Jensen Huang – Will Nvidia’s moat persist?

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。

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。

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?

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

AI is a five-layer cake, if you will.

AI 可以说是一个五层蛋糕。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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?

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?

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。

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。」

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。

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?

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。

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。

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。

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。

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。

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。

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。」

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。

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,

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。

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...

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。

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。

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。

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。

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。

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。

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。

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。

And 2x-ing at this scale now is really incredible.

确实如此。

Exactly.

确实如此。

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。

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%。

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%。

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?

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?

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?

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

It obviously has been tremendous for the company.

这对公司来说显然是巨大的成功。

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。

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。

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。

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。

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。

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,

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。

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。

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。

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。

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。

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。

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。」

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。

Guess what we're short of? Radiologists.

你猜我们现在缺什么?放射科医生。

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?

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。

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?

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。

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。

These things are not hard to replicate.

这些东西复制起来并不难。

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?

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?

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

The beautiful thing is you're talking to the expert.

好在你现在正在和专家对话。

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。

What does that mean for Nvidia going forward?

这对 Nvidia 未来意味着什么？

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。

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。

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。

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。

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，

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。」

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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……

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。

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。

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。

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?

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

That makes a lot of sense.

这很有道理。

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。

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。

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……

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。

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?

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。

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。

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?

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。

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。

These architectures are not as general purpose as a CPU.

这些架构不像 CPU 那样通用。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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?

You would choose an architecture that's most abundant.

你会选择最普及的架构。

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。

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。

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。

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。

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。

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。

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。

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?

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。

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。

No, I think your premise is wrong.

不,我认为你的前提是错的。

Maybe. But let me ask you a slightly different question.

也许吧。但让我换个问题问你。

Come back and make me correct your premise.

好的。让我换个问题问你。

Okay. Let me just ask you a different question.

好的。让我换个问题问你。

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——

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。

Yeah.

好。

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?

Obviously for Google, TPU is a majority of compute.

显然对 Google 来说,TPU 占了算力的大部分。

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。

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?

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。

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。

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。

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。

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。

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?

Sometimes you've got to be reminded of it.

有时候需要被提醒一下。

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。

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。

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?

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。

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。

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?

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。

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。

So, you asked the question why.

所以,你问了为什么这个问题。

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

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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?

We did it as soon as we could have.

我们是在能做的时候就尽快做了。

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。

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。

It wasn't in our sensibility to do so.

怎么说?是资金的问题吗?

How so? Was it like a cash thing?

怎么说?是资金的问题吗?

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。

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。

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。

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。

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。

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。

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。

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?

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

Why do you go out of your way not to pick winners?

你为什么特意不挑选赢家?

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。

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。

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。

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。

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?

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,

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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?

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?

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。

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?

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。

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。

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,

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。

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。

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。

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。

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。

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。

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……?

We never do that.

我们从不那样做。

Okay.

我们从不那样做。

We never do.

我们从不那样做。

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。

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。

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。

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。

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?

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。

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。

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。

The year after that, Feynman will come. And the year after that,

再下一年,Feynman 会推出。再下一年,

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。

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。」

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。

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。

You can say that about Nvidia today.

但今天你可以这样评价 Nvidia。

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。

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。

Or just one graphics card, okay, no problem.

或者只是一块显卡,好的,没问题。

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。

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。

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。

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。

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。

Okay. I want to ask about China.

好的。我想问问关于中国的问题。

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。

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?

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。

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。

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

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。

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。

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？

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。

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。

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。

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？

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。

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。

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。

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，

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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

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。

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。

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。」

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。

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。

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。

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。

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。

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。

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。」

I'm about to tell you.

我正要告诉你。

Okay.

好。

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?

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。

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。

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。

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。

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。

They have over-capacity, they have too much capacity.

他们产能过剩,产能太多了。

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。

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?

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。

They have plenty of compute already.

他们已经有足够的算力了。

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。

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。

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。

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。

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,

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。

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。

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。

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。

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?

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。

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——

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。

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。

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。

Huawei is a networking company.

Huawei 是一家网络公司。

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。

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。

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。

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。

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%。

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。

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。

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。

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?

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。

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。

Why would that stop being the case in the future?

为什么未来就不会是这样了?

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。

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。

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。

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。

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。

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——

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。

But American labs do that.

但美国实验室确实在这么做。

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?

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。

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。

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。

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?

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。

Right.

这不是好事。所以我们不能让它发生。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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?

There's no guarantee either way.

无论哪种方式都没有保证。

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

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?

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。

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。

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?

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。

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。

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。

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。」

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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?

We're doing everything we can to do that.

我们正在竭尽全力做到这一点。

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?

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。

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?

Yes.

Nvidia 呢,你认为 Nvidia 是美国公司吗?好吧。

Nvidia. You consider Nvidia a United States company? Okay.

Nvidia 呢,你认为 Nvidia 是美国公司吗?好吧。

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?

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?

The chip industry is part of the American ecosystem.

芯片产业是美国生态系统的一部分。

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。

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?

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。

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。

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。

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。

Who's sending enriched—

谁在运送浓缩——

The analogy is that enriched uranium is like compute.

这个类比是说浓缩铀就像算力。

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?

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?

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。

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。

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。

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。

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。

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?

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。

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

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。

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。

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。

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——

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。

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。

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。

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。

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。

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。

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。

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。

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。

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——

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。

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。

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。

No, I think what my argument is—

不,我认为我的论点是——

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。

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。

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。

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?

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。

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。

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。

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?

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?

Yes.

是的。

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?

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%。

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。

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。

It makes no sense to me.

这对我来说毫无意义。

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。

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。

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。

How can both of those things be true at the same time?

这两件事怎么可能同时为真?

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。

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。

Yeah.

是的。

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。

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。

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。

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。

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。

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……

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。

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。

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。

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。

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。

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。

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。

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。

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?

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。

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

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——

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。

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。

We're going to do our best.

我们会尽力而为。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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——

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。

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。

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。

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。

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。

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?

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。

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。

We should always have the best technology here.

我们应该始终把最好的技术留在国内。

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。

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。

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。

The world is just not absolutes.

世界本来就不是非黑即白的。

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。

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。

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。

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。

Their models have to be so far optimized for

他们的模型必须针对

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。

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。

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%。

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。

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。

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,

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。

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。

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,

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。

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。

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,

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。

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

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,

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

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。

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。

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?

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。

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。

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。

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。

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?

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。

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。

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。

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。

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。

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。

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。

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。

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。

But that market has only recently emerged.

但这个市场是最近才出现的。

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。

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。

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。

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。

That's the reason why we did it.

这就是我们这么做的原因。

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。

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。

The segmentation of it.

对,市场的细分。

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——

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。

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

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

If there were no AI, I would be very sad.

如果没有 AI,我会非常难过。

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。

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。

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。

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。

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

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。

Jensen, thank you so much.

Jensen,非常感谢你。

You're welcome. I enjoyed it.

我也是。

Me too.

我也是。