英伟达：OpenAI、计算的未来与美国梦 | BG2对话比尔·格利与布拉德·格斯特纳

我认为OpenAI很可能将成为下一个市值数万亿美元的超大规模企业。

I think that OpenAI is likely going to be the next multi trillion dollar hyperscale company.

Jensen，很高兴再次相聚，当然还有我的搭档Clark Tang。你知道吗，我简直不敢相信

Jensen, great to be back, of course, with my partner Clark Tang. You know, I can't believe it's

欢迎来到NVIDIA。

Welcome to NVIDIA.

噢，眼镜不错嘛。你戴着真的很合适。问题是现在大家都会要求你一直戴着它了。他们会说，就戴那副红框眼镜吧。这点我可以作证。

Oh, and nice glasses. Those actually look really good on you. The problem is now everybody's going want you to wear them all the time. They're going say, wear the red glasses. I can vouch for that.

距离我们上次录播客已经过去一年多了。是啊。如今你们超过40%的收入来自推理业务。但由于推理链的存在，推理技术即将迎来爆发。

So it's been over a year since we did the last pod. Yeah. Over 40% of your revenue today is inference. But inference is about ready because of chain of reasoning.

没错。对吧？规模即将增长十亿倍。对。就是就是就是

Yeah. Right? It's about to go up by a billion times. Right. By by by by

百万倍的量级

a million x by a

十亿级别，X。这是大多数人尚未完全理解的部分。这就是我们之前讨论的行业，但这是工业革命。

billion right. X. That's the part that most people have, you know, haven't completely internalized. This is that industry we were talking about, but this is the industrial revolution.

说实话，感觉自从那次之后，我们每天都在延续播客的对话。在AI时代里，这大概相当于一百年。我最近重听那期播客，我们讨论的许多突出观点中，对我影响最深的是你拍桌强调的那件事。记得当时预训练领域正处于低谷，人们都在说‘预训练的终结’。

Honestly, it's felt like you and I have had a continuation of the pod every day since then. Know, in AI time, it's been about one hundred years. I was rewatching the pod recently, and the many things that we talked about that stood out, the one that was probably most profound for me was you pounding the table. Remember at the time, there was kind of a slump in terms of pre training? And people were like, oh my The end of pre training.

没错，预训练的终结。我们过度建设了。这大概是一年半前的事。

Right. The end of pre training. We're overbuilding. This about is a year and a half ago. Yeah.

而你当时说推理不会增长100倍、1000倍，而是会增长10亿倍。嗯。这让我们走到了今天。你宣布了那个重大合作，我们该从这里说起。

And you said inference isn't going to 100x, 1000x. It's going to 1,000,000,000 x. Mhmm. Which brings us to where we are today. You know, you announced this huge deal, We ought to start there.

我低估了。我要正式声明：我确实低估了。现在我们有三条缩放定律对吧？首先是预训练缩放定律。

I underestimated. Let me just go on record. I underestimated. We now have three scaling laws, right? We have pre training scaling law.

其次是训练后缩放定律。训练后阶段本质上是AI的练习过程——反复尝试不同方法直到掌握技能。要实现这点，就必须进行推理。

We have post training scaling law. Post training is basically like AI practicing. Practicing a skill until it gets it right. And so it tries a whole bunch of different ways. And in order to do that, you've got to do inference.

所以现在训练和推理通过强化学习被整合在一起，非常复杂。这就是所谓的训练后阶段。第三条是推理定律——传统推理方式是单次完成的。

So now training and inference are now integrated in reinforcement learning. Really complicated. And so that's called post training. And then the third is inference. The old way of doing inference was one shot.

但我们所推崇的新推理方式是思考。所以在回答前要先思考。现在你有了三条扩展定律。思考时间越长，得到的答案质量越高。在思考过程中，你会进行研究。

But the new way of doing inference, which we appreciate, is thinking. So think before you answer. And so now you have three scaling laws. The longer you think, the better the quality answer you get. While you're thinking, you do research.

你会去核实一些基本事实。学到一些东西。再深入思考。继续学习更多。然后才生成答案。

You go check on some ground truth. And you learn some things. You think some more. You go learn some more. And then you generate an answer.

不要不假思索就直接生成答案。因此在训练后、训练前的思考阶段，我们现在有三条扩展定律，而非一条。

Don't just generate right off the bat. And so thinking, post training, pre training, we now have three scaling laws, not one.

你们去年就知道这个，但今年你们对推理能力将达到10亿倍增长的信心程度如何？这对智能水平意味着什么？信心更高了吗？相比去年，你们今年是否更有把握？

You knew that last year, but is your level of confidence this year in the inference is going to 1,000,000,000 x and where that will take the levels of intelligence? Is it higher? Are you more confident this year than you were a year ago?

我今年更有信心了。原因在于看看现在的代理系统。AI已不仅是语言模型，而是由多个语言模型组成的系统。它们都在并行运行，可能还使用各种工具。

I'm more confident this year. And the reason for that is because look at the agentic systems now. And AI is no longer a language model. And AI is a system of language models. And they're all running concurrently, maybe using tools.

我们中有些在使用工具，有些在进行研究。还有一大堆其他功能。没错，这都是多模态的。

Some of us are using tools. Some of us are doing research. And There's a whole bunch of stuff. Okay. It's all multi modality.

看看已经生成的所有视频内容。

Look at all the video that's been generated.

简直疯狂 这确实让我们联想到本周那个所有人都在谈论的关键时刻——几天前你宣布与OpenAI Stargate达成重大协议，你们将成为首选合作伙伴，在未来一段时间内向该公司投资1000亿美元。他们将建设10个超级计算中心。如果他们采用英伟达的技术来构建这10个中心，英伟达可能获得超过4000亿美元的收入。请帮我们解读一下，简单谈谈这次合作对你的意义，以及为何这项投资对英伟达如此明智。

Just crazy It really brings us to the seminal moment this week that everybody's talking about the massive deal you announced a couple days ago with OpenAI Stargate, where you're going to be a preferred partner, invest $100,000,000,000 in the company over a period of time. They're going to build 10 gigs. And if they used NVIDIA for those 10 gigs, that could be upwards of $400,000,000,000 in revenue to NVIDIA. So help us understand. Just tell us a little bit about that partnership, what it means to you, right, and why that investment makes so much sense for NVIDIA.

首先我要先回答最后一个问题，然后再以我的方式展开讨论。我认为OpenAI很可能会成为下一个市值数万亿美元的超级规模企业。你我为何称其为超级规模企业？

So first of all, I'll answer that last question first. Then I'll come back and raise it in my way. Good. Think that OpenAI is likely going to be the next multi trillion dollar hyperscale company. I think you and I Why do you call it a hyperscale company?

就像Meta和谷歌这样的超级规模企业。他们将同时拥有消费者和企业服务，极有可能成为全球下一个市值数万亿美元的超级巨头。我想你也会认同这个观点。

Hyperscale like Meta's a hyperscale. Google's a hyperscale. They're going to have consumer and enterprise services. And they are very likely going to be the world's next multi trillion dollar hyperscale company. And I think you would agree with that.

我

I

同意。若真如此，在他们达到那个规模前进行投资，这可能是我们能想象到的最明智的投资之一。你必须投资你熟悉的领域，而事实证明我们恰好了解这个领域。因此这项投资的机会，其回报将会非常可观。

agree. If that's the case, the opportunity to invest before they get there this is some of the smartest investments we can possibly imagine. And you had to invest in things you know. And it turns out we happen to know this space. And so the opportunity to invest in that, the return on that money is going to be fantastic.

所以我们很珍惜这个投资机会。我们并非必须投资，也没有被强制要求，但他们给予我们这个机会实在难得。现在让我从头说起。

So we love the opportunity to invest. We don't have to invest. And it's not required for us to invest, but they're giving us the opportunity to invest. Fantastic thing. Now let me start from the beginning.

我们正与OpenAI合作多个项目。首个项目是微软Azure的扩建工程，我们将持续推进，目前合作进展非常顺利。我们还有数年的扩建工作要完成，仅这部分就涉及数千亿美元的业务规模。

So we're partnering with OpenAI in several projects. The first project is the build out of Microsoft Azure. We're going to continue to do that. And that partnership is going fantastically. We have several years of build out to do, hundreds of billions of dollars of work just to do there.

其次是OCI的扩建项目。我认为大约有五、六、七个吉瓦的容量即将建成。我们正与OCI、OpenAI和软银合作推进这些已签约项目，目前正在实施中。

The second is the OCI build out. And I think there's some five, six, seven gigawatts that are about to be built out. And so we're working with OCI and OpenAI and SoftBank to build that out. Those projects are contracted. We're working on it.

还有很多工作要做。第三部分是Core Weave。关于CoreWeave四点——我仍在讨论OpenAI。是的，所有内容都在OpenAI的语境下展开。

Lots of work to do. And then the third is Core Weave. And so of CoreWeave four, I'm talking about OpenAI still. Yes. Okay, everything in the context of OpenAI.

那么问题在于：这个新合作是什么？这是首次协助OpenAI建立自主AI基础设施的合作伙伴关系。我们将直接在芯片层、软件层、系统层和AI工厂层面与OpenAI合作，帮助他们成为全面运营的超大规模企业。这将持续相当长时间，用于补充他们正在经历的双重指数级增长。

And so the question is, what is this new partnership? This new partnership is about helping OpenAI, working partnering with OpenAI to build their own self build AI infrastructure for the first time. And so this is us working directly with OpenAI at the chip level, at the software level, at the systems level, at the AI factory level to help them become a fully operated hyperscale company. This is going to go on for some time. It's going to supplement the amount of they're going through two exponentials, as you know.

第一个指数增长是客户数量呈指数级上升。这是因为AI技术持续进步，应用场景不断优化。现在几乎每个应用程序都与OpenAI对接，因此他们正经历使用量的爆发式增长。

The first exponential is the number of customers is growing exponentially. And the reason for that is the AI is getting better. The use case is getting better. Just about every application is connected to OpenAI now. And so they're going through the usage exponential.

第二个指数是每次使用的计算量激增，明白吗？不再是一次性推理，现在系统会在回答前进行思考。这两个指数效应叠加放大了他们的算力需求。我们能够完成所有这些不同项目，而最新合作是在已公布项目和现有合作基础上的增量补充。

The second exponential is the computational exponential of every use, right? Instead of just a one shot inference, it's now thinking before it answers. And so these two exponential is compounding their compute requirements. And so we could build out all these different projects. And so this last one is an additive on top of everything that they've already announced, all the things that we're already working on with them.

这是原有基础上的叠加，将支撑这种惊人的指数级增长。

It's additive on top of that, And it's going to support this incredible exponential growth.

你提到的观点让我很感兴趣——你认为他们极有可能成为市值数万亿美元的公司，是个绝佳投资。但与此同时，他们正在自主建设数据中心，而你们正协助这个过程。要知道此前他们一直将数据中心建设外包给微软。

One of the things you said there that's really interesting to me is kind of they're going to be a high probability, multi trillion dollar company in your mind. You think it's a great investment. At the same time, they're self building. You're helping them self build their data centers. So heretofore, they've been outsourcing to Microsoft to build the data center.

现在他们想自己建立全栈工厂。

Now they want to build full stack factories themselves.

他们本质上希望与我们建立类似埃隆与X公司那样的关系。没错，我是说埃隆和X公司销售成品。

They want to basically have a relationship with us the way that Elon and X has relationships. Correct. I mean, Elon and X sell built.

正是如此。但我认为这

Exactly. But I think that's

这一切都是件大事。

This all is a very big deal.

想想巨像公司的优势，他们正在构建全栈。那就是超大规模供应商。因为如果他们不使用这些产能，可以转售给其他人。

When think about the advantage that Colossus had, they're building full stack. That is a hyperscaler. Because if they don't use the capacity, they could sell it to somebody else.

没错。

That's right.

同样地，星际之门项目他们正在建设庞大产能。虽然预计会自用大部分，但也具备了对外销售的条件。听起来很像AWS、GCP或Azure的模式。你就是这个意思吧。

And the same way Stargate they're building monstrous capacity. They think they'll need to use most of it, But it puts them in a position to sell it to somebody else as well. Sounds very much like AWS or GCP or Azure. That's what you're saying.

是的。我认为他们很可能会自己使用它。就X的情况而言，他们很可能也会自行使用。但他们希望与我们建立同样的直接关系，包括直接的工作关系和直接的采购关系。就像扎克和Meta与我们之间的关系一样，完全是直接的。

Yeah. I think they'll likely use it themselves. And just in the case of X, they'll likely use it themselves. But they would like to have the same direct relationship with us, direct working relationship and direct purchasing relationship. Meta, just as what Zuck and Meta has with us, it's exactly a direct.

