黄仁勋谈 AI 资本支出、2027 年瓶颈与奇点

这对我们来说是一个生存问题。我们很容易停下来，但如果我们今天停下来，我们就输了。功能性、能力、安全性、保障——它们都处在完全相同的位置。头等大事。

This is an existential issue for us. We could easily stop, but if we stop today, we lose. Functionality, capability, safety, security—they're all at exactly the same place. Priority number one.

这是我们第三次一起做播客了。

It's our third pod back together.

在过去 10 天里，你出席了白宫的国宴。你参加了白宫峰会，我们在那里签署了关于超级智能的协议，而就在昨天，你和你的父母一起在那里，总统向你颁发了国家科学奖章。

In the last 10 days, you've been at the state dinner at the White House. You were at the White House summit where we did the accord on superintelligence, and just yesterday you were there with your parents when the president presented you with the National Medal of Science.

你是唯一一位与会的投资者。

You were the only investor at the table.

现在回想起来，你是唯一一位与会的投资者。

You were the only investor at the table, now that we reflected on it.

是的。

Yeah.

那跟我说说昨天和你父母一起在那里的情况吧。

Well, tell me about being there with your parents yesterday.

对，让我印象深刻的是——台上六个人中有五个都出生在美国以外。那是什么感觉？

Right, it struck me—five out of the six people on stage were all born outside of the United States. How did that feel?

嗯，首先，这本身就是极大的荣誉，对吧？这也反映了公司所做的贡献，对吧？显然，我们创造——我们发明了一种新的计算模型。这可能是当今世界上应用最广泛的计算模型。它是唯一一种存在于每个云中的计算架构。它存在于每家计算机公司。它被用于每个科学领域。它被游戏玩家使用。它被创作者和构建者使用，也被科学家使用。它驱动智能，对吧？我的意思是，如果你退一步看看公司取得的成就——我们最初是为游戏玩家制造图形芯片，当然我至今仍为此感到自豪。我们现在已经出货了 16.7 亿台，对吧？我培养了几代科学家、孩子和设计师。所以我认为其影响——现在我们就是驱动社会智能的基础设施。因此，代表公司获得这个奖项对我来说是无比自豪的时刻。至于我的父母，我想让他们去，我给了他们仅仅 24 小时的通知时间，他们不太常旅行，对吧？他们不太常旅行。我父亲甚至没有一套西装。

Well, first of all, it's a great honor all by itself, right? And it's a reflection of the contribution that the company has made, right? And obviously, we created—we invented a new computing model. This is probably the most pervasive computing model in the world today. It's the only computing architecture that's in every cloud. It's in every computer company. It's used in every field of science. It's used by gamers. It's used by creators and builders, and it's used by scientists. It powers intelligence, right? I mean, just if you take a step back and you look at what the company has accomplished—we started out building a graphics chip for gamers, and I'm still proud of that, of course. And we now shipped 1.67 billion units, right? That's how many generations of—I mean, I raised generations of scientists and kids and designers, you know. And so I think the impact—and here we are, we're the infrastructure that powers the intelligence of society. So it's an incredibly proud moment for me to receive the award on behalf of the company. Now for my parents, I wanted them to go, and I gave them literally 24 hours lead time, and they don't travel that much, right? They don't travel that much. And my dad didn't even have a suit.

哇。

Wow.

是的。所以他跑出去买了一套西装。他看起来很棒。他们看起来很棒。

Yeah. So he ran out and bought a suit. And he looked great. They looked great.

是的。

Yeah.

那是一套全新的西装，他出现了。我没想到他有西装。我以为他只会穿件衬衫。他出去买了一套西装。我妈妈买了一条新裙子，真的就在 24 小时内。

It was a brand new suit and he showed up. I didn't realize he had a suit. I thought he was just going to wear some shirt. And he went out and bought a suit. My mom got a new dress, literally in 24 hours.

是的。哦，我知道那是怎么回事。

Yeah. Oh, I know how it works.

哦，天哪。太不可思议了。所以我们让他们——我们为他们包了一架飞机，把他们送到那里，他们非常高兴见到总统。特朗普总统扶着我妈妈——因为她走路困难——他们两人一起走下舞台，他扶着她，带她走下舞台，他们非常自豪能见到他。但对他们来说，这是他们的美国梦。你知道，我爸爸——我还记得，我想我当时大概 5 岁左右。他刚从美国出差回来。他在一家空调公司工作。

Oh my gosh. It's incredible. And so we got them on a—we chartered a plane for them, got them there, and they were so happy to meet the president. And President Trump walked my mom—because she has a hard time walking—and the two of them were walking off stage and he held her and walked her down the stage, and they were just incredibly proud to meet him. But for them, it's their American dream. You know, my dad—I still remember, I think I was like 5 years old or something like that. And he had just come back from the United States on a business trip. He was working for an air conditioning company.

当时你们住在台湾。

And you were living in Taiwan.

我们——当时我们刚搬到泰国。泰国。

We were—at the time we had just moved to Thailand. Thailand.

哦不，不，我们正要搬到泰国。当时我们在台湾。他回来了——

Oh no, no, we were just about to move to Thailand. And we were in Taiwan. He had come back—

他在美国为一家叫开利的公司工作，一家空调公司。他来这里是接受培训。世界上那么多地方，他们偏偏把他送到了纽约。

He was in the United States working for a company called Carrier, an air conditioning company. And he was here for training. And of all the places in the world, they landed him in New York.

好的。

Okay.

所以你可以想象——一个 20 多岁的小伙子，我想是 25 到 29 岁左右，从未离开过台湾，去了美国，偏偏是纽约市。他看到了未来的未来。他回来时，你知道，被震撼了，他在机场买的所有东西都带了回来。他买了巧克力。他带了那种老式管装奶酪。有点像白色的奇怪奶酪。它装在管子里。

And so here you can just imagine—a kid in the late 20s, I guess mid-late 20s, and never been outside of Taiwan, goes to United States, New York City of all places. And he saw the future of the future. And he came back, you know, mind-blown, all the things that he bought at the airport brought it back. He bought chocolate. He brought that cheese that you have in a tube back in the old days. It's kind of like white strange cheese. It comes in a tube.

很棒。

It's great.

他带回了各种奇怪的东西。他说：‘这——我们有一天应该去那里。’他们一有机会，就把我和我哥哥送到了美国。你知道，这是他的梦想。我想很多人没有意识到你比你的父母先来到美国。

And he brought back all kinds of strange things. And he says, 'This is—we should go there someday.' And the first opportunity they had, they sent my older brother and I to the United States. And you know, it was his dream. I don't think many people realize that you came to the United States ahead of your parents.

是的。

Yeah.

对。他们几年后也来了，但昨天台上六个人中有五个出生在美国以外，这让我深受震撼。你和我讨论过，在所有其他事情发生的情况下，我们需要一个真正的战略和移民政策，确保我们仍然是世界上所有伟大的建设者都想来建设的地方。这是美国的超级力量，对吧？在很多方面，我们国家的知识产权创造能力是世界上独一无二的。原因就是我们吸引了世界上最优秀的人才。没错。你知道，我们是他们想要来建立生活的地方，因为它提供的机遇、它提供的自由、它提供的可能性、发明和创业精神以及‘我能行’的态度。你知道，在美国，一切皆有可能。

Right. And they joined a few years later, but it just hit me like a ton of bricks that five out of the six people on that stage yesterday were born outside the United States. You and I have talked about the need for—with all the other stuff going on—a real strategy and an immigration policy that makes certain that we're still the place that all the great builders in the world want to come build. This is America's superpower, right? In a lot of ways, the intellectual property creation capability of our nation is unlike anything else in the world. And the reason for that is we attract the world's best. That's right. You know, we are the place where they want to come and build their lives because of the opportunity it provides, the freedom that she offers, the opportunity that it provides, the spirit of invention and entrepreneurialism and the can-do attitude. You know, in the United States, everything is possible.

