI 将会消灭所有软件工程工作而把所有人吓得不敢从事软件工程工作——结果我们就没有任何软件工程师了——那我们就是在损害美国的利益。

**Core structure:**
- If we scare everybody out of doing jobs, we're doing a disservice.  
  如果我们把所有人吓得不敢做这份工作,我们就是在造成损害。

**Structure tree:**
```
conditional sentence: If we scare..., we're doing...
condition clause: If we scare everybody out of doing... jobs
reason clause: because it's going to kill every... job
parenthetical clause: and we don't have any... as a result
main clause: we're doing a disservice to the United States
```

**Grammar points:**
- **scare sb out of doing sth** - 把某人吓得不敢做某事
- **插入语** - 破折号中的内容是对前面条件的结果补充说明
- **do a disservice to** - 对...造成损害/帮倒忙

### [01:29:14]
**Original:** If we misunderstand that so profoundly and we scare everybody out of going to radiology school, we're not going to have enough radiologists and good enough healthcare.

**Translation:** 如果我们如此深刻地误解了这一点,并且把所有人吓得不敢去放射学院学习,我们就不会有足够的放射科医生和足够好的医疗保健。

**Core structure:**
- If we misunderstand and scare everybody out of going to school, we're not going to have enough radiologists and healthcare.  
  如果我们误解并把所有人吓得不敢去上学,我们就不会有足够的放射科医生和医疗保健。

**Structure tree:**
```
conditional sentence: If we misunderstand... and scare..., we're not going to have...
two parallel condition clauses: we misunderstand / we scare everybody out of...
main clause: we're not going to have enough radiologists and good enough healthcare
```

**Grammar points:**
- **并列条件从句** - 两个 we 开头的从句用 and 连接,共同构成条件
- **enough 的位置** - enough 修饰名词时放在名词前(enough radiologists),修饰形容词时放在形容词后(good enough)

### [01:30:07]
**Original:** But in a few years' time, I'm making you the prediction that when we want the American tech stack, when we want American technology to be diffused around the world—out to India, out to the Middle East, out to Africa, out to Southeast Asia—when our country would like to export, because we would like to export our technology, we would like to export our standards, on that day, I want you and I to have that same conversation again.

**Translation:** 但在几年后,我向你预测,当我们想要美国技术栈,当我们想要美国技术扩散到世界各地——印度、中东、非洲、东南亚——当我们国家想要出口,因为我们想要出口我们的技术,我们想要出口我们的标准,在那一天,我希望你和我再进行同样的对话。

**Core structure:**
- I'm making you the prediction that on that day, I want you and I to have that conversation again.  
  我向你预测,在那一天,我希望你和我再进行那次对话。

**Structure tree:**
```
main clause: I'm making you the prediction that...
that-clause: when...when...when..., on that day, I want...
multiple when-clauses (temporal conditions)
parenthetical: —out to India...Southeast Asia—
because-clause (reason)
```

**Grammar points:**
- **多重时间状语从句嵌套** - 三个when从句层层递进,最后引出主句,造成理解延迟。
- **破折号插入语** - 列举地区打断句子主干,需暂存前文信息。
- **复杂宾语从句** - prediction后接that从句,从句内又包含多层结构。

### [01:31:34]
**Original:** The argument hinges on this: They've built models that are specified for the best chips that they make in a few years, those chips get exported around the world, that sets the standard.

**Translation:** 论点的关键在于:他们构建的模型是为几年后制造的最好芯片而设计的,这些芯片出口到世界各地,这就设定了标准。

**Core structure:**
- The argument hinges on this: They've built models, chips get exported, that sets the standard.  
  论点关键在于:他们构建了模型,芯片被出口,这设定了标准。

**Structure tree:**
```
main clause: The argument hinges on this
colon introduces explanation
three parallel clauses connected by commas
relative clause: that are specified for...
relative clause: that they make...
```

**Grammar points:**
- **冒号引出解释性并列句** - 冒号后三个独立分句用逗号连接,非典型并列结构。
- **嵌套定语从句** - models后接that从句,从句内chips又有that从句修饰。

### [01:33:12]
**Original:** It's an ecosystem, a computing architecture that allows for so much flexibility that if you wanted to change an architecture completely—create something like MoE, create something like diffusion, create something that's disaggregated—you could do so.