我们与桑达尔及谷歌之间的关系是直接的。我们与萨提亚及Azure的合作也是直接的。难道不是吗？他们已经发展到足够大的规模，认为现在是时候开始建立这些直接关系了。我很高兴能支持这一点。

Our relationship between us and Sundar and Google, direct. Our partnership with Satya and Azure, direct. Isn't that right? And so they've gotten to a large enough scale that they believe it's time for them to start building these direct relationships. I'm delighted to support that.

萨提亚知道这一点，拉里也知道，每个人都清楚正在发生的事情，并且都非常支持。

Satya knows it, and Larry knows it, and everybody's aware of what's going on, and everybody's very supportive of it.

所以我觉得有件事很神秘，对吧？你刚才提到了Oracle的3000亿，Colossus，他们在建造什么。我们知道主权国家在建造什么。我们知道超大规模供应商在建造什么。山姆谈论的是以万亿计的规模。

So one of the things I find mysterious, right? You you just mentioned Oracle 300,000,000,000, Colossus, what they're building. We know what the sovereigns are building. We know what the hyperscalers are building. You know, Sam's talking in terms of trillions.

但在华尔街覆盖你们股票的25位卖方分析师中，如果我看共识预期，基本上认为你们的增长将从2027年开始趋于平缓。2027年到2030年2月增长8%。明白吗？这就是那25个人的唯一工作。他们拿钱就是为了预测英伟达的增长率。

But of the 25 sell side analysts on Wall Street who cover your stock, if I look at the consensus estimate, it basically has your growth flatlining starting in 2027. 8% growth 2027 through 02/1930. Okay? That is the 25 people in their only job. They get paid to forecast the growth rate for NVIDIA.

所以显然

So clearly

顺便说一句，我们对这种情况感到满意。听着，我们对此很坦然。我们经常能轻松超越这些数字。

We're comfortable with that, by the way. Right. Look, we're comfortable with that. Okay. We have no trouble beating the numbers on a regular basis.

对，不，我明白这一点。但这里存在一个有趣的脱节现象，对吧？

Right. No. I understand that. But there is this interesting disconnect. Right?

我每天都在CNBC和彭博社听到这种论调。我认为这涉及到关于短缺最终会导致过剩的质疑，他们并不相信这种说法。他们会说，好吧，我们姑且认可你们对2026年的预测，但2027年可能会出现产能过剩，到时候你们就不需要那么多了。但我觉得很有意思的是，必须指出你们的共识预测认为这种情况不会发生，对吧？

I hear it every day on CNBC and Bloomberg. And I think it goes to some of these questions around shortages leading to a glut that they don't believe. They say, Okay, we'll give you credit for '26, but '27, maybe we'll have too much and you're not going to need that. But it is interesting to me, and I think it's important to point out that your consensus forecast is that this won't happen. Right?

我们还综合考虑所有这些数据为公司制定了预测。这向我表明，尽管我们已进入AI时代两年半，但在Sam Altman、你、Sundar、Satya所表述的观点与华尔街仍然相信的内容之间，仍存在巨大的认知分歧。而且，你也清楚，你对此并不感到困扰。

And we also put together forecasts for the company, taking into account all of these numbers. And what it shows me is still, even though we're two and a half years into the age of AI, a massive divergence of belief between what we hear Sam Altman saying, you saying, Sundar saying, Satya saying, and what Wall Street still believes. And, you know, again, you're comfortable with that.

我也不认为这存在矛盾。

I also don't think it's inconsistent.

好的，那请解释一下

Okay. So explain that a

首先，作为建设者，我们应该为机遇而建设。我们是建设者。让我提出三个要点供你思考，这些要点希望能让你对英伟达的未来更有信心。第一点是关于物理定律的。

little So first of all, for the builders, we're supposed to be building for opportunity. We're builders. Let me give you three points to think through. And these three points, it'll help you hopefully be more comfortable with NVIDIA in this future. So the first point and this is the laws of physics point.

这是最重要的一点：通用计算的时代已经结束，未来属于加速计算和AI计算。这是第一点。由此可以想到，全球价值数万亿美元的计算基础设施都需要更新换代，对吧。

This is the most important point that general purpose computing is over and the future is accelerated computing and AI computing. That's the first point. And so the way to think about that is there's how many trillions of dollars of computing infrastructures in the world that has to be refreshed. Right.

当它刷新时，将会被加速。没错。

And when it gets refreshed, it's going to be accelerated That's right.

所以首先要认识到的是，通用计算已终结——这一点无人质疑。所有人都表示赞同。通用计算时代结束了，摩尔定律已死。人们常这么说。

And so the first thing you have to realize is that general purpose computing, and nobody disputes that. Everybody goes, yeah, we completely agree with that. General purpose computing is over. Moore's Law is dead. People say these things.

这意味着什么？通用计算将转向加速计算。我们与英特尔的合作正是认识到：通用计算需要与加速计算融合，为他们创造机遇。对吗？因此第一点，通用计算正在向加速计算和AI转型。

And so what does that mean? So general purpose computing is going to go to accelerated computing. Our partnership with Intel is recognizing that general purpose computing needs to be fused with accelerated computing to create opportunities for them. Is that right? And so one, general purpose computing is shifting to accelerated computing and AI.

第二点，AI的首个应用场景其实已无处不在——在搜索、推荐引擎中，对吧？在购物领域。超大规模计算基础设施过去由CPU处理推荐系统，现在正转向GPU处理AI。

Two, the first use case of AI is actually already everywhere. It's in search, recommender engines Isn't that right? In shopping. The basic hyperscale computing infrastructure used to be CPUs doing recommenders. It's now going to GPUs doing AI.

传统计算正向加速计算AI转型。超大规模计算正从CPU转向加速计算和AI。这是第二点。仅帮助Meta、谷歌、字节跳动、亚马逊将其传统超大规模计算方式转向AI，就涉及数千亿美元。

So you just take classical computing, it's going to accelerated computing AI. You take hyperscale computing, it's going from CPUs to accelerated computing and AI. And then now that's the second point. Just feeding the metas, the Googles, the ByteDances, the Amazons and take their classical traditional way of doing hyperscaling and moving it to AI, that's hundreds of billions of dollars.

考虑到TikTok、Meta和谷歌的用户，全球可能有40亿人已在需要由加速计算驱动的工作负载。确实如此。

And because that may be 4,000,000,000 people on the planet today if you take TikTok, Meta into account That's right. Google into account who are already demanding workloads that are driven by accelerated

完全正确。甚至无需考虑AI创造的新机遇，关键在于AI正在改变传统做事方式。现在让我们谈谈未来——目前为止我讨论的都只是基础层面。非常基础的内容。

That's exactly right. And so without even thinking about AI creating new opportunities, it's about AI shifting how you used to do something to the new way of doing something. And then now let's talk about the future. So far I've only spoken largely about just mundane stuff. Just mundane stuff.

旧方法现在行不通了。你们将不再使用燃油灯，而是转向电力。就这样。

The old way is now wrong. You're gonna go you're no longer gonna use fuel light lanterns. You're gonna go to electricity. That's all. Right.

明白吗？你们也不能再依赖螺旋桨飞机了，要改用喷气机。仅此而已。目前我只谈了这些。

Okay? And you can no longer, you know, prop planes. You're going go to jets. That's all. And so far, that's all I've talked about.

而最不可思议的是，当你转向人工智能，转向加速计算时会发生什么？会涌现哪些新应用？这就是我们讨论的所有AI相关内容。这个机会看起来如何？简单来说，就像电动机取代体力劳动一样，我们现在有了AI——这些我常说的AI超级计算机和AI工厂。

And then now that the incredible thing is when you go to AI, when you go to accelerated computing, then what happens? What are the new applications that emerge as a result? And that's all the AI stuff that we're talking about. And that opportunity, what does that look like? Well, the simple way of thinking about that is where motors replace labor and physical activity, we now have AI, these AI supercomputers, these AI factories that I talk about.

它们将生成标记来增强人类智能，对吧？人类智能约占全球GDP的55%到65%，大约50万亿美元。这50万亿美元将被某种力量增强。以个人为例：假设我雇佣一名10万美元的员工，再花1万美元用AI增强他，结果这个AI让员工效率提升两到三倍，我会这么做吗？

They're going to generate tokens to augment human intelligence, right? And human intelligence represents, what, 55%, 65% of the world's GDP. Let's call it $50,000,000,000,000 And that $50,000,000,000,000 is going to get augmented by something. And so let's come back to a single person. Suppose I were to hire a $100,000 employee and I augmented that $100,000 employee with a $10,000 AI, and that $10,000 AI, as a result, made that $100,000 employee twice more productive, three times more productive, would I do it?

毫无疑问。我现在正对公司每个员工实施这个方案，对吧？

Heartbeat. I'm doing it across every single person in our company right now, right?

他们都有AI协作者。

They all have co agents.

没错。就是同事。完全正确。我们公司每个软件工程师、每个芯片设计师都已配备AI助手，覆盖率100%。

That's right. Co workers. That's right. Every single software engineer, every single chip designer in our company already has AIs working with them. 100% coverage.

因此，我们制造的芯片数量更优，且持续增长。我们的推进节奏恰到好处，公司发展也因此更为迅猛，进而需要招募更多人才。

As a result, the number of chips we're building is better. The number is growing. The pace at which we're doing it is right. And so we're growing faster as a company. As a result, we're hiring more people.

我们的生产效率更高，营收规模更大，盈利能力更强。这有什么不值得欣喜的呢？现在将英伟达的案例映射到全球GDP上。

Our productivity is greater. Our top line is greater. Our profitability is greater. What's not to love about that? Now apply the NVIDIA story to the world's GDP.

未来可能发生的是，现有的50万亿美元规模将额外增加——假设增加10万亿美元。这新增的10万亿美元需要运行在机器上。AI与过去IT的根本差异在于：传统软件是预先编写后在CPU上运行，由人工操作；而未来AI将持续生成数据标记（tokens），由机器自主完成思考过程。

And so what's likely to happen is that that $50,000,000,000,000 is augmented by let's pick a number $10,000,000,000,000 That $10,000,000,000,000 needs to run on a machine. Now the reason that AI is different than IT in the past, in a way software was written a priori and then it runs on a CPU. And it doesn't it runs a person would operate it. In the future, of course, AI is generating tokens. But a machine has to generate the tokens and it's thinking.

这意味着软件在持续动态运行。过去的软件是一次性编写的产物，而现在的软件实际上在实时创作——它具备思考能力。而AI要实现思考，就需要配套的工厂设施。

So that software is running all the time. Whereas in the past, the software was written once. Now the software is in fact writing all the time. It's thinking. In order for the AI to think, it needs a factory.

假设这10万亿美元标记生成业务有50%毛利率，其中5万亿美元需要建设工厂——即AI基础设施。若告诉我全球年度资本支出约为5万亿美元，我会认为这个推算合乎逻辑。这就是未来图景：从通用计算转向加速计算，用AI替代所有超大规模计算，进而增强全球GDP中的人类智能。

And so let's say that that $10,000,000,000,000 of token generated, 50% gross margins, and 5,000,000,000,000 of it needs a factory, needs an AI infrastructure. So if you told me that on an annual basis, the CapEx of the world was about $5,000,000,000,000 I would say the math seems to make sense. And that's kind of the future, right? Going from general purpose computing to accelerated computing, replacing all the hyperscales with AI, and then now augmenting human intelligence for the world's GDP.

目前这个市场规模据我们估算约为每年4000亿美元。是的，所以总潜在市场（TAM）将是当前规模的4到5倍。

And today, that market is about our estimate is about 400,000,000,000 annually. Yep. So the the TAM, you know, is is a four to five x increase over where it is today.

没错。昨晚阿里巴巴的吴泳铭提到，从现在到本年代末——抱歉纠正下，是到2030年底——他们将把数据中心算力提升10倍。对吧？确实如此。

Yeah. Eddie last night, Eddie Wu at Alibaba said, between now and the end of the year and excuse me. Now and the end of the decade, they're going to increase their data center power by 10x. Right? Right.

你刚才说多少倍来着？4倍。就是这样。

You just said how much? 4x. There you go.

就是这样。对。

There you Yeah.

他们打算将性能提升10倍，而我们则把性能与营收挂钩。英伟达的营收几乎与性能成正比。没错吧？是的。

They're gonna increase power by 10 x, and we we correlate the power. NVIDIA's revenue is almost correlated to power. Isn't that right? Yeah.