对。

Right.

我们从‘当然可能，当然我们可以做到’开始，而不是——我经常旅行——在其他地方是‘当然你做不到，当然’。我们把这视为理所当然。但这里创造的文化所带来的精神、能量——怎么强调都不为过。所以我认为流入这个国家的知识产权是世界上前所未有的。

We start with 'of course it's possible, of course we could do that' instead of—I do a lot of travel—'of course you can't do that, of course' in other places. And we take it for granted. But the spirit, the energy that comes around with the culture that we've created here—it just can't be understated. And so I think the inflow of intellectual property into this country is unlike anything else the world's ever seen.

确实如此。

It's true.

我们想要这样。我们希望世界上所有伟大的头脑都来这里，带来他们伟大的想法，在这里实现它们。

And we want that. We want all of the world's great minds to come here, bring their great ideas, make them realize them here.

在这里实现它们，让我们保持这个飞轮转动。

Realize them here and let us keep that flywheel going.

你知道，我真的——这是一个微妙的系统。这是一个微妙的平衡。我明白这是一个微妙的平衡。所以我们必须合法地做。必须正当地做，对吧？但我们当然想这样做。

You know, I really—it's a delicate system. It's a delicate balance. I get it's a delicate balance. And so we have to do it lawfully. You have to do it rightfully, right? But we surely want to do it.

是的，当然。

Yeah, for sure.

所以，我想先从宏观层面开始，谈谈我们现在所处的位置。然后我想深入探讨一下安全和监管，以及如何把这件事做对，还有我们在算力建设方面的进展。最后，我们会稍微深入一下 Nvidia 的增长和那边的情况。但你一直很擅长预测未来，对吧？两年前，我们坐在这里，你说市场严重低估了未来五年所需的算力建设规模。你说不会是 1 万亿，至少会是 2 万亿的累计算力建设。当然，现在我们已经超过了。在所有人还没想到的时候，你就说推理——算力会因为推理时的思考而增长 10 亿倍。然后去年，当然，在我们看到 Opus 4.6 之前，在我们看到 Claude Code 之前，你说我们正在进入智能体的时代，对吧，我们将解锁很多这些能力。当然，我们看到了 Opus 4.6 和 Claude Code 真的起飞了。那么，我们现在在哪里？

So, I want to start high level just kind of where we are. Then I want to dig down a little bit into the safety and regulation and getting this right, where we are in terms of the compute buildout. And then finally, we're going to finish up drilling in a little bit to Nvidia's growth and what's going on there. But you've been really good at calling the future, right? Two years ago, we were sitting here and you said that the market was dramatically underestimating the amount of compute buildout that was going to be required and was going to happen over the course of the next five years. You said it's not going to be a trillion, it's going to be two trillion at a minimum in terms of cumulative compute buildout. And of course, now we're over that. You said before everybody was thinking about it that inference, right, compute was going to 1 billionx because of inference time reasoning and thinking. And then last year, of course, before we saw Opus 4.6, before we saw Claude Code, you said we're entering the age of agents, right, where we're going to unlock a lot of these capabilities. And of course, we saw Opus 4.6 and Claude Code really took off. So, where are we now?

对，我们有 Elon 和 Sam，而我说过的事情只有三件成真了。还有另外 200 件呢。你为什么不保留那个？就让那成为播客吧。

Right, we have Elon and Sam and only three things that I said that came true. There's another 200. Why don't you just keep that? Let's just let that be the podcast.

所以，我们几周后的投资者日是关于进入奇点，这种加速。我们有 Elon、Sam 和 Demis 都说，“嘿，我们正在山脚下或正在进入奇点。”我们有模型解决 Navier-Stokes 方程，如你所知。那么，当你看向你的水晶球时，我们在哪里，我们是在拐点，是在加速进入超级智能的不同阶段吗？

So, our investor day in a couple weeks is on entering the singularity, this acceleration. We have Elon and Sam and Demis all say, "Hey, we're in the foothills or entering the singularity." We have model solving Navier Stokes as you know. So when you look into your crystal ball where are we and are we inflecting are we accelerating into a different phase of super intelligence.

让我们先解开这一点。首先，为什么我们需要的算力会比我们最近，比如去年，多 10 亿倍，这非常合乎逻辑。原因在于，思考和解决问题比重复、 regurgitating 你记住的东西需要更多的推理。

Let's unpack that first. First of all it's very fairly logical why the amount of compute that we're going to need is a billion times more than where we were just recently, say last year. And the reason for that is because thinking and problem solving takes a lot more inferences than repeating something, regurgitating something that you memorized.

对。

Right.

所以当你使用 ChatGPT 时，你给它一个提示，大部分知识都编码在大语言模型中。不幸的是，当然，它会幻觉，所以下一步是生成，但也基于或受限于真实数据。

So when you're a ChatGPT, you give it a prompt, most of that knowledge was encoded into the large language model. Unfortunately, of course, it hallucinated and so the next click was it was generative, but it was also grounded or conditioned by ground truth.

是的。

Yes.

所以它必须出去搜索，快速搜索，带回真实数据，然后从那种条件中，加上它自己在预训练中已经学到的东西，它会生成新的答案。嗯，我们真正想做的有趣的事情大多需要你逐步推理，想出几个选项计划，你可能会决定让我们看看这些如何展开，然后基于所有这些答案，我反思到目前为止最好的答案，最好的路径，我可能选择一两个或三个，展开它们。

And so it would have to go out and do a search, do a quick search, comes back with the ground truth, and then from that conditioning augmented with its own what it already learned pre-training and it would generate new whatever the answer is. Well, most of the things that we want to really do that are interesting requires you to reason step by step to come up with several options plan and you might decide let's see how these things all roll out and then based on the answers of all of those I reflect on the best answer so far the best path so far I might choose one or two or three of them roll them out.

对。

Right.

好，修剪其余的，然后你继续这样做，你知道人们我们这样做，这就是为什么我们说伟大的战略家喜欢战略选择权。一方面你不想太快修剪东西，但你也必须优化你的资源。嗯，AI 正在这样做。好。所以它们保持战略选择权。它们展开东西。它们修剪。它们反思，它们修剪，等等。所有这些思考，所有这些探索都需要更多的算力，而且不仅如此，因为它现在能够做生产性的事情，我们会用它做更多事情。所以两者复合：它使用更多算力来完成工作，但它做得很好，因此我们会应用，我们会雇佣它做更多工作。

Okay prune the rest of it and then you just keep on doing this and you know people we do this that's the reason why we say great strategists love strategic optionality. You don't want to prune things too fast on the one hand, but you also have to optimize your resources. Well, AIs are doing that. Okay. So they're keeping their strategic optionality. They're rolling things out. They're pruning. They're reflecting, they're pruning, so on and so forth. All of that thinking, all of that exploration just takes a lot more compute and not on top of that because it's able to now do productive things, we're going to use it for more things. And so the compounding of both it uses more compute to do its job, but it's doing a good job and therefore we're going to apply we're going to hire it to do more jobs.

正是如此。

Exactly.