**Translation:** 这是一个生态系统,一个计算架构,它提供了如此大的灵活性,以至于如果你想完全改变架构——创建像MoE这样的东西,创建像扩散这样的东西,创建分解式的东西——你都可以做到。

**Core structure:**
- It's an architecture that allows flexibility that you could change architecture.  
  这是一个允许灵活性的架构,你可以改变架构。

**Structure tree:**
```
main clause: It's an ecosystem, a computing architecture
relative clause: that allows for flexibility
result clause: that if you wanted..., you could do so
conditional clause: if you wanted to change...
parenthetical: —create...create...create—
```

**Grammar points:**
- **so...that结果状语从句** - 表示'如此...以至于',that从句内又嵌套if条件句。
- **破折号插入并列结构** - 三个create短语打断if从句,需跳过插入语找到主句。

### [01:36:03]
**Original:** You're doing so much engineering and packaging and stacking, and the numerics and the system architecture, when you run out of capacity, to easily go back to another node… That's a level of R&D that no one could afford.

**Translation:** 你在做大量的工程、封装和堆叠,以及数值计算和系统架构,当你用完产能时,要轻松回到另一个节点...那是一个没人能负担得起的研发水平。

**Core structure:**
- When you run out of capacity, to go back to another node is a level of R&D that no one could afford.  
  当你用完产能时,回到另一个节点是没人能负担的研发水平。

**Structure tree:**
```
fragment: You're doing so much...
temporal clause: when you run out of capacity
infinitive phrase: to easily go back...
ellipsis: ...
main clause: That's a level of R&D
relative clause: that no one could afford
```

**Grammar points:**
- **句子碎片与省略** - 前半句无完整谓语,用省略号连接后句,口语化结构。
- **不定式短语作主语** - to go back作真正主语,由That指代。

### [01:36:42]
**Original:** One question somebody I was talking to had is, why doesn't Nvidia run multiple different chip projects at the same time with totally different architecture, so you could do something like a Cerebras-style wafer scale, you could do a Dojo-style huge package, you could do one without CUDA, you have the resources and the engineering talent to do all of these in parallel, so why put all the eggs in one basket, given who knows where AI might go and architectures might go?

**Translation:** 我交谈过的某人提出的一个问题是,为什么英伟达不同时运行多个完全不同架构的芯片项目,这样你可以做像Cerebras风格的晶圆级,可以做Dojo风格的大封装,可以做一个没有CUDA的,你有资源和工程人才并行做所有这些,那么为什么把所有鸡蛋放在一个篮子里,考虑到谁知道AI和架构会走向何方?

**Core structure:**
- The question is, why doesn't Nvidia run multiple projects, so why put all eggs in one basket?  
  问题是,为什么英伟达不运行多个项目,那为什么把所有鸡蛋放一个篮子?

**Structure tree:**
```
main clause: One question...is
embedded clause: somebody I was talking to had
predicative clause: why doesn't Nvidia run...
so-clause: so you could do...(multiple parallel clauses)
so-clause: so why put...
given-clause: given who knows where...
```

**Grammar points:**
- **多层嵌套从句** - 主语中嵌套定语从句,表语从句后接多个并列so从句。
- **given引导让步状语** - given表'考虑到',后接间接疑问句who knows where。
- **省略连词的并列句** - 多个you could do用逗号连接,无and/or,口语化表达。

### [01:37:32]
**Original:** If the workload were to change dramatically—and I don't mean the algorithms, I actually mean the workload, and that depends on the shape of the market—we may decide to add other accelerators.