没错。对。因为更多的

That's right. Yeah. Because more

他还提到哪些方面？对。他说代币生成量每隔几个月就翻一番。是啊。这话什么意思？

What slots else did he say? Yeah. He said token generation is doubling every few months. Yeah. What's that saying?

每瓦性能必须持续呈指数级增长。这就是英伟达拼命提升每瓦性能的原因。而每瓦营收本质上就是未来的总收入。

The perf per watt has to keep on going exponentially. That's why NVIDIA's cranking it out with perf per watt. And revenue per watt watt is basically revenues in this future.

在这个假设中，我发现从历史背景看非常耐人寻味。两千年来，GDP基本没有增长。明白吗？然后工业革命来了，GDP开始加速增长。

Embedded in this assumption, I find it very fascinating in historical context. For two thousand years, basically GDP did not grow. Okay? And then we get the Industrial Revolution. GDP accelerates.

我们迎来了数字革命。GDP加速增长。正如你和斯科特·贝松所言，他预测明年GDP增长率将达到4%。本质上，你们的意思是全球GDP增长将提速，因为现在我们正为世界提供数十亿的‘数字同事’来替我们工作。

We get the Digital Revolution. GDP accelerates. And basically what you're saying and Scott Besson has said it. He said, I think we're going to have 4% GDP growth next year. Basically what you're saying is the world's GDP growth is going to accelerate because now we are giving the world billions of coworkers that will do work for us.

如果GDP代表固定劳动力和资本投入下的产出总量，那么它必然加速增长。

And if GDP is an amount of output for a fixed amount of labor and capital, it has to accelerate.

必须加速。看看AI的现状——由于AI技术（我们姑且称之为大语言模型和各类AI代理）的发展，它正在催生一个全新的AI代理产业，这是毋庸置疑的。

It has to. It has to. Look at what's going on with AI. As a result of the technology of AI and that technology of AI, let's just call it the large language models and all the AI agents, it's now creating a new industry of AI agents. There's no question about that.

OpenAI已成为史上营收增长最快的公司，且呈指数级扩张。AI本身就是一个高速发展的行业，因为它需要背后的工厂和基础设施支撑，这个行业在壮大，我的行业也在壮大。

OpenAI is the fastest growing revenue company in history. And they're growing exponentially. And so AI itself is a fast growing industry. Because AI needs a factory behind it, an infrastructure behind it, This industry is growing. My industry is growing.

正因我的行业在增长，我们底层的基础产业也随之增长。能源行业正在蓬勃发展。

And because my industry is growing, the industry underneath us is growing. Energy is growing.

电力、壳牌、这个

Power, Shell, This

对能源行业而言就像文艺复兴，不是吗？核能、燃气轮机...看看我们底层基础设施生态中所有这些企业，它们表现惊人，每个领域都在增长。

is like renaissance for the energy industry, isn't that right? Nuclear energy, gas turbines. I mean, look at all of those companies in infrastructure ecosystem underneath us. They're doing incredibly well. Everybody's growing.

这些数字让所有人都在讨论供应过剩或泡沫，对吧？扎克伯格上周在播客中说过，听着，我认为很可能在某个时点我们会遇到空中陷阱，Meta实际上可能会超支100亿美元或更多。但他说这没关系。这对他们业务的未来至关重要，是他们必须承担的风险。

These numbers have everybody talking about a glut or a bubble, Right? Zuckerberg said last week on a podcast, you know, he said, listen. I think it's quite possible at some point that we will have an air pocket, and Meta may in fact overspend by $10,000,000,000 or whatever. But he said it doesn't matter. It's so existential to the future of his business that it's a risk that they have to take.

但当你细想时，这听起来有点像囚徒困境，对吧？请再给我们梳理一遍

But when you think about that, it sounds a little bit like Prisoner's Dilemma, right? And walk us again through

这些可是非常快乐的囚徒。

These are very happy prisoners.

再带我们过一遍，根据我们今天的预估，到2026年AI收入将达到1000亿美元（不包括Meta和运行推荐引擎的GPU）。这里还包括

Walk us again through, right, today our estimate is that we're going to have $100,000,000,000 of AI revenue in 2026, excluding meta and excluding the GPUs running recommender engines. There's In

搜索或修正。

Search or Correct.

所以还有其他部分。是的。但我们就按1000亿美元来算

So there's other stuff. Yeah. But let's call it $100,000,000,000

那到底是什么行业？已经是超大规模的那个行业？在万亿级别之间的超大规模到底是什么。没错，顺便说一句，那个行业正在向AI转型。

What is that industry anyways? What is the industry already in hyperscale? What is the hyperscale between Trillions. Exactly. By the way, that industry is going to AI.

在任何人从零开始时，你必须从那里起步。

Before anybody starts at zero, you've got to start there.

但我想怀疑论者会说，我们需要从2026年1000亿美元的AI收入增长到至少2030年2月的1万亿美元AI收入。明白吗？你刚才提到全球GDP大约5万亿美元。如果从底层做起，你能看到在未来五年内AI驱动的收入从1000亿增长到1万亿美元的路径吗？我们的增长速度有那么快吗？

But I think the skeptics would say we need to go from $100,000,000,000 of AI revenue in 'twenty six to at least $1,000,000,000,000 of AI revenue in 02/1930. Okay? You just were talking a minute ago about $5,000,000,000,000 when you look at kind of global GDP. If you did a bottoms up, can you see your way to $1,000,000,000,000 of AI driven revenues from $100,000,000,000 over the course of the next five years? Are we growing that fast?

是的。

Yes.

而且我还要说，我们已经达到了这个水平。

And I would also say we're already there.

好的，请解释一下。

Okay. So explain that.

因为超大规模服务商已经从CPU转向了AI。他们的整个收入基础现在都是由AI驱动的。没错，没有AI就无法运行TikTok。

Because the hyperscalers, they went from CPUs to AI. Okay. Their entire revenue base is all now AI driven. Correct. You can't do TikTok without AI.

没错，没有AI就无法制作YouTube Shorts。没有AI，所有这些都无法实现。Meta正在做的定制化内容、个性化内容这些惊人的事情，没有AI就办不到。所有这些过去都是由人类预先制作内容，创造四个选项再由推荐系统选择。而现在是由AI生成无限数量的选择，对吧？

Correct. You can't do YouTube Short without AI. You can't do any of this stuff without AI. The amazing things that Meta's doing for customized content, personalized content, you can't do that without AI. All of that stuff used to be humans doing content a priori, creating four choices that are then selected by a recommender And now it's infinite number of choices generated by an AI, right?

但这些事情已经像是我们经历了从CPU到GPU的转型，主要是为了那些推荐引擎。而现在它们正在

But those things are already like we had the transition from CPUs to GPUs largely for those recommender engines. And now they're going

嗯，那相当新。会

Well, that's fairly new. Would

在过去三四年里，是的。

In the last three or four years, Yeah.

扎克会告诉你我在SIGGRAPH大会上，扎克也会承认他们在采用

Zuck would tell you I was at SIGGRAPH, and Zuck would tell you they were late getting to

GPU方面。为了

GPUs. For

确实。Meta使用GPU才多久？两年？一年半？相当新。用GPU进行搜索？

sure. GPUs for Meta is what, a couple of years? A year and a half? It's pretty new. Search with GPUs?

毫无疑问。崭新崭新的。确实。确实。崭新崭新的。

For sure. Brand spanking new. For sure. For sure. Brand spanking new.

在GPU上搜索GPU？

Search for GPUs on GPUs?

所以你的论点是，到2030年我们将拥有万亿美元规模的AI收入几乎是确定无疑的，因为我们几乎已经达到了。已经达到了。好吧。我们就来谈谈增量增长。增量。

So your argument would be the probability that we're going to have a trillion dollars of AI revenues by 2030 is near certain because we're almost Already there. Already there. Okay. Let's just talk about incremental growth. Incremental.

现在我们可以讨论增量增长了。正是如此。

Now we can talk about incremental Incremental growth. Exactly.

对。你可以自下而上或自上而下分析，我刚听你提到全球GDP占比的自上而下观点。你认为未来三到五年内出现供应过剩的概率百分比是多少？这是一个我们无法预知未来的概率分布问题。

Right. You do your bottoms up or your tops down, I just heard your tops down about percentage of global GDP. What is the percentage probability that you think will have a glut, will run into a glut in the next three or four or five years? It's a distribution of we don't know the future. It's a distribution of probability.

在我们把所有通用计算完全转化为加速计算和AI之前，我认为出现这种情况的概率极低。

Until we fully convert all general purpose computing to accelerated computing and AI, until we do that, I think the chances are extremely low.

好的。这还需要几年时间。

Okay. And that will take a few years.

确实还需要几年时间。

That'll take a few years.

让我再问一个问题。

Let me ask one more.

直到所有推荐引擎都基于人工智能，直到所有内容生成都基于人工智能，因为面向消费者的内容生成在很大程度上依赖于推荐系统等等。所有这些都将由AI生成。直到所有这些传统意义上的超大规模计算都转向AI领域——从购物到电子商务的一切，直到所有领域都完成转型。

Until all recommender engines are AI based, until all content generation is AI based, because content generation, consumer oriented content generation is very largely recommender systems and so on and so forth. And all of that's going to be AI generated. Until all of this stuff, what classically was hyperscale, now transitions to AI. Everything from shopping to e commerce to all that stuff until everything goes over.

但这些新建设投资，当我们谈论数万亿美元时，我们是在超前投资。这是可以随时调整的吗？即使看到增速放缓或供应过剩迹象，你们是否仍必须继续投资？还是说这更像是对生态系统的号召，告诉大家去建设，而如果出现放缓迹象，我们随时可以收缩投资规模？

But all this new build, right, when we're talking about trillions, we're investing ahead of where we are. You know, is that like at will? Are you obliged to invest the money even if you see a slowdown or a kind of a glut coming? Or is this one of these things that you're just waving the flag to the ecosystem to say, get out and build. And at some point in time, if we see some of this slow down, we can always pull back on the level of investment.

实际上恰恰相反，因为我们处于供应链末端，明白吗？我们是根据需求做出反应的。现在所有风投都会告诉你——你们也知道——当前存在算力短缺。不是因为全球GPU短缺（虽然确实短缺），而是只要接到订单我们就会建设。过去几年我们已彻底打通了供应链。

Actually, it's the other way because we're at the end of the supply chain, right? And so we respond to demand. And right now, all the VCs will tell you, and you guys know, the demand there's a shortage of compute in Not the because there's a shortage of GPUs in the world, Okay? If they give me an order, I'll build it. Over the last couple of years, we've really plumbed the supply chain.

从晶圆投料到Co-op封装再到HBM内存，整个后端供应链技术我们都已准备就绪。如果需要产能翻倍，我们就能翻倍。供应链已整装待发，现在只等待需求信号。当云服务提供商、超大规模企业和客户提交年度计划预测时，我们就会响应。

So all of the supply chain behind me from wafer starts to co op to HBM memories, all of that technology, we've really geared up. If we need to double, we'll double. So the supply chain is ready. Now we're just waiting for demand signals. And when the CSPs and the hyperscalers and our customers do their annual plan and they give us their forecast, we respond to that.

我们按预测进行建设。但现实情况是，他们提供的每个预测最终都被证明是低估的。所以我们总是处于紧急赶工状态，这种状态已持续两年。无论收到什么预测，相比去年都有显著增长，但永远不够。

And we build to that. Now what's going on, of course, is that every one of their forecasts that they provide us turns out to have been wrong because they underforecasted. And so now we're always in a scramble mode. And so we've been in the scramble mode now for a couple of years. And whatever forecast we've been given has been always significant increase from last year, but not enough.

去年萨提亚似乎有所收敛，有人称他是清醒的决策者，在抑制某些预期。但几周前他说今年已建成两个超算中心，未来还将加速。你是否看到那些传统超大规模企业——可能比Core Weave、Elon X或Stargate动作稍慢的——现在听起来他们都在加大投入，并且都在积极探索？

Satya last year seemed to be pulling back a little bit. Seemed to be, some people called him the adult in the room, tamping down kind of some of these expectations. A few weeks ago he said, Hey, I've also built two gigs this year, and we're going to accelerate in the future. Do you see some of the traditional hyperscalers that may have been moving a little slower than, let's call it, a Core Weave or Elon X or maybe a little slower than Stargate. Do you see them all it sounds like to me they're all leaning in more now and they're all also exploring.

因为第二个指数级增长。好的。我们已经经历了一个指数级增长，那就是AI的采用率。AI的参与度正在呈指数级增长。而突然出现的第二个指数级因素是推理能力。

Because of the second exponential. Okay. We've already had one exponential we were experiencing, which was the adoption rate of AI. The engagement of AI was growing exponentially. The second exponential that kicked in was reasoning.