好。所以这一切都复合。在这一点上，我会说去年的重大观察是，一方面我们希望模型越来越好，我们仍然希望模型变得更好，模型的缩放定律将继续。然而，附加一个框架，基本上是大语言模型的外骨骼，对吧？

Okay. And so that all compounds. At this point I would say that the big observation from last year is that on the one hand we want the models to get better and better and we still want the models to get better and the scaling laws of models are going to continue. However, the attachment of a harness, basically the exoskeleton of the large language model, right?

你知道，有点像大脑的身体。

You know, kind of the body of the brain.

这允许它管理工作记忆，管理知识。你如何从工作记忆到短期到长期，基本上制定你的知识图谱，这就是我们所有人所做的？我们有点想，嘿，你知道，我学到了什么，它如何融入我过去知道的一切，你知道当我们睡觉时我们做压缩，那些不合理或无关紧要的事情我们最终会忘记，但那些重要的事情不知何故会重新连接，所以我们在离线做那种事情，所以我们有工作记忆，我们有离线知识提取和创建知识图谱创建，记忆的整个部分都是大事，另一件事当然是工具使用，我们已经教这些工具这些 AI 如何与人类制造的最重要的工具互动。你知道，这就像如果一个机器人来到我们家，它更有可能使用我家里所有的工具，而不是从零开始建造。所以 spice 仍然有用，Verilog 会有用，模拟器会有用，你知道设计编译器、Photoshop 和 AutoCAD，你知道 cell kia 和所有 seammen 的工具，所有这些我们用来运营公司、建立公司、设计东西、操作东西的不同工具，它们都将成为智能体使用的伟大工具，但智能体必须学会如何使用它，所以这两个领域知识和 工作记忆和工具使用，去年取得了巨大突破。

That allows it to manage working memory, manage knowledge. How do you go from working memory to short-term to long term to basically formulating your knowledge graph, which is what we all do? we kind of think about, hey, you know, what did I learn and how does it fit into everything that I know in the past and you know how when we go to bed we do compaction things that don't make sense we kind of or don't matter we kind of eventually forget but the things that kind of matter somehow it gets reconnected and so we're doing that kind of offline and so we got working memory we have offline knowledge extraction and creation knowledge graph creation that whole part of memory is a big deal the other thing is of course tool use and we've taught these tools these AIs how to interact with the most important tools that the humans have ever made. You know, it's no different than if you were if you were a robot comes into our house, it's more likely you're going to use all the tools that I have in the house than to go off and build it from scratch. And so spice is still going to be useful and verilog will be useful and simulators will be useful and you know design compilers and Photoshop and AutoCAD and you know the cell kia and all the seammen's tools all of these different tools that we've been using to run our companies build our companies design things operate things they're all going to be great tools for the agents to use but the agents had to learn how to use it and so these two areas knowledge and working memory and tool use huge unlocks in the last year.

所以现在突然之间，你可以把这些大语言模型封装在它们的框架或外骨骼中，你可以把它们放入你的公司，将它们连接到你拥有的信息、真实数据，将它们连接到它们可以使用的工具，突然之间它们开始做惊人的事情。

So now all of a sudden you could put these large language models encapsulated in their harness or exoskeleton and you can put them into your company, connect them to the information, the ground truth that you have, connect them to the tools that they can use and all of a sudden they start doing amazing things.

坦白说，最简单的部分是编程。这算是通往有用工作的第一个入口。在很多方面，编程之所以容易，是因为客户和他们使用的工具都相当习惯于那些用途巨大但并非极其易用的东西。在编程的早期，部署并不容易，但软件工程师找到了办法。

Now the easiest part, frankly, was coding. That's kind of the first on-ramp to useful work. In a lot of ways, coding is easy because both the customer and the tools they use are fairly accustomed to things that have tremendous use but are not extremely easy to use. In the early days of coding, it wasn't extremely easy to deploy, but software engineers found a way.

所以现在我认为我们的下一个突破点就是——你们可能已经在玩 Astra 了。OpenAI 的 Astra 真的非常棒，我们在各种不同的地方都在用它。所以这些智能体的下一个阶段是，它们能够做设计工作和很多创意营销工作。我们正在用它来——你拍一张照片给 Astra，它真的能把它重建成 CAD。

So now I think the next click for us is now—and you guys have probably been playing with Astra. OpenAI Astra is really fantastic, and we're using it in all these different places. So now the next phase of these agents is they're able to do design work and a lot of creative marketing work. We're using it to—you take a picture and give it to Astra, it literally recreates it into CAD.

是的。

Yeah.

它把它重建成 CAD，这太不可思议了，因为我们需要 CAD。

It recreates it into the CAD, which is incredible because we need the CAD.

对。

Yes.

这样我们就能与 CAD 交互，进行微调和调整。所以 Astra 真的能极其高效地给我们提供最初的实现。现在你拿 Astra 和最新版本，把它应用到所有这些不同的领域。我认为除了软件工程，我们还会看到设计师、建造者、创作者，以及整个领域——科学家和所有这些被解锁的东西。

So that we can interact with the CAD and make fine-tuning and adjust. So Astra could really give us the first initial implementation incredibly productively. Now you take Astra and the latest release and you apply it to all of these different domains. I think that besides software engineering, we're going to see designers and builders and creators and this whole area—scientists and all of this whole unlocking.

当然，挑战在于我们如何将智能体安全可靠地部署到全球的企业中。

Now the challenge, of course, is how do we deploy agents safely and securely into the world's enterprise.

这又完全是同一个关于安全与安保的问题。我们会谈到——我们会讨论这一点。

This is exactly the same question again with safety and security. And we'll talk about—we'll come in on that.

让我印象深刻的是，这一切都发生在今年，OpenAI 和 Anthropic 各自拥有，我不知道，两到三千兆瓦的算力。好吧。所以相对于即将到来的算力，这还算是相对较小的量。好吧。所以我感兴趣的是，经济通常运作的方式——或者人们建模经济、建模公司的方式——是线性减速。他们不太擅长建模加速。但这里我们面对的是双重指数。我们有大量新算力上线。所以如果 OpenAI 和 Anthropic 明年把算力扩大 3 到 5 倍，这合理吗？我的意思是，我们解决了纳维-斯托克斯方程，我们会回顾这一点并说那只是很小的算力和智能，是我们从此往后最差的智能，却解决了一个千禧年难题。所以当你看到明年即将上线的东西时，你预期我们会在长时间运行的智能体问题上取得重大突破，让我们能够使用这些智能体集群吗？

What I'm struck by is that all of this occurred this year with OpenAI and Anthropic having, I don't know, two, three gigawatts of compute each. Okay. So relatively small amount of compute relative to what's coming. Okay. And so again what I'm interested in is the way the economy generally works—or people model the economy, or they model a company—is linear deceleration. They're not very good at modeling acceleration. But here we have this double exponential. We have massive amount of new compute coming on board. So if OpenAI and Anthropic are, I don't know, 3 to 5xing their compute next year, does it stand to reason? I mean, we solved Navier-Stokes, we'll look back at this and say that was a tiny amount of compute and intelligence, the worst intelligence we're ever going to have from here forward, solving a millennium problem. So when you look at what's coming online next year, do you expect that we're going to have major breakthroughs in long-running agentic problems that we can use these swarms of agents on?

嗯，纳维-斯托克斯确实是个很好的例子。记住，OpenAI 并不是从零解决纳维-斯托克斯的。

Well, Navier-Stokes is really a great example. Remember, OpenAI did not solve Navier-Stokes from nothing.

没错。

Correct.