**Translation:** 如果工作负载发生巨大变化——我指的不是算法，我实际上指的是工作负载，而这取决于市场的形态——我们可能会决定添加其他加速器。

**Core structure:**
- If the workload were to change dramatically, we may decide to add other accelerators.  
  如果工作负载发生巨大变化，我们可能会决定添加其他加速器。

**Structure tree:**
```
conditional clause: If the workload were to change dramatically
parenthetical insertion: —and I don't mean... shape of the market—
main clause: we may decide to add other accelerators
```

**Grammar points:**
- **虚拟语气 (were to)** - 表示对未来不太可能发生情况的假设
- **破折号插入语** - 中断主句流程插入澄清信息，增加理解难度

### [01:38:17]
**Original:** Because the customers make so much money—for example, our software engineers—if I can give them much more responsive tokens so that they're even more productive than they are today, I would pay for it.

**Translation:** 因为客户赚了很多钱——例如，我们的软件工程师——如果我能给他们响应更快的tokens，让他们比现在更高效，我愿意为此付费。

**Core structure:**
- If I can give them more responsive tokens, I would pay for it.  
  如果我能给他们响应更快的tokens，我愿意为此付费。

**Structure tree:**
```
reason clause: Because the customers make so much money
parenthetical: —for example, our software engineers—
conditional: if I can give them more responsive tokens
purpose clause: so that they're even more productive
main clause: I would pay for it
```

**Grammar points:**
- **多层嵌套从句** - 原因从句+条件从句+目的从句，层层嵌套
- **比较级强化 (even more...than)** - even 强调比较程度

### [01:38:46]
**Original:** That's the reason why we decided to expand the Pareto frontier and create a segment of inference that is faster response time, even though it's lower throughput.

**Translation:** 这就是我们决定扩展帕累托前沿并创建一个响应时间更快的推理细分市场的原因，尽管它的吞吐量较低。

**Core structure:**
- That's the reason why we decided to expand the frontier and create a segment.  
  这就是我们决定扩展前沿并创建一个细分市场的原因。

**Structure tree:**
```
main clause: That's the reason why we decided to...
reason clause: why we decided to expand... and create...
relative clause: that is faster response time
concessive clause: even though it's lower throughput
```

**Grammar points:**
- **reason why 定语从句** - why 引导定语从句修饰 reason
- **让步状语从句 (even though)** - 表示尽管存在相反情况但不影响主要决定

### [01:40:42]
**Original:** The reason for that is fairly fundamental, which is that the ability for general purpose computing to continue to scale has largely run its course.

**Translation:** 其原因相当根本，那就是通用计算继续扩展的能力已经基本走到尽头了。

**Core structure:**
- The reason is that the ability has run its course.  
  原因是这种能力已经走到尽头了。

**Structure tree:**
```
main clause: The reason is fairly fundamental
non-restrictive clause: which is that...
predicative clause: that the ability... has run its course
noun phrase: the ability for general purpose computing to continue to scale
```

**Grammar points:**
- **非限制性定语从句 + 表语从句** - which 引导定语从句，内含 that 引导的表语从句，双重嵌套
- **复杂名词短语** - the ability for... to... 结构，for 引出逻辑主语

### [01:41:14]
**Original:** Our mission was really to bring accelerated computing to the world and advance the type of applications that general purpose computing can't do, and scale to the level of capability that helps break through certain fields of science.

**Translation:** 我们的使命实际上是将加速计算带给世界，推进通用计算无法完成的应用类型，并扩展到能够帮助突破某些科学领域的能力水平。

**Core structure:**
- Our mission was to bring computing to the world, advance applications, and scale to the level of capability.  
  我们的使命是将计算带给世界，推进应用，并扩展到能力水平。

**Structure tree:**
```
main clause: Our mission was to...
three parallel infinitives: to bring... / and advance... / and scale...
relative clause 1: that general purpose computing can't do
relative clause 2: that helps break through certain fields
```

**Grammar points:**
- **并列不定式结构** - 三个 to 不定式并列作表语，最后一个省略 to
- **多个定语从句修饰** - 两个 that 从句分别修饰不同的先行词