这就是我们一年前的那场对话。对。我们当时说，听着，当AI从单次记忆答案和泛化——对吧？记忆和泛化本质上就是重新训练——转变的那一刻。

That was the conversation we had one One year year ago. Yeah. We said, hey, listen. The moment you take AI from one shot, memorizing an answer and general right? Memorizing and generalizing, that's basically retraining.

比如记住一个答案，八乘八是多少？直接记住就行。明白吗？所以记忆答案和泛化，那是单次AI。而一年前，推理能力出现了。

So memorizing an answer, what's eight times eight? Just memorize it. Okay? And so memorizing an answer and generalizing, that was one shot AI. Now a year ago, reasoning came about.

研究能力出现了。工具使用能力出现了。现在你面对的是会思考的AI。十亿倍的提升。这将消耗更多的计算资源。

Research came about. Tool use came about. And now you're a thinking AI. 1,000,000,000 X. It's gonna use a lot more compute.

正如你所说，某些超大规模客户内部有必须从通用计算迁移到加速计算的工作负载。所以他们挺过了这个周期。可能有些超大规模服务商的工作负载不同，所以不太确定能多快消化这些变化。

Certain hyperscale customers, to your point, had internal workloads that they had to migrate anyways from general purpose computing to accelerated computing. So they built through the cycle. I think maybe some hyperscalers had different workloads, so they weren't quite sure how quickly they could digest it.

没错。

That's right.

现在所有人都意识到他们严重低估了建设需求。

Everyone has now concluded that they dramatically underbuilt.

我最喜欢的应用之一就是传统的数据处理，包括结构化数据和非结构化数据。纯粹的传统数据处理。很快，我们将宣布一个加速数据处理的重大计划。如今，全球绝大多数CPU都用于数据处理，而这仍完全依赖于CPU运行。

One of the applications that my favorite is just good old fashioned data processing, Structured data and unstructured data. Just good old fashioned data processing. And very soon, we're going to announce a very big initiative of accelerated data processing. Data processing represents the vast majority of the world's CPUs today. It still completely runs on CPUs.

如果你去Databricks，主要是CPU。去Snowflakes，主要是CPU。Oracle的SQL处理，也主要是CPU。所有人都在用CPU处理SQL和结构化数据。未来，这些都将转向AI数据处理。

If you go to Databricks, it's mostly CPUs. If you go to Snowflakes, mostly CPUs. SQL processing at Oracle, mostly CPUs. Everybody's using CPUs to do SQL, structured data. In the future, that's all going to move to AI data.

这是一个我们将进入的庞大市场。但NVIDIA做的每件事都需要加速层和特定领域的数据处理库——就像菜谱一样。我们必须去构建这些，但它们即将到来。

That is one gigantic massive market that we're going to move to. But everything that NVIDIA does requires acceleration layers and requires domain specific data processing libraries. Recipes. We got to go build that. But that's coming.

昨天我打开CNBC时听到一些反对声音，他们说‘泡沫、过剩’。换到Bloomberg，又在讨论循环收入和虚假交易。所以为了帮助

So one of the pushbacks, I turned on CNBC yesterday, they were like, Oh, glut, bubble. When I turned on Bloomberg, it was about round tripping and circular revenues. And so for the benefit

家里的人

of

理解，这些操作是指公司通过误导性交易虚增收入，缺乏实际经济实质。换句话说，收入是靠财务工程而非真实客户需求支撑的。典型案例就是25年前泡沫时期的思科和北电。当你们或微软、亚马逊投资那些同时是你们大客户的公司时——比如你们投资OpenAI，而OpenAI正采购数百亿芯片——请告诉我们和所有人，Bloomberg那些分析师在炒作循环收入或虚假交易时到底错在哪里？

people at home, know, these arrangements are when companies enter into a misleading transaction that artificially inflates revenue without any underlying economic substance. So in other words, gross propped up by financial engineering, not by customer demand. And the canonical case everybody's referencing, of course, is Cisco and Nortel from the last bubble twenty five years ago. So when you guys or Microsoft or Amazon are investing in companies that are also your big customers in this case, you guys investing in OpenAI. While OpenAI is buying tens of billions of chips, just remind us and remind everybody else, what are the analysts on Bloomberg and otherwise getting wrong when they're hyperventilating about circular revenues or about round tripping?

10吉瓦相当于约4000亿美元。这4000亿美元主要需通过他们的承购收入（正呈指数增长）、自有资本（股权融资所得）以及可获得的债务融资来支撑。就这三个渠道。而他们能募集的股权和债务资金，很大程度上取决于市场对其可持续收入的信心。这是肯定的。

10 gigawatts is like $400,000,000,000 something And like that $400,000,000,000 will have to be largely funded by their offtake, their revenues, which is growing exponentially. It has to be funded by their capital, the money they've raised through equity, and whatever debt they can raise. Those are the three vehicles. And the equity that they could raise and the debt that they could raise has something to do with the confidence of the revenues that they could sustain. For sure.

因此，精明的投资者和贷款方会综合考虑所有这些因素。从根本上说，这正是他们将要

And so smart investors and smart lenders will consider all of these factors. Fundamentally, that's what they're going to

做的。

do.

那是他们的公司，与我无关。当然，我们必须与他们保持紧密联系，以确保我们的建设能支持他们的持续发展。明白吗？所以这是收入方面的事，与投资方面毫不相干。

That's their company. It's not my business. And of course, we have to stay very close to them to make sure that we build in support of their continued growth. Okay? And so there's the revenue side of it, and it has nothing to do with the investment side of it.

投资方面不受任何限制。这是投资他们的机会。正如我们之前提到的，这家公司很可能成为下一个价值数万亿美元的超大规模企业。谁不想成为这样的投资者呢？我唯一的遗憾是他们早期邀请我们投资时。

The investment side of it is not tied to anything. It's an opportunity to invest in them. And as we were mentioning earlier, this is likely going to be the next multi trillion dollar hyperscale company. And who doesn't want to be an investor in that? My only regret is that they invited us to invest early on.

我记得那些对话。

I remember those conversations.

当时我们太穷了。穷到没投够钱。我本该把全部身家都投给他们。

And we were so poor. We were so poor we didn't invest enough. And I should have given them all my money.

而现实是，如果你们不履行职责，跟不上节奏，如果维瓦里姆芯片表现不佳，他们完全可以换用其他芯片装进这些

And the reality is if you guys don't do your jobs and keep up with if vivarium doesn't turn into a good chip, they can go get other chips and put them in these

没错。

That's right.

他们并没有义务必须使用你们的芯片。

There's no obligation that they have to use your chips.

而且

And

正如你所说，你们将此视为机会主义的股权投资。我还想说

like you said, you're looking at this as opportunistic equity investment. The other thing I would say

而且我们已经做了一些很棒的投资。我必须提一下。我们投资了XAI。我们投资了CoreWeave。非常了不起。

And we've made some great investments. I've got to put it out there. We invested in XAI. We invested in CoreWeave. Incredible.

这决策多明智啊？

How smart was that?

回到这个话题，在我看来另一个根本问题是，你们公开表明了立场。你们在宣告这就是我们的行动。而背后的经济实质是什么？并非只是两家公司之间来回输送收入。我们有用户每月为ChatGPT付费，十五亿月活跃用户在使用这个产品。

As I go back to this, the other fundamental thing, it seems to me, is you're putting it out there. You're saying this is what we're doing. And the underlying economic substance here, right? It's not that you're just somehow sending revenues back and forth between the two companies. We've got people sending money every month for ChatGPT, a billion and a half monthly users using the product.

你刚刚说世界上每家企业要么这么做，要么就会消亡。每个主权国家都将此视为关乎国家安全和经济安全的生死存亡问题，如同核武器

You just said every enterprise in the world is either going to do this or they will die. Every sovereign views this as existential to their national security and economic security as nuclear

力量或西方世界。有哪个个人、公司或国家会说情报对我们来说基本是可选的？我的意思是，这对他们而言是根本。这就是情报的自动化。

power or West. What person, company, or nation says intelligence is basically optional for us? I mean, it's fundamental to them. It's the automation of intelligence.

我已经把需求问题讨论透了。现在让我们稍微深入系统设计，我稍后会请克拉克来谈这个。但在2024年，你们转向了年度发布周期对吧，从Hopper开始。然后在2025年进行了需要数据中心大规模改造的重大升级Grace Blackwell。2026年下半年我们将迎来Vera Rubin。

I beat the demand question to death. So let's jump in a little bit to system design, and I'm going to turn to Clark here in a second on that. But in 2024, you switched to your annual release cycle right, with Hopper. You then had a massive upgrade, which required significant data center overhaul with Grace Blackwell in 2025. And in the back half of 'twenty six, we're going to get Vera Rubin.

2027年会有Ultra，2028年是费曼。年度发布周期进展如何？主要目标是什么？

'twenty seven, we'll get Ultra. And 'twenty eight, Feynman. How is the annual release cycle going? Okay? What were the main goals of going to an annual release cycle?

英伟达内部的人工智能是否帮助你们实现了年度发布周期？

And did AI inside NVIDIA allow you to execute the annual release cycle?

是的，最后一个问题的答案是肯定的。没有它，英伟达的速度、节奏和规模都会受限。如今没有人工智能，根本不可能完成我们的建设。为什么要这么做？记得埃迪在财报电话会议或他的会议上说过。

Yeah, the answer is yes on the back of the last question. Without it, NVIDIA's velocity, our pace, our scale would be limited. And so without AI these days, it's just simply not possible to build what we build. Now, why do we do it? There's something that, remember, Eddie said it at his earnings call or his conference.

萨提亚说过，山姆也说过。代币生成率正在呈指数级增长。客户使用量也在指数级上升。我记得他们周活跃用户数达到了8亿左右。

Satya has said it. Sam has said it. The token generation rate is going up exponentially. And the customer use is going up exponentially. I think they were at 800,000,000 weekly active users or something like that.

是的。我是说，那距离ChatGPT发布已经超过两年了，对吧？

Yes. Mean, that's than two years from ChatGPT, right?

而且每位用户都在生成海量更多的token，因为他们正在使用推理时间进行推理。

And each of those users is generating massively more tokens because they're using inference time reasoning.

没错，完全正确。首要问题是，由于token生成速率正以惊人速度增长——两个指数级增长叠加在一起，除非我们以不可思议的速度提升性能，否则token生成成本将持续攀升，因为摩尔定律已经失效了，对吧？因为现在晶体管成本每年基本持平，功耗也大致不变。

That's right. Exactly. And so the first thing is because the token generation rate is going up so incredibly, two exponentials on top of each other, we have to unless we increase the performance at incredible rates, the cost of token generation will keep growing because Moore's law is dead, right? Because transistors basically cost the same every single year now. And power is largely the same.

在这两条基本定律制约下，除非我们发明新技术来降低成本，即使毛利率有微小差异——比如给予几个百分点的折扣，又怎能抵消双重指数增长？因此我们必须以匹配这种指数增长的速度逐年提升性能。以英伟达的AI发展历程为例，从Kepler架构一路发展到Hopper架构，性能可能提升了10万倍。这是英伟达AI征程的起点——十年间实现10万倍提升。

And between those two fundamental laws, unless we come up with new technologies to drive the cost down, even if there's a slight difference in gross, you give somebody a discount of a few percent, how is that going to make up for two exponentials? And so we have to increase our performance annually at a pace that keeps up with that exponential. So in the case of going from, I guess, Kepler to all the way to Kepler all the way to Hopper was probably 100000x. That was the beginning of the AI journey for NVIDIA. 100000x in ten years.

明白吗？从Hopper到Blackwell架构，得益于NVLink72技术，我们一年内实现了30倍提升。接着Rubin架构会带来又一个数量级提升，Feynman架构会再次实现飞跃。我们能做到这点是因为晶体管本身已帮不上什么忙了，对吧？

Okay? Between Hopper and Blackwell, we increased, because of NVLink72, 30x in one year. And then we'll get another x factor again with Rubin. And then we'll get another X factor with Feynman. And the way we do that is because the transistors aren't really helping us very much, right?

摩尔定律基本上...虽然晶体管密度在增加，但性能并未同步提升。因此我们面临的挑战是：必须在系统层面彻底重构整个问题，同步革新每颗芯片、整个软件栈以及所有系统——这是极致的协同设计。此前从未有人实现过这种级别的协同设计。

Moore's law is largely. The density is going up, but the performance is not. And so if that's the case, one of the challenges that we have to do is we have to break the entire problem down at the system level and change every chip at the same time and all the software stack and all the systems all at the same time. The ultimate extreme co design. Nobody's ever co designed at this level before.