他们是在许多伟大科学家、伟大数学家和大量伟大前人工作的基础上解决的。我之前说过，我们正处在许多不同领域的临界点，因为我们在纳维-斯托克斯上正处于那个阶段，大量基础工作已经完成。如果我们能把 AI 的穷举探索应用到先前完成的工作上，我们很快就会看到一个又一个突破。但要做到这一点，我们确实需要大量算力。原因在于，大量的探索是必要的。如果只是某人解决某个基本原理性理论就能发现突破，如果不需要对搜索空间——不是太空，是搜索空间——进行大量探索，如果不需要那样，那么早就有人找到了，对吧？人是很聪明的。所以我认为有很多这样的问题正处在悬崖边上，真的只需要那种大规模的智能探索就能带我们到达那里。

They solved it from the foundation of a lot of great scientists and a lot of great mathematicians and a lot of great previous work. Earlier I said we're at the tipping point of a lot of these different fields because we're at that phase with Navier-Stokes, with a great deal of foundational work already done. If we could just apply the exhaustive exploration of AI to the previously done work, we're going to see one breakthrough after another rather quickly. But we do need a lot of compute to do that. And the reason for that is because a large amount of exploration was necessary. If it was simply somebody resolving some fundamental first-principles theory in order to discover the breakthrough, if it didn't require a vast amount of exploration of search space—not space space, but search space—if it didn't require that, then somebody would have found it earlier, right? People are pretty smart. And so I think that there are many problems like this that are just on a precipice and really want to just have that massive amount of intelligence exploration to get us there.

在未来几年，我预期智能的成本会相当快地下降。原因很明显，我们正在上线更多算力。这些模型变得越来越聪明。当你更聪明时，你可以更高效。所以这两者的结合——智能的成本会下降，这当然正是他们想要的，对吧？这和摩尔定律没什么不同。如果你回到三十多年前，我们有三十年的计算成本每年下降一半。我们希望智能成本每年下降一半或二十分之一，因为对它的消费会远远弥补这一点。

In the coming years, my expectation is that the cost of intelligence will fall fairly quickly. And the reason for that is because obviously we're bringing more compute online. These models are getting smarter. When you're smarter, you can be more efficient. And so the combination of these two things—the cost of intelligence will come down, which is of course what they want, right? And this is no different than Moore's law. If you go back 30-some years, we have three decades of computing cost decreasing by a factor of two every single year. And we want intelligence cost to decrease by a factor of two or 20 every single year because the consumption of it is going to more than make up for it.

没错。远远弥补。

That's right. More than make up for it.

所以世界上有一个巨大的空白空间等待被填补。机会的真空，我们只需要把成本降下来。

And so there's a massive empty space in the world that wants to be filled. Empty vacuum of opportunity, and we just got to get the cost down.

关于纳维-斯托克斯，让我印象深刻的是他们有一万个智能体一起工作。在我看来，这就像一个超大规模公司，基本上把所有智能体都指向一个目标协同工作。我认为在早期阶段，现在有很多是瘫痪式的搜索探索，但我认为目标是更多这样的智能体能够为一个共同目标一起工作，我们能够解决更多问题。正如你所说，我认为我们的数学计算表明，解决这个纳维-斯托克斯实例大概需要 400 万美元的算力。未来几年做类似事情的成本还会再下降 5 倍、10 倍。所以我们将能够做的事情的数量正在迅速扩大。

I guess what stands out to me about Navier-Stokes is that they had 10,000 agents working together on this. It looks to me like a hyperscale company basically all pointing agents to work together towards a goal. And I think in the early days right now it's a lot of paralyzed search exploration, but I think the goal is that more of these agents can work together to a common goal and that we can solve a lot more. And to your point, I think our math says probably cost about $4 million of compute to solve this instantiation of Navier-Stokes. The cost to do something like that over the next few years is going to drop by another 5x, 10x. So the number of things that we're going to be able to do is rapidly expanding.

我们的雄心会直冲云霄。最初的 GPT 大概花了——我们随便说个数——也许 2000 万美元。如果你想现在训练 GPT-1，大概要花多少，一千，也许 100 美元？

Our ambition is going to go sky-high. The original GPT probably cost—let's pick a number—say maybe $20 million to do. If you wanted to train GPT-1 now, it'll probably cost you what, a thousand, $100 maybe?

大概一百美元。是的。

It's like a hundred bucks. Yeah.

好吧。所以你从 2000 万美元，就算 1000 万美元，降到了 100 美元。而我们是在 10 年内做到的。

Okay. And so you went from $20 million, let's say $10 million, to 100 bucks. And we did that in 10 years.

是的。

Yeah.

好吧。所以我完全预期——与此同时我们的雄心变得如此之大，我们现在训练的模型是 3 万亿、10 万亿、30 万亿参数。我们现在谈论的是巨型模型。而我们对技术能力的雄心也提高了。

Okay. And so I fully expect—meanwhile our ambition got so large that the models that we're training now are 3 trillion, 10 trillion, 30 trillion. We're talking about gigantic models now. And our ambition for the capabilities of technology has gone up.

与此同时，我们对它的使用量从 10 年前到现在大概增长了一百万倍。所以我认为这个趋势不会停止。我认为对智能的需求是相当无止境的。你越聪明，你就有越多应用智能的想法。

Meanwhile, our use of it has gone up between 10 years ago to now probably a million times. And so I don't think this trend is going to stop. I think there's a fairly insatiable demand for intelligence. And the more intelligent you are, the more ideas of applying intelligence you have.

你知道，我在生活中越学习，就越意识到有更多事情要做。这是因为每当你学到新东西，就会打开一扇新门。每次学习，都会打开一扇新门。每当你打开一扇新门，你就会看到更多机会。你看到更多挑战。你看到更多问题。所以这就是未来，对吧？

You know, it's over the more I've learned in life, the more I realize there's more to do. And that's because every time you learn something, it opens a new door. Every time you learn, it opens a new door. And every time you open a new door, you see more opportunity. You see more challenges. You see more problems. And so that's kind of the future, right?

你知道，人们说我们会有更多时间。我完全不同意。我认为对大多数人来说，这种智能将解锁更多机会。我们只是会，他们会想要追求它，因为他们有雄心。

You know, people say we're going to have more time. I completely disagree. I think for most people this intelligence is going to unlock more opportunities. We're just going to and they're going to want to pursue it because they're ambitious.

是的。你知道，给我举个例子，一个家庭，一个父亲，一个儿子，你知道，一个丈夫，一个妻子，一个母亲，谁不想要它，谁没有让家庭拥有更多的雄心。只是他们只是没有机会，而这些事情将打开机会。

Yes. You know, give me an example of a family, a father, a son, you know, a husband, a wife, a mother who doesn't want it, who don't have ambition for their families to have more. It's just they just don't have the opportunity and these things are going to open up opportunity.

如果你回到 1930 年，凯恩斯谈到了他对工业革命的预测，他说，你知道，一百年后经济将增长 8 倍，但每周工作时间将减少到 15 小时。他说对了。经济确实增长了 8 倍，对吧？我们从中获得了巨大的富足和生产力提升。

If you go back to 1930, Keynes talked he made a prediction about the industrial revolution and he said, you know, in a hundred years the economy will have 8xed, but the work week will go to 15 hours a week. He was right. The economy did add 8x, right? We had massive abundance and productivity gains from it.

是的。只不过我们每周工作 80 小时，但我们没有变成每周 15 小时。他错在没理解人类精神，对吧？他没有意识到工作就是目的。发现就是目的，对吧？我认为这深深编码在我们的 DNA 中。我现在比以往任何时候都更努力工作，也比以往任何时候都更快乐。不是因为我必须像我的祖父母那样，对吧？我这样做是因为我想做。所以我同意你的观点。

Yeah. Except we're working 80 hours, but we didn't go to a 15 hour week. And what he got wrong was the human spirit, right? He didn't realize that work is purpose. Discovery is purpose, right? It's deeply encoded in I think our DNA. And I'm working harder than I ever have, having more fun than I ever have. Not because I have to, like my grandparents did, right? I'm doing it because I want to. And so I agree with you on that.

我对这个想法很着迷。你知道，奇点假设认为，我们现在有所有这些智能体可以一起在长期任务上工作。它们可以保持专注。我喜欢你的让彼此保持专注，作为一个团队工作。嘿，你偏离了。

I'm fascinated with this idea. You know, the singularity postulates that, you know, we've kind of now have all these agents that can work together on long-running tasks. They can stay on task. I love your keep each other on task, working as a team. Hey, you're off.

没错。我喜欢你的想法，你知道，我们在所有已经出现的东西之上构建，对吧？所以如果我以全自动驾驶为例，我们讨论了十年，我们真的没有达到，对吧？但突然之间，就像你上 Twitter，每个人都在体验特斯拉 FSD，体验 Waymo，我们到了。有人问我：“今天早上开车过来怎么样？”我说：“我不知道。”因为我上了车，启动了 FSD，直到我们开进停车场我才注意。对吧？我一路都在工作。但如果我在科学领域思考这一点，制药和遗传学领域将出现的突破，对吧？我们已经铺设了 30 年的轨道，我们将在世界上许多大规模治愈方法上取得突破。当然，ChatGPT 已经在帮助父母规划他们的一天，解决孩子的健康问题，并帮助他们获得更好的教育。

That's right. I love your idea that, you know, we build on top of all the things that have come, right? And so if I think about full self-driving as an example, we talked about it for a decade and we really didn't get there, right? But all of the sudden, like you go on Twitter, everybody experiencing Tesla FSD, experiencing Waymo, we're here. Somebody asked me, "How was the drive here this morning?" I said, "I don't know." Cuz I got in my car and I hit FSD and I didn't pay any attention until we pulled into the parking garage. Right? I was working the whole way here. But if I think about that in the fields of science, the breakthroughs that are going to come in the fields of pharma and genetics, right? We've been laying down the tracks for 30 years, we're going to break through on a lot of massive cures for the world. Of course, ChatGPT is already helping parents plan their day and solve their kids' health issue and helping them get better educated.

不过请跟我们谈谈为什么它这么难突破。Jensen，我们有所有这些正在改善生活、即将到来的东西。所以对我来说有点好奇，为什么它没有突破。

Talk to us a little bit though as to why that's having such a hard time breaking through. Jensen, we have all of these things that are out there improving lives and that are right around the corner. And so that's kind of curious to me as to why that's not breaking through.

我想让我试着回答一下。是的。你知道，我们最终不会真正知道。我和很多人谈过。他们告诉了我他们的想法。我的意思是，其中一个原因当然是 AI 或超级智能是软物质。是的。你看不到它。所以全世界都在谈论这项技术，但你看不到它。你摸不到它。即使他们已经从中受益，当他们搜索时，他们已经从中受益。当他们购物时，他们从中受益。当他们使用 Netflix 时，他们从中受益。你知道，你去沃尔玛，你从中受益。联邦快递把东西送到我们这里，你知道，今天这一切都在利用超级智能和加速计算。但我们没有很好地讲述这个故事。

I guess let me take a shot at it. Yeah. And you know, we'll ultimately not really know. I've spoken to a lot of people. They told me what they thought. I mean, one of the things, of course, is is AI or superintelligence is soft matter. Yeah. You can't look, you can't see it. And so you got the whole world talking about this technology, but you can't see it. You can't touch it. And even though they're already benefiting from it, when they do search, they're already benefiting from it. When they go shopping, they're benefiting from it. When they're using Netflix, they're benefiting from it. You know, you go to Walmart, you're benefiting from it. The fact that FedEx is getting stuff to us, you know, it's all leveraging superintelligence and accelerated compute today. But we don't do a very good job telling that story.

所以生活是，他们从中受益，只是它是无形的。

So life is they're benefiting from it's just it's invisible.

对。所以如果我们关闭数据中心，关闭超级智能一个周末，我想每个人都会注意到。

Right. So if we shut off the data centers and shut off superintelligence for a weekend, I think everybody would notice.

对。我知道你特别有一个关于机器人技术的问题。

Right. I know you had a question about robotics in particular though.

是的，我想你今年多次说过，机器人技术的 ChatGPT 时刻已经到来，我还认为你说过它可能是你的下一个千亿美元市场。所以我想你现在看到了什么，今年的转折点是什么？

Yeah, I guess you've you've said notably said a few times this year that the ChatGPT moment for robotics is here and I think you've also said that it could be your next hundred billion dollar market. So I guess what are you seeing now in the inflection point in this year?

是的。这即将到来。是的。我低估了千亿美元，因为今年前六个月我们在物理 AI 上做了 100 亿美元。物理 AI 基本上是，你知道，以某种方式与物理世界交互的 AI。它可以是光子计数 CT 机。可以是超声波。可以是新的成像雷达。5G 下一代。它不仅理解无线电波，还将理解各种信号。它是一个传感网络。当然还有机器人技术、自动驾驶汽车、制造系统、视觉检测、地理空间卫星、对世界成像并将其重建为 3D。所以基本上，如果你看整个电磁频谱，从可见到不可见频谱都有 AI 在发生。

Yeah. That this is around the corner. Yeah. I understated the hundred billion dollars because we're already the first six months of this year we did $10 billion in physical AI. Physical AI is basically you know AI that's somehow interacting with the physical world. And it could be photon counting CT machines. It could be ultrasounds. It could be new imaging radars. 5G next generation. Not only does it understand radio waves, it's going to understand signals of all kinds. It's a sensing network. Of course robotics, self-driving cars, manufacturing systems, visual inspections, geospatial satellites, imaging the world and reconstructing it into 3D. And so there's basically if you look at the entire electromagnetic spectrum, there's AI stuff going on from visible to invisible spectrums.

好的。

Okay.

所以我们正在整个领域开展工作。作为 AI，有感知部分，有推理部分，然后是规划和执行部分。当你最终来到规划和执行时，那就是自动驾驶汽车和机器人技术之类的东西，你知道，在仓库送货车辆中，拾放机器人，你不必编程得那么精确，因为它自己学习，你知道，一直到人们谈论的消费机器人，你知道，进入你的房子，你知道，整理你的床等等。所以，或者医疗机器人，医院机器人，有很多很好的例子。

And so we're doing work across that entire thing. And as an AI, there's the perception part of it, there's the reasoning part of it, and then there's the planning and actuation part of it. When you finally come to planning and actuation, then that's self-driving cars and robotics and things like that, you know, in warehouse delivery vehicles, pick and place robots that you don't have to program as precisely because it learns by itself, you know, all the way to people talking about consumer robots, you know, coming into your house, you know, making your bed and things. And so, or healthcare robots, hospital robots, lots of really good examples there.

精准地把你的处方送到病人手中，让机器人护理真正做到 24/7、每一秒都不间断，以保障病人的安全。这样的例子太多太多了。整个手术室都将完全机器人化，对吧？大量伟大的技术正在被部署。

Precisely delivering exactly your prescription to the patient, having robotic care literally 24/7 every single second to keep patients safe. And so just lots and lots of great examples. Entire surgical rooms that are going to be completely robotic, right? And so lots of great technology being deployed.