我们改造了CPU，革命性地升级了GPU，重新设计了网络芯片，通过NVLink实现纵向扩展，通过Spectrum X实现横向扩展。有人说——我听到有人说'不就是以太网嘛'。好吧，Spectrum X以太网绝非普通以太网。人们开始发现：天啊，这些X因子简直不可思议。

We change the CPU, revolutionize the CPU, a GPU, the networking chip, the NVLink scale up, the Spectrum X scale out. Somebody said, I heard somebody said, Oh yeah, it's just ethernet. Yeah, right. Okay, so Spectrum X ethernet is not just ethernet. And people are starting to discover, oh my god, the X Factors is pretty incredible.

英伟达的以太网业务，即纯以太网业务，是全球增长最快的以太网业务。因此我们需要横向扩展。当然，现在我们必须构建更庞大的系统，以便将多个相互连接的AI工厂进行跨系统扩展。我们每年都在推进这项工作。因此，从技术层面而言，我们正经历着指数级叠加指数级的增长。

NVIDIA's ethernet business, the Just Ethernet business, is the fastest growing ethernet business in the world. And so scale out. And of course, now we have to build even larger systems so we scale across multiple AI factories connected together. And then we do this at an annual pace. And so we now have an exponential of exponentials going ourselves from technology.

这使得我们的客户能够降低token成本，通过预训练、后训练及持续优化让这些token变得越来越智能。最终当AI变得更聪明时，它们的使用率就会提升；而使用率越高，其增长就会呈现指数级爆发。

And that allows our customers to drive the cost of tokens down, keep making those tokens smarter and smarter with pre training and post training and thinking. And as a result, when the AI gets smarter, they get more used. When they get more used, they're going to grow exponentially.

对于不太了解的听众，什么是极致协同设计？

For people who may not be as familiar, what is extreme co design?

极致协同设计意味着必须同步优化模型、算法、系统及芯片。你需要跳出框架进行创新——因为摩尔定律认为只需让CPU越来越快，所有东西都会加速，这其实是在既定框架内创新。

Extreme co design means that you have to optimize the model, algorithm, system, and chip at the same time. You have to innovate outside the box. Because Moore's law said, you just have to keep making the CPU faster and faster. Everything got faster. You were innovating within a box.

框架。只要让芯片更快就行。但如果芯片速度无法再提升呢？那就必须突破框架创新。英伟达真正改变了游戏规则，因为我们做了两件事：

Box. Just make that chip faster. Well, if that chip doesn't go any faster, then what are you going to do? Innovate outside the box. And so NVIDIA really changed things because we did two things.

我们发明了CUDA和GPU，并开创了大规模协同设计的理念。这就是为什么我们能涉足众多行业——我们创建了各种库并实施协同设计。首先是全栈能力。'极致'甚至超越了软件和GPU层面，现已延伸至数据中心级别，包括交换机、网络设备、其中所有软件、网卡，纵向扩展与横向扩展，实现全局优化。

We invented CUDA, invented GPUs, and we invented the idea of co design at a very large scale. That's why there's all these industries we're in. We're creating all these libraries and co design. Number one, full stack. Extreme is even beyond software and GPUs, it's now at the data center level, switches and networking and all of that software in the switches and the networking and the NICs, scale up, the scale out, optimizing across all of that.

正因如此，Blackwell相比Hopper实现了30倍提升。任何摩尔定律都不可能达成这种飞跃。这就是'极致'的含义。

As a result of that, Blackwell to Hopper is 30x. No Moore's Law could possibly achieve that. And so that's extreme.

而这正是源于极致的协同设计。

And that comes from the extreme co design.

这正是英伟达涉足网络、交换机、纵向扩展、横向扩展、跨平台扩展，以及研发CPU、GPU和网卡的原因。我们拥有如此丰富的软件资源——全球开源代码贡献量我们仅次于另一家公司（可能是AI领域的）。这还只是AI领域，别忘了我们在计算机图形学、数字生物学和自动驾驶领域的软件积累。

That's because NVIDIA has That's why we got into networking and switching and scale up and scale out and scale across and building CPUs and building GPUs and building NICs. That's the reason why NVIDIA is so rich in software and people We check-in more open source software in the world than just about anybody except one other company. I think it's AI too or something like that. And so we have such enormous richness of software and that's just in AI. Don't forget computer graphics and digital biology and autonomous vehicles.

我们公司开发的软件数量惊人，这使我们能够进行深度极致的协同设计。

Amount of software we produce as a company is incredible. That allows us to do deep and extreme co design.

听你们某位竞争对手说，这样做确实能降低token生成成本，但你们每年发布新品的节奏让对手几乎无法跟进。供应链也被你们锁定得更紧——因为给供应链提供三年可见性后，他们就能明确知道该生产什么了。

I heard from one of your competitors, yes, he's doing this because it helps drive down the cost of token generation, but at the same time, your annual release cycle makes it almost impossible for your competitors to keep up. The supply chain gets locked up more because you're giving three year visibility to your supply chain, so now the supply chain has confidence as to what they can build to.

你想想——提问前先思考下：要实现每年数千亿美元的AI基础设施部署，我们一年前就要开始筹备多少产能？这涉及数百亿美元的晶圆投片和DRAM采购。

Do you Think about this. Wait before you ask the question. Think about this. In order for us to do several $100,000,000,000 a year of AI infrastructure build out, think about how much capacity we had to go start a year ago. We're talking about building hundreds of billions of dollars of wafer starts and DRAM buys.

各位明白吗？这种规模几乎没有企业能维持运转。

Are you guys talking? This is now at a scale that hardly any company can keep open.

所以你认为现在的竞争壁垒比三年前更高了吗？

So would you say your competitive moat is greater today than it was three years ago?

是的。首先，现在的竞争比以往任何时候都要激烈。但难度也前所未有。我之所以这么说，是因为晶圆成本越来越高，这意味着除非你在极大规模上进行协同设计，否则根本无法实现指数级增长，这是第一点。因此，除非你每年研发六、七、八颗芯片，那才称得上非凡成就。

Yeah. First of all, there's just more competition than ever before. But it's harder than ever before. And the reason why I say that is because wafer costs are getting higher, which means that unless you do co design at an extreme scale, you're just not going to be able to deliver the X factor growth, number one. And so unless you're working on six, seven, eight chips a year, that's an amazing thing.

关键不在于制造一块ASIC芯片，而是构建一个AI工厂系统。这个系统包含大量芯片，它们全部采用协同设计，共同实现我们几乎定期获得的10倍性能提升。

It's not about building an ASIC. It's about building an AI factory system. And this system has a lot of chips in it. And they're all co designed. And together, they deliver that 10x factor that we get almost regularly.

明白吗？所以第一，协同设计达到极致。第二是规模达到极致。当客户部署千兆瓦级设施时，意味着40万到50万块GPU。让50万块GPU协同工作本身就是奇迹。

Okay? So number one, the co design is extreme. The second thing is that the scale is extreme. When your customers deploy a gigawatt, that's 400,000, 500,000 GPUs. Getting 500,000 GPUs to work together is a miracle.

这简直是个奇迹。客户承担巨大风险采购所有这些设备时，你必须思考：什么样的客户会为一种架构下50亿美元的采购订单？

I mean, it's just a miracle. And so your customers are taking enormous risk on you to go buy all of this. You've got to ask yourself, what customer would place a $50,000,000,000 PO on an architecture?

而且还是未经验证的架构。

On an unproven architecture.

没错，全新的架构。

That's right. A new one.

对，全新架构。

Right, a new architecture.

没错，你刚拿出一款全新的芯片。你对它充满热情，大家也为你兴奋不已。当你首次展示硅片样品时，谁会立刻给你500亿美元的采购订单？

Yeah, you just take out a whole new chip. You're as excited as you are about it. And everybody's excited for you. And you just show the first silicon. Who's going to give you $50,000,000,000 PO?

为什么你会为一个刚完成流片的芯片启动价值500亿美元的晶圆生产？但对英伟达来说，我们可以做到，因为我们的架构久经考验。我们的客户规模庞大到难以置信，如今供应链规模也同样惊人。除非确信英伟达能兑现承诺，否则谁会为一个公司预先启动所有生产环节？

And why would you start $50,000,000,000 worth of wafers for a chip that just taped out? But for NVIDIA, we could do that because our architecture is so proven. So the scale of our customer is so incredible. Now the scale of our supply chain is incredible. Who's going to start all of that stuff, pre build all of that stuff for a company unless they know that NVIDIA can deliver through.

难道不是这样吗？他们相信我们能服务全球所有客户，所以愿意一次性启动数百亿美元的生产规模。这种量级确实令人震撼。

Isn't that right? And they believe that we can deliver through to all of the customers around the world. They're willing to start several $100,000,000,000 at a time. This the scale is incredible.

说到这个，目前全球最关键的争论焦点之一就是GPU与ASIC之争。谷歌的TPU、亚马逊的Tranium，从ARM到OpenAI再到Anthropic，似乎所有公司都在传闻要自研芯片。

To that point, you know, one of the biggest key debates and controversies in the world is this question of GPUs versus ASICs. Google's TPUs, Amazon's Tranium, and it seems like everyone from ARM to OpenAI to Anthropic are, you know, rumored to be building one.

嗯。

Mhmm.

去年你说过，英伟达构建的是系统而非单一芯片，你们通过整个技术栈的每个环节来提升性能。你还提到这些项目中许多可能永远达不到量产规模。但考虑到，比如大多数...

Last year, you said, you know, we're building systems, not chips, and you're driving performance through every single part of that stack. You also said that many of these projects may never get to production scale. But given, like, the Most of

它们中的大多数。大多数项目。

them. Most of them.

是的。考虑到谷歌TPU看似取得的成功

Yeah. Given the seeming success of Google's TPUs

没错。

Yeah.

你如何看待当前这个不断演变的格局？

How are you thinking about this evolving landscape today?

首先，谷歌的优势在于远见。记住，他们在一切开始前就启动了TPU项目。这与初创企业没有区别。你应该在市场规模壮大之前创立一家初创公司。

First of all, the advantage that Google had is foresight. Remember, they started TPU one before everything started. This is no different than a startup. You're supposed to build a startup. You're supposed to create a startup before the market grows.

而不是等到市场规模达到万亿美元时才成立初创公司。这个谬误所有风投都明白——认为只要在庞大市场中占据几个百分点份额就能成为巨头企业。这本质上是错误的。你应该像英伟达那样，在微型行业里占据100%份额，TPU走的正是这条路。

You're not supposed to come up as a startup when the market's $1,000,000,000,000 large. This fallacy and all VCs know this this fallacy that a large market, if you could just take a few percent market share, you could be a giant company. That's actually fundamentally wrong. You're supposed to take 100% of a tiny company, a tiny industry, which is what NVIDIA did, right? Which is what TPUs did.

当时只有我们两家。

There were only the two of us.

但你最好希望那个行业能真正壮大起来。

But you better hope that that industry gets really big.

工业。没错。

Industry. That's right.

我是说，英伟达的故事。所以这就是

And I mean, the NVIDIA story And so that's the

当前正在构建ASIC芯片的人们面临的挑战。这看起来是个诱人的市场。但请记住，这个诱人的市场已从一种名为GPU的芯片，发展成了我刚才描述的AI工厂。你们刚才也看到了，我刚刚宣布了一款名为CPX的芯片，专用于上下文处理和扩散视频生成。这是一个非常专业化的负载，但在数据中心内部却至关重要。

challenge for people who are building ASICs now. It looks like a juicy market. But remember, this juicy market has evolved from a chip called a GPU to, I just described, an AI factory. And you guys just saw, I just announced a chip called CPX for context processing and diffusion video generation. A very specialized workload, but an important workload inside a data center.

我刚才的铺垫可能是关于AI数据处理处理器。因为你知道吗？你需要长期记忆，也需要短期记忆。键值缓存处理的强度非常大。

I just prelude it to maybe AI data processing processors. Because guess what? You need long term memory. You need short term memory. The KV cache processing is really intense.

AI内存至关重要。你肯定希望你的AI拥有良好的记忆能力。而仅仅处理系统中所有的键值缓存就是极其复杂的事情，或许它需要一个专门的处理器。可能还有其他需求，对吧？所以你会发现，英伟达现在的视角已不再是GPU。

AI memory is a big deal. You kind of like your AI to have good memory. And just dealing with all the KV caching around the system, really complicated stuff, maybe it wants to have a specialized processor. Maybe there's other things, right? So you see that NVIDIA's our viewpoint is now not GPU.