所以如果你看这整件事，我们是全球最大的物理 AI 公司。我们前六个月做了 100 亿美元。去年我们做了 100 亿美元。

And so if you look at this whole thing, we're the world's largest physical AI company. We did $10 billion in the first six months. We did $10 billion in the last year.

嗯。

Mhm.

那是我们第一个 100 亿美元的年份。再往前一年我们做了 50 亿。

And that was our first $10 billion year. The year before that we did five.

对。

Yeah.

好。所以是 50 亿、100 亿。假设我们做到 200 亿。谁知道呢？所以也许那个 100 亿说得还有点保守了。

Okay. And so it went five, 10. Let's say we did 20. Who knows? So maybe the hundred was a little understated.

对。

Yeah.

我认为其中的逻辑在于，归根结底，对我们重要的一切都触及物理世界，而我们生活在物理世界里。所以我认为，把软件、计算、智能带入一个历史上主要是机械、电气的行业——这是一个巨大的机会。

And I think the reasoning part of it is everything that matters to us in the final analysis touches the physical world, and we live in the physical world. And so I think the market opportunity for bringing software computing intelligence into an industry that largely historically was mechanical, electrical — that's a big opportunity.

我的意思是，看看 Elon 对汽车行业做了什么。他把它变成了本质上的计算行业、技术行业。所以我认为每一家汽车公司都会成为技术公司。

I mean, look at obviously what Elon has done to the automotive industry. Turned it into essentially a computing industry, a technology industry. And so I think every car company is going to be a technology company.

嗯。

Mhm.

每一家航运和物流公司都会成为技术公司。我为 FedEx 感到非常兴奋。他们显然会成为一家 AI 公司。他们显然会成为一家技术公司。沃尔玛显然会成为一家技术公司、一家 AI 公司。所以我为他们感到非常兴奋。

Every shipping and logistics company is going to be a technology company. I'm super excited for FedEx. They're obviously going to be an AI company. They're obviously going to be a technology company. Walmart obviously going to be a technology company, an AI company. So super excited for them.

是的，整个电信行业将再次被重塑为技术行业。很久以前它们是技术行业，后来变成了服务行业。它们将再次成为技术行业。我很高兴能与 AT&T、Verizon 以及欧洲和世界各地的公司合作，帮助它们重塑自我。

Yeah, the entire telecommunications industry will be reinvented again into a technology industry. They were a technology industry a long time ago, then became a services industry. They're going to be a technology industry again. And I'm delighted to work with AT&T and Verizon and in Europe and around the world as they reinvent themselves.

所以你看看每一个触及物理世界的行业，都有用 AI 重塑它们的绝佳想法。我刚去过华盛顿。

And so you take a look at every one of these industries that touch the physical world, and there's a great idea for reinventing them with AI. I was just in Washington.

真奇怪，你知道，人们能把我们放进芯片里。

So weird, you know, people can put us in chips.

是的。所以我们实际上涉足仓储和物流。

Yes. And so we're actually in warehouse and logistics.

是的。

Yes.

我们涉足医疗。我们涉足汽车。我们涉足电信。地球上每一个行业。我们还涉足生命科学，还有所有这些——娱乐。我的意思是，显然很长一段时间里，Nvidia 在很大程度上是一家娱乐公司。所以这些垂直领域中的每一个，Nvidia 内部都有庞大的技术栈，我们深深扎根于生态系统中，在那里有极好的合作伙伴关系。所以总有一天，人们可能不再把我们归类为芯片公司了。

We're in healthcare. We're in automotive. We're in tele. Every industry on the planet. And so we're in life sciences, and so all these — entertainment. I mean obviously for a long time Nvidia was really vastly an entertainment company. And so each one of these verticals, Nvidia has a massive stack that's inside of it, and we're very deep in the ecosystem and we have excellent partnerships there. And so one of these days people might not categorize us as a chip company anymore.

这个打法大概就是把汽车行业的打法，插入到这些行业中的每一个。

The playbook would be to take the automotive playbook, I guess, and insert it into each one of these industries.

是的，那是最终的贡献。但在每一个领域，构建一个物理 AI 平台都需要三台计算机。你需要一台计算机来创造大脑。那就是训练系统。但非常重要的是，这也是你会看到实验室也这么做的地方，他们需要构建一个互联网的数字孪生，这样就不必把他们的 AI 智能体放到互联网上去学习如何使用它。

Yeah, that's the final contribution. But in each one of these, it takes three computers to build a physical AI platform. And so you need the computer to create the brain. That's the training system. But very importantly, and this is where you're going to see the labs also do this, they need to build a digital twin of the internet so that they don't have to put their AI agents out into the internet to learn how to use it.

所以就像我们创建 Omniverse 作为仓库、建筑、城市和物理世界的数字孪生一样，我们把机器人放进数字孪生里。数字孪生学会如何成为机器人，然后我们把它部署到最后一台计算机里，也就是例如机器人、汽车、手术器械或任何东西内部的那台计算机。所以需要三台计算机和三个完整的技术栈，我们为每一个都构建完整的技术栈。我们是世界上唯一这样做的公司。

And so just like we created Omniverse as the digital twin of warehouses and buildings and cities and the physical world, we put our robot into the digital twin. The digital twin learns how to be a robot, and then we deploy it into the last computer, which is the computer that's inside, for example, a robot or a car or a surgical instrument or whatever it is. And so it takes three computers and three complete stacks, and we build complete stacks for each one of them. We're the only company in the world that does that.

然后如果你是一家想要端到端构建它的公司，你可以用两台、三台或我们的一台计算机。我们不在乎。所以每一台计算机都有自己的软件栈。我们在所有垂直领域都这样做。

And then if you're a company that wants to build it end to end, you could use two or three or one of our computers. We don't care. And so each one of the computers has its own software stack. We do the same across all of our verticals.

但当然，这些能力是双刃剑，这就是为什么我们听到很多人在谈论安全。你一直对末日论叙事持高度批评态度。只有负责任地构建技术，你才能达到那里。

But of course these capabilities cut both ways, which is why we hear a lot of people talking about safety. You've been highly critical of the doomer narrative. You can only get there if you build the technology responsibly.

没错。

That's right.

所以第一件事就是负责任地构建技术。

And so the first thing is build the technology responsibly.

对。注意“构建”这个词。这是一个工程学的词，对吧？它关乎技术。它关乎工程。工程师、科学家、研究人员。我们必须安全地、负责任地构建这项技术。问题是你能否做到？答案是绝对可以。因为如果你做不到，那你根本就不应该构建它。

Right. Notice the word build. It's an engineering word, right? It's about technology. It's about engineering. Engineers, scientists, researchers. We have to build this technology safely, responsibly. And the question is, can you? And the answer is absolutely yes. Because if you can't, then you shouldn't build it at all.

当然。

Of course.

所以我们要么相信我们能做到，要么就不应该构建它。好。我相信我们能做到。那么在这里，让我和你一起推理一下。第一部分是，他们说——有些人说 AI 是不可控的，因为不可能完美对齐。给所有听众解释一下，对齐是这个意思。你告诉某人，我要你做这件事。所以任务是清晰的。

And so we either have to believe we can or we shouldn't build it. Okay. And so I believe we can. And so here, let me just reason about it with you. So the first part of it is they say — some people say that AI is uncontrollable because it's impossible to align perfectly. Alignment means this for all the audience. You tell somebody, I want you to do this. And so the mission is clear.

任务是清晰的，而且对你来说是清晰的，因为你用英语说了出来。

The mission is clear and it's clear to you because you said it in English.

好。所以可能是，我要你解这个二次方程，对吧？所以任务是解这个二次方程。对齐是你怎么做？有很多方法可以做。例如，你背下来了，对吧？

Okay. And so it could be, I want you to solve this quadratic equation, right? And so the mission is to solve the quadratic equation. The alignment is how do you do it? There are many ways to do it. For example, you memorized it, right?

好。

Okay.

所以答案是，你知道，随便是什么。或者你在教科书里看到了答案，所以你查了一下，对吧？好。你参考了参考资料，或者你走到一个抽屉前，你知道那个抽屉里有所有答案。你打开抽屉，挑出答案。好。有些人管那叫作弊。我管那叫打开抽屉取出答案。

And so the answer is, you know, whatever it is. Or you saw the answer in a textbook, so you looked it up, right? Okay. You referred to a reference, or you went to a drawer and you know that drawer had all the answers. You open a drawer, you picked out the answer. Okay. Some people call that cheating. I call that opening the drawer and retrieving the answer.

好吗？这就是计算机所做的。你可以用很多不同的方法解决那个问题。好吗？现在，当然，那个问题非常明确。这是另一个。在生活中取得成功。这个任务说起来容易，事实上我们一直对彼此这么说。你知道，我们对孩子这么说。我们对朋友说，让我们尽力为社会做贡献。

Okay? And that's what computers do. There's a lot of different ways you could solve that problem. Okay? Now, of course, that one is very explicit. Here's another one. Be successful in life. The mission is easy to say and in fact we say it to each other all the time. You know we say it to our children. We say to our friends, let's do our best to contribute to society.

现在的问题是，有太多方式了。你怎么合理地统一每个人的文化和背景？有些人认为贡献社会意味着慷慨、善良。有些人可能认为贡献社会意味着赚很多钱然后做慈善。有些人可能说等等等等。难道不是吗？

Now the question is, there are so many ways. How do you reasonably align everybody's culture and background? Some people think that contributing to society means being generous, being kind. Some people might mean that contributing to society means making a lot of money and then being philanthropic. Some people might say so on and so forth. Isn't that right?

是的。

Yes.

好的。有些人可能只是认为，仅仅建立一家公司，或者独自做某件事，就足以贡献社会。创作一件美丽的艺术品，或者所有这些都对。

Okay. And some people might just think that building a company alone, or doing something alone, is a sufficient contribution to society. Creating a beautiful piece of art, or all of it is true.

正确。

Correct.

所以问题是，什么是对齐？这取决于你的背景、你的文化。也许与你住在哪里、你来自哪里有关。有无数种方式。

And so the question is, what's the alignment? And it just depends on your background, your culture. Maybe it has something to do with where you live, where you came from. There's just an infinite number of ways.

所以我的观点是：对齐将永远是一个问题。

And so my point is this: alignment is going to be a forever problem.

是的。

Yes.

如果我们依赖完美的对齐来确保安全，那么我们永远也达不到。所以答案是：我们对待婴儿的方式，我们抚养孩子的方式，你知道，我们养宠物或其他什么，员工，好吗？你从遏制开始。你给他们权利，使用工具的机会，获取信息的机会，获取任何责任、能力的机会，资金、资源。你根据你要求他们做的工作给予他们访问权限。你从零信任开始。当你雇佣某人时，你从零信任开始。当你有一个一岁的婴儿时，你把他们放在游戏围栏里。你假设零信任。

If we depend on perfect alignment in order to ensure safety, then we're never going to get there. And so the answer is this: the way we treat, the way we think about babies and when we're raising kids, and you know, we have pets or whatever it is, employees, okay? You start with containment. You give them the rights, the access to tools, the access to information, the access to whatever responsibility, capabilities that they have, funds, resources. You give them access based on the job that you ask them to do. You start out with zero trust. When you hire somebody, you start out with zero trust. When you have a baby that's one year old, you put them in a playpen. You assume zero trust.

所以你零信任。你给他们最低限度的权利。

And so you zero trust. You give them minimum rights.

嗯。随着你了解更多，你扩大那个边界。所以每当我们有一个新的人工智能时，我们必须假设同样的事情。每当你测试人工智能时，零信任，最小边界，最小权利，你随着测试扩大它并给自己信心。所以遏制，非常严格的遏制，对任何自主系统都非常重要。第二件事，监控，监控必须是带外的。原因是，如果你告诉一个聪明人，嘿，听着，我会坐在这里看着你工作，确保你做得对。

Mhm. And you expand that boundary as you learn more. And so whenever we have a new AI, we have to assume the same thing. Whenever you're testing an AI, zero trust, minimum boundaries, minimum rights, and you expand it as you test and give yourself confidence. So containment, very tight containment, is really important to any autonomous system. Second thing you do, monitoring, and the monitoring has to be out of band. And the reason for that is because if you tell an intelligent person, hey, listen, I'm going to sit here and watch you do your work to make sure you're doing it right.

对。

Right.

好的。我会坐在这里看着你，确保你正确地服务客户，并且你工作足够努力。好的。在那个自主系统中，它的行为会与你不看时不同。

Okay. And I'm going to sit here and watch you make sure that you're serving customers properly and that you're working hard enough. Okay. That in that autonomous system is going to behave differently than when you're not watching.

就像布拉德在不在房间里一样。

It's like Brad's in the room with me or not.

正是。正是。正是。好的。是的。所以监控必须是带外的，并且必须有能力观察自主系统是否在做与政策不一致或与对齐不一致的事情。并且你有能力升级或关闭它，对吧？

Exactly. Exactly. Exactly. Okay. Yeah. And so the monitoring has to be out of band and it has to have the ability to observe that the autonomous system is doing something that was inconsistent with the policy or inconsistent with alignment. And you have the ability to either escalate or shut it down, right?

直接关掉。

Just turn it off.

所以这些能力必须存在。那是工程，那是技术。而且它是可能的。我们最近宣布了一个叫 Open Shelf 的东西。你知道，现在有超过 100 家公司加入我们。世界各地有大量的活动，因为每个人都想要同样的东西。

And so those kind of capabilities have to exist. That's engineering, that's technology. And it's possible. And we announced it recently called Open Shelf. We have, you know, over 100 companies now joining us. Lots and lots of activity around the world because everybody wants the same thing.

从最高层开始，你知道，当我在白宫环顾桌子时，是的，嗯，桌子周围的每个人，对吧，他们的父母，他们有孩子。我认为这个国家有一种观点，认为一些超级智能的领导者只是在 yolo，他们并不真正关心。他们鲁莽。他们只想抓住所有的钱。然而当我环顾他们的桌子时，那张桌子，我在那里没有看到一个人想要鲁莽，想要 yolo。事实上，那里的每个人都想在快速和安全之间找到正确的平衡。那是总统的指示。我们稍微辩论了那意味着什么，对吧？有不同的实施版本。但是，你知道，你会对害怕我们有一个鲁莽的总统或鲁莽的领导人的美国公众说什么？你看到任何证据吗？你有孩子。你担心你通过超级智能带来的未来吗？

Starting at the highest level, you know, when I looked around the table at the White House, yeah, um, everybody around that table, right, their parents, they have kids. I think there's this view in the country that there are some leaders of superintelligence that are just yoloing, that don't really care. They're being reckless. They just want to grab all the money. And yet when I looked around their table, that table, there wasn't a person in there that I saw that wants to be reckless, that wants to yolo. In fact, everybody there wanted to find the right balance between fast and safe. That was the president's directive. And we debated a little bit what that meant, right? There are different versions of how to implement that. But, you know, what would you say to the American public who fears that we have a reckless president or reckless leaders? Did you see any evidence of that? You have kids. Do you worry about the future that you're bringing into existence with superintelligence?