我们的视角是审视整个AI基础设施，以及这些卓越的公司需要什么才能让所有多样化且不断变化的负载通过它。看看Transformer架构，它正以惊人的速度演变。如果不是因为CUDA易于操作和迭代，他们如何尝试海量实验来决定使用哪个Transformer版本、哪种注意力算法？如何实现解耦？

Our viewpoint is looking at the entire AI infrastructure and what does it take for these incredible companies to get all of their workload through it, which is diverse and changing. Look at the transformer. The transformer architecture is changing incredibly. If not for the fact that CUDA is easy to operate on and iterate on, how do they try all of their vast number of experiments to decide which one of the transformer versions, what kind of attention algorithm to use. How do you disaggregate?

CUDA能帮助你完成所有这些，因为它的可编程性极强。所以现在看待我们业务的方式是：当所有这些ASIC公司或ASIC项目在三、四、五年前启动时，我得说，那个行业超级可爱又简单。当时只有GPU参与其中。但现在它变得庞大而复杂。再过两年，它将变得完全庞大无比。

CUDA helps you do all that because it's so programmable. And so the way to think about our business now is you look at when all of these ASIC companies or ASIC projects start three, four, five years ago, I got to tell you, that industry was super adorable and simple. There was a GPU involved. But now it's giant and complex. And in another two years, it's going to be completely massive.

规模将会非常庞大。所以我认为，作为一个新兴参与者进入一个巨大市场的竞争本身就很难，你们懂的。

The scale is going to be so large. And so I think that the battle of getting into a very large market as a nascent player is just hard, you know, as you guys know.

即便是对那些可能在使用ASIC上取得成功的客户来说

Even for the customers who perhaps are successful with ASICs

是啊。

Yeah.

他们的计算资源组合难道不存在最优平衡吗？我是说，投资者往往是非黑即白的生物，他们只想要一个明确的是或否的答案。

Isn't there an optimal balance in their compute fleet? Like, it's you know, I think investors are very much binary creatures. They just want a a yes or no black and white answer.

没错。

Yeah.

但即便ASIC能正常工作，考虑到我在采购英伟达平台——CPX即将推出用于预填充，视频生成可能需要解码器，或者视频转码平台之类的，难道不存在最优平衡吗？确实如此。

But even even if you get the ASIC to work, isn't there an optimal balance because you think, I'm buying the NVIDIA platform. CPX is gonna come out for prefill Yeah. For, you know, for video generation, maybe a decode, you know, a you know, a platform A video transcoder. Exactly. Yeah.

没错。随着新工作负载的出现，英伟达生态系统里会需要许多不同的芯片或部件来扩充加速计算资源。而那些现在正在流片新芯片的人，其实无法准确预测一年后的技术需求。

Yeah. So there will be, like, many different chips or parts to add to the NVIDIA ecosystem, accelerated compute fleet Yeah. Right, as new workloads are, you know, are are born. That's right. And, you know, people trying to tape out new chips today are not really anticipating what's happening a year from now.

他们只是想让一块芯片正常工作。

They're just trying to get a chip to work.

没错。

That's right.

换句话说，谷歌是个大GPU客户。

Say it another way, Google's a big GPU customer.

谷歌是个大GPU客户。如果你观察的话，谷歌是个非常特殊的案例。我的意思是，该尊重的地方必须尊重——我真心认为，TPU已经发展到第七代了。对吧？

Google's a big GPU customer. If you look at and Google is a very special case. I mean, just have to show respect where respect is I really mean, TPU is on TPU seven. Yes. Right?

这对他们也是个挑战，不是吗？他们所做的工作极其艰难。所以我认为首先要明确的是，记住芯片分三类：第一类是架构型芯片，比如x86 CPU、ARM CPU、NVIDIA GPU这类具有架构特性的。

And it's a challenge for them as well, right? And so the work that they do is incredibly hard. So I think the first thing to let me do it. Remember, there are three categories of chips. There's the category of chips that are architectural, x86 CPUs, ARM CPUs, NVIDIA GPUs, architectural.

这类芯片上方有完整的生态系统。架构本身拥有丰富的知识产权和生态系统，技术极其复杂。是由像我们这样的所有者构建的。明白吗？然后是ASIC芯片。

And it has an ecosystem above. And the architecture allows has rich IP and rich ecosystem, very complicated technology. It's built by the owners like us. Okay? There's ASICs.

我曾就职于发明ASIC概念的LSI Logic原公司。如你所知，LSI Logic已不复存在。原因在于当市场规模不大时，ASIC确实很出色——很容易找到承包商帮你完成封装制造，他们收取50到60个百分点的利润。

I worked for the original company, LSI Logic, who invented the idea of ASICs. As you know, LSI Logic is not here anymore. And the reason for that is because ASICs is really fantastic when the market size is not very large. It's easy to have somebody be a contractor to help you put the packaging of all that stuff together and do the manufacturing on your behalf. And they charge you $50.60 points of margin.

但当ASIC市场规模扩大时，出现了一种称为COT（客户自有工具）的新模式。谁会采用这种方式？以苹果智能手机芯片为例，其产量如此巨大，他们绝不会支付50%、60%的毛利率让别人代工ASIC。他们采用客户自有工具模式。那么当TPU业务规模扩大后，将何去何从？

But when the market gets large for an ASIC, there's a new way of doing things called COT, customer owned tooling. And who would do something like that? Apple's smartphone chip, the volume is so large, they would never go pay somebody else 50%, 60 gross margin to be an ASIC. They do customer owned tooling. And so where will TPUs go when it becomes a large business?

毫无疑问会转向客户自有工具。但ASIC仍有其生存空间。视频转码器的市场规模永远不会太大，智能网卡也是如此。

Customer owned tooling. There's no question about it. But there's a place for ASICs. Video transcoders will never be too large. Smart NICs will never be too large.

因此当一家ASIC公司同时进行十到十五个ASIC项目时，我并不感到意外——可能包含五个智能网卡和四个转码器。这些全是AI芯片吗？当然不是。如果有人要为特定推荐系统打造嵌入式处理器，且采用ASIC方案，这当然可行。但你会将其作为持续演变的AI基础计算引擎吗？

And so when there's ten, twelve, 15 ASIC projects going on at an ASIC company, I'm not surprised by that because there are probably five Smart NICs and four transcoders. Are they all AI chips? Of course not. And if somebody were to build an embedding processor for a specific recommender system and that was an ASIC, of course you could do that. But would you do that as the fundamental compute engine for AI that's changing all the time?

存在低延迟工作负载，也有高吞吐量工作负载。既有聊天对话的token生成，也有思维型工作负载，还有AI视频生成工作负载。

You've got low latency workload. You've got high throughput workload. You have token generation for chat. You have thinking workload. You have AI video generation workload.

你现在讨论的是

Is there a now you're talking about

这正是你们加速计算架构的中流砥柱

a That's the very workhorse backbone of your accelerated

这正是NVIDIA的核心价值所在。

That's what NVIDIA is all about. Again,

简单来说，这就像下国际象棋和跳棋的区别，对吧？事实上，现在开始做ASIC的人，无论是Tranium还是其他各种加速器，他们都在打造一个更大机器的组件。你们已经构建了一个非常复杂的系统、平台或工厂，随你怎么称呼。现在你们稍微开放了一些，对吗？

dumb this down. It's like playing chess and checkers, right? The fact of the matter is the folks who are starting ASICs today, whether it's Tranium or whether it's some of these other different accelerators, etc, they're building a chip that's a component of a much larger machine. You've built a very sophisticated system, platform, factory, whatever you want to call it. And now you're opening up a little bit, right?

你刚才提到了CPX GPU对吧？在我看来，你们某种程度上是在将工作负载分解到最适合特定需求的硬件模块上。

So you mentioned CPX GPU, right? It seems to me that in some ways you're disaggregating the workloads to the best slice of the hardware for that particular demand.

看看我们的成果。我们发布了名为Dynamo的分布式AI工作负载协调系统，并开源了它，因为未来的AI工厂必定是分布式架构。

Look what we did. We announced this thing called Dynamo, disaggregated AI workload orchestration, and we open sourced it because the future AI factory is disaggregated.

你们还推出了NV Fusion，甚至对包括刚投资的英特尔在内的竞争对手说：'参与我们建造工厂的方式是这样的'——毕竟没人会疯狂到想独自建造整个工厂。但只要你的产品足够优秀，能让终端用户说'我们想用这个替代ARM GPU'或'替代你们的推理加速器'，就能接入这个体系。是这样吗？

And you launched NV Fusion that even said to your competitors, including Intel, which you just invested in, the way in which you participate in this factory that we're building, because nobody else is crazy enough to try to build the entire factory. But you can plug into that if you have a product that's good enough, compelling enough that the end user says, Hey, we want to use this instead of an ARM GPU, or We want to use this instead of your inference accelerator, etcetera. Is that correct?

是的，我们非常乐意将你们接入系统。

Yeah, we're delighted to connect you in.

好的，请再详细说说。

Yeah. Tell us a little bit more.

D Link Fusion这个构想太棒了。我们很高兴能与英特尔合作，这整合了英特尔生态——毕竟全球大多数企业仍运行在英特尔平台上。

D Link Fusion. Such a great idea. And we're so happy to partner with Intel on that. It takes the Intel ecosystem. Most of the world's enterprise still runs on Intel.

这需要英特尔生态系统，需要英伟达的人工智能生态系统，加速计算，我们将它们融合在一起。我们与ARM合作实现了这一点。未来还将与其他几家伙伴达成类似合作。这为我们双方都打开了机遇。这是双赢的局面。

It takes the Intel ecosystem, takes the NVIDIA AI ecosystem, accelerated computing, and we fused it together. And we did that with ARM. And there are several others we're going be doing it with. And that opens up opportunities for both of us. It's a win for both of us.

非常、非常重大的胜利。我将成为他们的大客户，而他们将为我们打开一个更广阔的市场机遇。是的。

Great, great win. I'll be a large customer of theirs, and they're going to expose us to a much, much larger market opportunity. Yeah.

与此密切相关的是您提出的一个令人震惊的观点：我们的竞争对手正在研发ASIC芯片。他们现在的芯片已经更便宜，甚至可以完全免费提供。但我们的目标是——即便他们免费赠送芯片，客户仍会选择英伟达系统，因为包括电力、数据中心、土地等在内的整体运营成本，其产出的人工智能效益，依然优于免费获取的芯片。

Deeply related to this idea is the argument you've made that kind of shocks some people, where you say, Our competitor's building ASICs. They could literally All their chips are cheaper already today, but they could literally price them at zero. Our objective is they could price them at zero, and you would still buy an NVIDIA system because the total cost of operating that system power, data center, land, etcetera the intelligence out is still a better bet than buying a chip even if it's given to you for free.

因为土地、电力和基础设施成本已经高达150亿美元。没错。

Because the land power and shell is already $15,000,000,000. Right.

是的。我们尝试过核算这笔账，但请您详细解释您的计算逻辑。因为对那些不常接触这方面的人来说，这完全不合常理——竞争对手芯片免费提供，而你们芯片价格昂贵，怎么可能仍是更优选择？

Yeah. So we've taken a crack at kind of the math on that, but walk us through your math. Because I think for people who don't spend as much time here, that they it just doesn't compute. How could it possibly be that you were pricing your competitors chips at zero, given the expense of your chips and it still is a better bet?

可以从两个角度思考。其一是从收入维度来分析。

There's two ways to think about it. One way is let's just think about it from a perspective of revenues.

所以

So

每个人的电力资源都是有限的。假设你能够额外获取两千兆瓦的电力，那么你自然希望这两千兆瓦能转化为收益。没错。因此，你的性能或者说每瓦特产生的代币量会是其他人的两倍，因为我进行了深度极致的协同设计。

everybody's power limited. And let's say you were able to secure two more gigawatts of power. Well, that two gigawatts of power you would like to have translate to revenues. Yes. So your performance or tokens per watt was twice as high as somebody else's token per watt because I did deep and extreme co design.

而我的单位能耗性能远高于他人，那么我的客户就能从他们的数据中心获得双倍收益。谁不想要双倍收益呢？即便有人给他们15%的折扣，我们的毛利率差距大约是75个百分点，而其他人的毛利率在50到65个百分点之间，这还不足以弥补Blackwall与Hopper之间30倍的差距。假设Hopper是一款卓越的芯片和系统，再假设别人的ASIC就是Hopper。

And my performance was much higher per unit energy, then my customer can produce twice as much revenues from their data center. Who doesn't want twice as much revenues? And if somebody gave them a 15% discount, the difference between our gross margins was, call it, 75 points and somebody else's gross margins, it 50 to 65 points, is not so much as to make up for the 30 times difference between Blackwall and Hopper. Let's pretend Hopper is an amazing chip, an amazing system. Let's pretend somebody else's ASIC is Hopper.

Blackwall是30倍的差距。所以，你在一千兆瓦的电力上要放弃30倍的收益。这个代价太大了。即使他们免费提供给你，你只有两千兆瓦可用，机会成本高得离谱，你总会选择每瓦特性能最优的方案。

Blackwall is 30 times. So you've got to give up 30x revenues in that one gigawatt. It's too much to give up. So even if they gave it to you for free, you only have two gigawatts to work with. Your opportunity cost is so insanely high, you would always choose the best perf per watt.

我从某超大规模公司的CFO那里听到，鉴于你们芯片带来的性能提升——再次强调，正是每千兆瓦代币数和电力成为限制因素——他们不得不升级到新周期。那么展望Rubin、Rubin Ultra和Feynman时，这种提升趋势会持续吗？

So I heard this from one of the CFOs at one of the hyperscalers that given the performance improvement that's coming out of your chips, again, precisely to that point tokens per gig and power being the limiting factor that they had to upgrade to the new cycle. So when you look ahead at Rubin, at Rubin Ultra, at Feynman, does that trajectory continue?

我们现在每年要造多少芯片？六七个吧？

We're building what, six, seven chips a year now?

每一个都是那个系统的一部分。

Each As one part of that system.

没错。而且那个系统中软件无处不在，需要整合优化这六七个芯片才能实现30倍性能的Blackwell。想象一下我每年都在这样做：砰、砰、砰、砰、砰、砰。如果你在这一堆ASIC、一堆芯片中只造一个ASIC，而我们正在整体优化，这确实是个难题。

That's right. And that system, software is everywhere, and it takes the integration and the optimization across all of those six, seven chips to deliver on the 30X Blackwell. Now imagine I'm doing this every single year. Bam, bam, bam, bam, bam, bam. And so if you build one ASIC in that soup of ASICs, in that soup of chips, and we're optimizing across that, it's a hard problem to solve.

这确实让我回到了我们最初讨论的竞争护城河问题。我们与投资者探讨这个话题已有时日。作为贯穿整个生态系统的投资者，从谷歌到博通都是你们的竞争对手。当我回归第一性原理思考时，不禁要问：你们是在拓宽还是缩窄竞争护城河？你们转向了年度发布节奏。

This does bring me back to where we started about the competitive moat. We've been covering this in investors for a while. We're investors throughout the ecosystem and competitors of yours from Google to Broadcom. When I really just first principles around this and say, are you increasing or decreasing your competitive moat? You move to an annual cadence.

你们与供应链共同开发。规模远超所有人预期，这需要资产负债表和研发能力的同步扩张。通过收购或内生增长（比如刚讨论过的NV Fusion和CPX），这些举措让我确信你们的竞争护城河正在相对拓宽——

You're codeveloping with supply chain. The scale is massively bigger than anybody anticipated, which requires scale, both of balance sheet and of development. Right? The moves you made both through acquisition and organically with things like NV Fusion or CPX, which we just talked about. All of those things together cause me to believe that your competitive moat is increasing vis a vis,

在

at

至少在建厂或系统扩展层面如此。

least insofar as building out the factory or system.

至少令人惊讶。但

It's at least surprising. But

有趣的是你们的市盈率远低于多数同行。我认为部分原因在于规模效应的限制——4.5万亿美元市值的公司怎么可能再增长？但一年半前我问过：面对AI工作负载将增长5-10倍的市场预期，已知资本支出的走势，在你们构想中，是否存在五年后营收规模达不到2025年2-3倍的可能性？考虑到这些优势，营收停滞的概率究竟有多大？

I think it's interesting that your multiple is much lower than most of those other people. And I think part of that has to do with this law of large numbers. A $4,500,000,000,000 company couldn't possibly get any bigger. But I asked you this a year and a half ago: As you sit here today, the market's going to AI workloads are going to 10x or 5x, we know what CapEx is doing, etcetera, Is there any conceivable world in your mind where your top line in five years isn't 2x or 3x bigger than it is in 2025? Like, what's the probability that it's actually not much higher than it is today, given those advantages?

我这么回答：正如我所描述的，我们的机遇远超出市场共识。

I'll answer it this way. Our opportunity, as I described it, is much larger than the consensus.

我要在这里说，我认为英伟达很可能成为第一家市值达到10万亿美元的公司。我在这行够久了，就在十年前，人们还说永远不可能有万亿美元市值的公司，现在我们已经有了10家，对吧？

I'll say it here. I think NVIDIA will likely be the first $10,000,000,000,000 company. And I've been here long enough it wasn't that long ago, just a decade ago, as you well remember that people said there could never be a trillion dollar company. Now we have 10. Right?

如今世界更大了。没错。现在又回到了围绕GDP和增长的指数级发展轨道上。

Today But the world's bigger. Right. And today, is back to the exponentials around GDP and the growth.

世界确实更大了。但人们误解了我们的业务。他们记得我们是芯片公司，我们制造芯片。天啊，我们确实制造芯片，而且是世界上最了不起的芯片。

The world is bigger. And people misunderstand what we do. They remember we're a chip company. And we build chips. Boy, do we build chips and build the most amazing chips in the world.

但英伟达实际上是一家AI基础设施公司。我们是你们的AI基础设施合作伙伴。我们与OpenAI的合作就是完美例证——我们是他们的AI基础设施伙伴。我们以多种方式与客户合作，不要求任何人必须全盘采购我们的产品。

But NVIDIA is really an AI infrastructure company. We are your AI infrastructure partner. And our partnership with OpenAI is a perfect demonstration of that, that we are their AI infrastructure partner. And we work with people in a lot of different ways. We don't require anybody to buy everything from us.

我们不要求客户购买整机柜。他们可以只买芯片，可以买某个组件，可以采购我们的网络设备，甚至有些客户只购买我们的CPU。

We don't require that they buy the full rack. They could buy a chip. They could buy a component. They could buy our networking. They could buy our we have customers buying only our CPU.

只买我们的GPU，搭配别人的CPU和别人的网络设备也行。我们接受任何采购方式。我唯一的请求就是——

Just buy our GPUs and buy somebody else's CPUs and somebody else's networking. We're kind of Okay selling any way you like to buy. My only request is just buy a

好歹买点我们的产品。你说这不只是关于更好的模型，我们还需要建设者。我们需要世界级的建设者。你说过，也许埃隆·马斯克就是我们国家最顶尖的建设者代表。

little something from us. You said this isn't just about better models. We also have to build. We have to have world class builders. And you said, you know, the most world class builder maybe that we have in the country is Elon Musk.

我们之前讨论过Colossus I及其功能，它当时能协调部署数十万——你知道的，就是那些H-100、H-200组成的集群。现在他正在研发Colossus II，可能达到50万GB的规模，相当于数百万个H100组成的协调集群。

And we talked about Colossus I and what he was doing there, standing up a couple 100,000, you know, the time, H-100s, H-200s in a coherent cluster. Now he's working on Colossus II, which may be 500,000 GBs, millions of H100 equivalents in a coherent cluster.

如果他率先实现千兆瓦级规模，我一点都不会感到惊讶。就在一个...

I would not be surprised if he gets to a gigawatt before anybody else does Right. In one

没错。稍微展开说说这点优势——作为建设者，不仅开发软件和模型，还深刻理解构建这些集群所需的一切。

Yeah. So say a little bit about that. The advantage of being the builder who isn't just building the software and the models but understands what it takes to build those clusters.

这些AI超级计算机确实复杂。技术本身复杂，由于融资问题采购过程复杂，获取土地电力及基础设施供电复杂，整个建造调试过程都复杂。

Well, these AI supercomputers are complicated things. The technology is complicated. Procuring it is complicated because of financing issues. Securing the land power and shell, powering it is complicated. Building it all, bringing it all up.

毫无疑问，这是人类迄今为止尝试过的最复杂系统难题。而埃隆的巨大优势在于，所有这些系统如何协同运作的蓝图都存在于他一个人的脑海中——包括融资在内的所有依存关系。

I mean, this is unquestionably, the most complex systems problem humanity has ever endeavored. And so Elon has a great advantage that, in his head, all of these systems are interoperating. And the interdependencies resides in one head, including the financing.

是的。他本身就是个大型GPT，是台人形超级计算机。

Yes. He's a big GPT. He's a big supercomputer himself.

他是终极GPU。因此他在这方面具有巨大优势，同时还具备强烈的紧迫感，怀抱着真正的建造欲望。当意志力与技能相结合，就能创造奇迹。

He's the ultimate GPU. And so he has a great advantage there. And he has a great sense of urgency. He has a real desire to build it. And so when will comes together with skill, unbelievable things can happen.

是啊，相当独特。

Yeah, quite unique.

我想和你深入探讨的是主权AI这个话题。我想聊聊中国以及正在进行的全球AI竞赛。回顾三十年前的你，恐怕无法想象这周你会与酋长、国王们在宫殿里会面，还频繁出入白宫。总统曾表示你和英伟达对美国国家安全至关重要。面对这种情况，首先请你帮我理清背景——若非各国将AI视为关乎存亡的要务，就像上世纪40年代对待核技术那样，你出现在这些场合简直难以置信。

Something you've been so involved in is I want to talk about sovereign AI. I want to talk about China and the global AI race that's going on. When I look back at you thirty years ago, you couldn't have imagined you were going to be hanging out in palaces with emirs and the king this week, and you're at the White House all the time. The president has said that you and NVIDIA are critical to US national security. So when you look at that, first just contextualize for me, like, it's hard to believe that you would be in those places if sovereigns didn't view this at least as existential, as important, as maybe we did nuclear in the 1940s.

对吧？如今我们至少没有政府资助的曼哈顿计划，但有英伟达、OpenAI、Meta和谷歌在投入。如今这些企业规模堪比城邦国家，感谢美国造就了这样的环境——这些公司正在资助的项目，在各国总统和国王眼中似乎关乎他们未来的经济与国家安全。你同意这个观点吗？

Right? We don't have a Manhattan Project today at least funded by the government, but it's funded by NVIDIA, it's funded by OpenAI, it's funded by Meta, it's funded by Google. We have companies today the size of nation states, and thank God for America, right, who are funding something that it appears to me presidents and kings think is existential to their future economic and national security. Would you agree with that?

没人需要原子弹。但人人都需要AI。说得好。确实如此。没错。

Nobody needs atomic bombs. Everybody needs AI. Well said. Here, here. Yeah.

确实如此。这就是天壤之别。如你所知，AI就是现代软件——这正是我的起点。从通用计算到加速计算，从逐行人工编码到AI自动编程，这个根基不容忽视。

Here, here. And so that's a very, very large difference. AI, as you know, is modern software. That's where I started. From general purpose computing to accelerated computing, from human written code line at a time to AI written code, that foundation can't be forgotten.

我们重塑了计算技术。地球上没有诞生新物种，我们只是革新了计算方式——而计算是所有人的刚需。必须普及这种技术，正因如此所有国家都意识到必须进军AI领域，因为谁都不能脱离计算时代。世界上没人会说：猜怎么着？

We've reinvented computing. There's not a new species on earth. We just reinvented computing and everybody needs computing. It needs to be democratized, which is the reason why all of these countries realize they have to get into the AI world because everybody needs to stay in computing. There's nobody in the world that says, guess what?

我昨天还在用计算机，但明天改用棍棒火把也挺好。所以每个人都必须拥抱计算技术，这只是现代化的必然进程。仅此而已。

I used to use computers yesterday. I'm pretty good with Clubs and Fire tomorrow. And so everybody needs to move into computing. It's just being modernized. That's all.

明白吗？第一点。为了参与人工智能，你必须将你的历史、文化和价值观编码进AI中。当然，AI正变得越来越聪明，甚至核心AI也能相当快速地学习这些内容。你不必从零开始。

Okay? Number one. It is the case that in order to participate in AI, you have to encode within AI your history, your culture, your values. And of course, AI is getting smarter and smarter so that even the core AI is able to learn these things fairly quickly. You don't have to start from the ground, from ground zero.

因此我认为每个国家都需要具备一定的主权能力。我建议他们都使用OpenAI，使用Gemini，使用这些开放模型。你们可以使用Brock。

And so I think that every country needs to have some sovereign capability. I recommend that they all use OpenAI. They all use Gemini. They all use these open models. You use Brock.

我建议他们都这么做。我建议他们使用Anthropic。但他们也应投入资源学习如何构建AI。原因在于，他们不仅需要为语言模型构建AI，还需为工业模型、制造模型、国家健康与安全模型构建AI。他们必须自行培育一整套智能体系。

And I recommend they all do that. I recommend they all use Anthropic. But they should also dedicate resources to learn how to build AI. And the reason for that is because they need to learn how to build it not just for language models, but they need to build it for industrial models, manufacturing models, health National and security models. There's a whole bunch of intelligence they had to go cultivate themselves.

所以他们应当拥有主权能力。每个国家都应发展这种能力。

So they ought to have sovereign capability. Every country should develop it.

这是你所看到的吗？这是你在世界各地听到的情况吗？

And is that what you see? Is that what you're hearing around the world?

他们都意识到了。他们都意识到了。他们都将成为OpenAI、Anthropic、Grok和Gemini的客户。但他们也都真正需要建设自己的基础设施。这就是NVIDIA所做的重大理念——我们正在构建基础设施，就像每个国家都需要能源基础设施一样。

They all realize it. They all realize it. And they all are going to be customers of OpenAI, Anthropic, and Grok, and Gemini. But they all really need to also build their own infrastructure. And this is the big idea that what NVIDIA does is we're building infrastructure just as every country needs energy infrastructure.

通信和互联网基础设施。如今每个国家都需要AI基础设施。

The communications and internet infrastructure. Now every single country needs AI infrastructure.

那么让我们从世界其他地区开始说起。我们的好朋友大卫·萨克斯，这位人工智能领域的负责人，他的工作表现非常出色。

So let's start with the rest of the world. Our good friend David Sachs, the AI czar, who's doing a heck of a job.

我们非常幸运能在华盛顿特区拥有大卫和斯里拉姆。人工智能和AI技术。特朗普总统将他们安排在白宫是多么明智的决定。因为在这个关键时刻，技术问题非常复杂。

We are so lucky to have David and Sriram in Washington, D. C. AI and the AIs are. What a smart move by President Trump to put them in the White House. Because during this pivotal time, the technology is complicated.

斯里拉姆是华盛顿特区唯一我认为懂CUDA的人，这本身就很奇怪。但我非常欣赏的是，在这个技术复杂、政策复杂的关键时刻，对我们国家未来影响如此重大的时期，能有这样一位头脑清晰、愿意花时间理解技术并深思熟虑帮助我们的人。

Sri Ram is the only person in Washington, DC that I think knows CUDA, which is strange anyways. But I just love the fact that during this pivotal time when technology is complicated, policy is complicated, the impact to the future of our nation is so great that we have somebody who is clear minded, dedicating the time to understand the technology and thoughtful to help us through that.

在我看来，再次回到曼哈顿计划的类比，我们有一位明白这事关国家存亡的总统。我们有像德克萨斯州州长格雷格·阿博特这样希望取消监管以加速发展的州长，因为他们理解其重要性。我们还有能源部长赖特、内政部长道格·伯格鲁姆和商务部长卢特尼克，他们也明白这有多重要。

And it would seem to me again, going back to the Manhattan Project analogy, right that you have a president who understands how existential this is. You have governors like Greg Abbott in Texas who want to remove regulations to accelerate because they understand how important it is. You have Secretaries Wright at energy and Doug Berggrum at Interior and Lutnick at Commerce who also understand how important this is

他们是多么支持能源发展啊。你能想象如果我们现在有一个不支持能源发展、不希望国家能源增长的政府会怎样吗？那样我们可能连AI都发展不了，我简直不敢想。

How pro energy they are. Could you imagine the alternative if we had an administration right now who is not pro energy and want energy to grow in our nation so that we could have AI I just can't even think about it.

我觉得很讽刺的是，就在几年前我们还在说中国在建100座核反应堆，他们遥遥领先于我们。这就像是AI发展的原始阶段。但现在当我们开始建设时，每个人都说'哦，这过剩了'，对吧？

I find it ironic that just a couple years ago, we were saying, China's building 100 nuclear reactors. They're so far ahead of us. Like, that's the primitive to AI. But now you have people, when we go to build it, everybody says, Oh, it's a glut. Right?

在我看来，政府在这件事上是有自身利益的，而且我们看到产业和政府以一种我很久没见过的程度在合作。你在政界多年，与特朗普总统关系密切。请帮我们理解，当前产业与政府关系的本质是什么？我们上周看到了那场与所有CEO共进的晚餐。

Like, it seems to me that this is something that the government it is in their interest, and we have industry and government working together in a way that I haven't seen in a long time. You've been around a long time. You're very close with President Trump at this stage. Help us understand, what is the nature of industry government relationships? We saw that dinner last week with all the CEOs.

你花了很多时间。这是独一无二的吗？在你过去三十年的职业生涯中，见过类似的情况吗？

You spent a lot of time. Is it unique? Have you seen anything like this in your career over the last thirty years?

如你所知，过去去华盛顿特区很困难。几乎不可能预约到时间。特朗普总统对想要前来并帮助他们理解未来的领导人敞开大门。这是一个相信增长的政府。从根本上说，特朗普总统希望美国增长。

It was hard to go to DC in the past, as you know. Getting an appointment is almost impossible. President Trump has an open door to leaders who wants to come in and help them understand the future. This is an administration that believes in growth. Fundamentally, President Trump wants America to grow.

如果我们能在经济上增长，我们将在军事上变得强大。如果我们能在经济上增长，我们将获得安全。我从未见过一个贫穷的人能感到安全。作为一个国家变得富裕是国家安全的重要组成部分，他明白这一点。他还希望美国赢得人工智能竞赛。

If we can grow economically, we will be strong militarily. If we could grow economically, we will be secure. I've never met somebody who is secure who's poor. Being rich as a nation is an essential part of national security, and he knows that. He also wants America to win the AI race.

这将是一场非常长期的竞赛。他明白这是一个关键时刻。他希望科技行业能够蓬勃发展。他希望全世界都建立在美国的技术之上。这些都是明智、合乎逻辑的事情。

This is going be a very long term race. And he understands that this is a pivotal time. He wants the technology industry to run. He wants everybody in the world to be built on American technology. These are sensible, logical things.

相反的观点对我来说很奇怪。如果我把一切都颠倒过来，我们希望我们的国家不要增长。因为我们不希望国家增长，所以我们不需要任何能源，因为我们知道增长需要能源。所以让我们不要有任何能源。事实上，我们不希望我们的科技行业领先。

The opposite is strange to me. If I take everything and I just reversed it, we want our country not to grow. And because we don't want our country to grow, we don't need any energy because we know we need energy to grow. And so let's not have any energy. And in fact, we don't want our technology industry to lead.

他明白我们的科技行业是我们的国宝。没错。像过去的玉米、钢铁等一样，技术现在是如此基础的贸易机会。它是贸易的重要组成部分。为什么你不希望美国的技术被所有人觊觎，以便用于贸易呢？

He understands that our technology industry is our national treasure. Correct. And that technology, like corn and steel and things in the past, are now such fundamental trade opportunities. It's an essential part of trade. And why would you not want American technology to be coveted by everyone so that it could be used for trade?

对。那么让我们谈谈互联网，谷歌遍布全球。我们通过搜索传播了民主价值观。谷歌不需要去华盛顿获得许可就能做到这一点。它自然而然地发生了。

Right. So let's talk about the internet, Google spread around the world. We had democratic values spread around the world by way of search. Google didn't have to go to Washington to get permission to do it. It just happened.

我们将技术推广至全球。大卫·萨克斯已明确表示需要加快出口许可审批，以确保美国的人工智能技术在全球占据优势。对吧？我们讨论芯片、模型、数据中心等等。一年半前我们清楚这并未实现。

We diffused our technology around the world. David Sachs has been crystal clear of the need to accelerate export licenses so that the American AI stack wins around the world. Right? We're talking chips, we're talking models, we're talking data centers, etcetera. We know a year and a half ago that wasn't happening.

曾有个概念叫'小院高墙'之类。讽刺的是，这个政策建议被描述为在美国周围筑起'小院高墙'。这才是奇怪之处。我认为特朗普总统的观点正确——我们要最大化出口，最大化美国在全球的影响力。

There was a concept that was called small yard tall fence or something like that. Small yard tall And the irony of it was was described in such a way and it was recommended in policy in such a way, it was a small yard tall fence around America. That was the strange part. I think President Trump's got it right that we want to maximize exports. We want to maximize American influence around the world.

我们本就应该最大化这些方面。

We're supposed to maximize those things.

那么你看到这些许可审批加速了吗？华盛顿方面有进展吗？高层虽在表态，但政府内部是否真正在推动我们全球布局的加速？

And do see those licenses coming? Are you seeing the acceleration in Washington? I know it's being said at the top, but are you seeing it flow down through government that's accelerating us around the world?

卢特尼克部长全程跟进此事。很好。是的。

Secretary Lutnick was all over it. Great. Yeah.

现在让我们谈谈中国。要知道，多数人可能没意识到，我认为你对中国的理解不亚于美国任何一位领导人。

So now let's talk about China. You know, what most people may not realize is I think you understand China as well as any leader in The United States.

我们在那里已有三十年。

We've been there for thirty years.

在那里已有三十年。大多数人没有意识到的是，直到几年前，在中国市场上还占据着主导份额，就

Been there for thirty years. What most people don't realize is up until a couple years ago, had dominant market share within China in terms of

95%的市场份额。

95% market share.

95%的市场份额，可以说是在最重要的领域。而你曾说过，我们国家可能犯下的最大乌龙，就是假借试图减缓他们发展的名义，单方面解除了武装，迫使英伟达退出中国，这反而让华为得以凭借在中国的垄断利润加速发展。今早我还看到，华为、阿里巴巴等公司宣布将在全球

95% market share, in the most important thing, arguably. And you have said that our biggest own goal that we as a country could have, under the guise of somehow trying to slow them down, is we've unilaterally disarmed, we've forced NVIDIA out of China, which has allowed Huawei to accelerate on the back of monopoly profits within China. And I just saw this morning, you're seeing announcements out of Huawei and BABA and others that they're going to build data centers around the

各地

world

建设数据中心。华为制定了三年计划，要超越英伟达，资金来源于全球最大AI市场的垄断利润。看来你警告过将垄断市场拱手让给中国是个巨大错误，正在应验。总统在H20芯片禁令后表示，现在可以向中国出售芯片，但要征收15%的出口税。但如今中国似乎因美国的言论感到冒犯，表示不再允许英伟达在此销售。

now. Huawei has a three year plan to pass NVIDIA, funded by the monopoly profits in the biggest AI market in the world. So it's looking like your admonition that this is a huge mistake to hand China monopoly markets is coming true. The president said, you know, after kind of the ban on H20s, now we have a situation where you can sell chips to China, but there's a 15% export tax. But now it appears that the Chinese, perhaps offended by statements out of The United States, are saying, No, NVIDIA is not allowed to sell here now.

目前英伟达与中国的关系处于什么状态？能否重申你认为我们国家应采取哪些最佳策略，以在全球AI竞赛中胜出？

Where do we stand today between NVIDIA and China? And can you reiterate kind of what you think we as a country should be doing to put ourselves in the best position to win the AI race around the world?

我们与中国存在竞争关系。应当承认，中国理所当然希望其企业表现出色。我对此毫无怨言。他们理应发展得好，也应当给予企业充分支持。

We have a competitive relationship with China. We should acknowledge that China rightfully should want their companies to do well. I don't for a second begrudge them for that. They should do well. They should give them as much support as they like.

这完全是他们的特权。别忘了，中国拥有世界上一些最优秀的企业家，因为他们毕业于顶尖的STEM院校，他们是全球最具饥饿感的一群人。是的，众所周知，九月是个高产期。

It's all their prerogative. And don't forget that China has some of the best entrepreneurs in the world because they came from some of the best STEM schools in They're the the most hungry in the world. Yes. September, as you know. Is a very Producing

九个月内培养最多AI工程师的

the most AI engineers in nine the

96，观众都知道，朝九晚九，每周六天。这就是他们的文化。明白吗？我们面对的是一个强大、创新、渴望、快速行动且监管宽松的对手。懂吗？

ninety six, the audience knows, nine in the morning to nine at night, six days a week. That is their culture. Okay? We're up against a formidable, innovative, hungry, fast moving, under regulated. Okay?

人们没意识到这点。他们的监管非常宽松。

People don't realize this. They are very lightly regulated.

讽刺的是，相比我们资本主义体系，他们的监管更少。

Less regulated ironically than we are in a capitalist system.