安全和安保技术，对吧，是人工智能的一部分，对吧？有一种奇怪的表述，认为人工智能技术和安全与安保是正交的。

Safety and security technology, right, are part of AI, right? There's some weird formulation that AI technology and safety and security are orthogonal.

是的。

Yeah.

这有点像安全气囊和刹车系统，对吧？还有雷达、摄像头、嗯、ABS、车道保持，所有这些技术不知何故独立于汽车。你们明白我在说什么吗？对汽车来说唯一重要的是发动机，对吧？事实上，整辆车都是汽车技术的一部分。当我说我们需要加速人工智能技术时，我的意思是我们需要加速安保技术。我们必须加速安全技术。它们都必须加速。

It's kind of like the airbags and the braking system, right? And the radars, the cameras, um, the ABS, the lane-keeping, all of that technology is somehow independent of the car. Do you guys understand what I'm saying? The only thing that matters to the car is the engine, right? In fact, the whole car is part of the car's technology. When I say we need to accelerate AI's technology, what I mean by that is we need to accelerate the security technology. We have to accelerate the safety technology. They all have to be accelerated.

对吧？如果你现在问我，我想每个人都会说，“让我们加速安保。让我们加速安全技术。”没有人说，“不，让我们放慢那个，对吧？我们应该加速那个。”这都是同一件事的一部分。你不能将能力、功能与安全和安保分开。它们是一体的。

Right? If you ask me right now, I think everybody would say, "Let's accelerate security. Let's accelerate safety technology." Nobody says, "No, let's pace that, right? We should accelerate that." It's all part of the same thing. You can't separate capability, functionality from safety and security. They go as one.

当你提供服务时，没有人会说，你知道，好，好，你知道，我真的很喜欢你的服务的功能，但是天哪，它，你知道，正在造成严重破坏。所以我认为这都是一件事，我们把它当作一件事来对待，行业必须把它当作一件事来对待，当我说加速时，我们必须意味着加速整个包，现在包括安全。

And when you deliver the service, nobody's going to say, you know, good, good, you know, I really like the functionality of your service, but boy, is it, you know, wrecking havoc. And so I think it's all one thing and we treat it as all one thing and the industry has to treat it as one thing and when we say accelerate we have to mean accelerating the entire package and right now including the safety.

包括尤其是安全。

Including especially the safety.

你提到了节奏，嗯，尤其是安全。

You mentioned pacing um especially safety.

你知道，有些人读了 Dario 的《Pacing the Frontier》，认为那意味着放慢速度。我认为房间里一个有趣的时刻是总统叫他的新好朋友 Dario。他用那种方式称呼他，他说，你知道，向我解释一下，你知道，你的快速和安全版本。Dario 一开始说，总统先生，你说谁赢得人工智能谁就赢，对吧？他说我们都在加速，他们也在加速，他们想快速前进，他们想，你知道，赢得与中国的竞赛，但也承认，一路上，正如你刚才指出的，我们必须确保安全系统也在加速，也在保持同步。

You know, some people read Pacing the Frontier by Dario and thought that meant slowing down. And I thought a funny moment in the room was when the president called on his new best friend Dario. He referred to him in that way and he said, you know, explain to me, you know, your version of fast and safe. And Dario started off by saying, Mr. President, you said whoever wins AI wins, right? And he said that we're all speeding up, that they're speeding up, that they want to go fast, that they want to, you know, win the race with China, but also acknowledge that along the way, just to your point you just made, we have to make sure the safety systems are also speeding up and also keeping pace.

所以我觉得那是一个重要的时刻，因为我认为人们把他定位成好像反对这件事，对吧，但在我看来你们其实非常一致。你和他、Greg 以及其他前沿实验室的人一起去内阁会议室签了这份协议。那个房间里有没有关于意见分歧的议论，还是说大家——显然你们把它签下来了。有没有人需要被强行按住，还是大家都是自愿的？

So, I thought that was an important moment where you had, you know, because I think people have positioned him as though he's in opposition, right, to this, but it seemed to me that you were all very much on the same page. You went to the cabinet room with him and Greg and the other frontier labs to sign this agreement. Was there any chatter in that room about differences of opinion, or was everybody—obviously you got it signed. Did anybody have to be put in a headlock, or did everybody go willingly?

不，每个人都非常有建设性、非常高效。我认为无一例外，每个人都致力于安全、可靠地推进这个行业。原因在于，如果不安全、不可靠，功能就永远没有机会发声，能力也永远没有机会闪耀。所以我们知道，任何时候的首要任务都是安全和可靠。

No, everybody was very constructive, very productive. And I think that, without exception, everybody is dedicated to advancing the industry safely and securely. And the reason for that is because if it's not safe, if it's not secure, the functionality will never get a chance to speak. And so the capability will never get a chance to shine. And so we know that the priorities are at all times safety and security.

当然，要认识到这个行业刚刚经历了一次转变，对吧？就在六个月前——这项技术还几乎不能用。

Of course, recognizing that this industry came through a transition just recently, right? It wasn't but six months ago—this technology barely worked.

正是如此。所以如果没有功能和能力，把它做得安全可靠又有什么意义？根本没人会在意。所以显然，AI 的第一要务是证明它能做一些有用的事，对吧？而它在六个月前真的达到了那个点。所以现在它真的想要被部署，被广泛部署。因此毫无疑问，整个行业现在都有一种战略上的紧迫要求，一种极其强烈的紧迫感，要确保我们部署的任何东西都是安全可靠的。所以我认为每个人都完全认同：能力、功能、安全、可靠——它们都处在完全相同的位置。

Exactly. And so without functionality and capability, what's the point in making it safe and secure? Nobody even cares about it. And so obviously the first job of AI is to demonstrate it could do something useful, right? And it really reached that point six months ago. And so now it really wants to be deployed and broadly deployed. And so there's no question that the entire industry now has a strategic imperative, an incredible imperative and great urgency, to make sure that whatever we deploy is safe and secure. And so I think everybody is completely convinced: capability, functionality, safety, security—they're all at exactly the same place.

头号优先事项。

Priority number one.

所以问题在于“赢”。我会重新定义“赢”——不是科技行业赢，不只是 AI 行业赢。美国赢意味着美国每一个行业都赢。医疗。

And so the question is winning. I would reframe winning—not of the technology industry winning, not of just the AI industry winning. America winning means every industry in America winning. Healthcare.

没错。没有什么比这更让我高兴的了。

That's right. Nothing would give me more joy.

美国赢的定义就是当我们谈到礼来赢、沃尔玛赢、联邦快递赢，每一个都赢，波音赢、美国军队赢，

The definition of America winning is when we talk about Lilly winning, Walmart winning, FedEx winning, every single one, Boeing winning, the US military winning,

每个公民都赢，

every citizen winning,

每个公民都赢。还有我们的大学赢、我们的高中赢。这才是赢的定义。如果只是技术赢了，那根本不算赢。

every citizen winning. And so our universities winning, our high schools winning. That's the definition of winning. There is no winning if it's just a technology that won.

正是如此。但没人用它。

Exactly. But nobody used it.

对。最大的风险当然是——我就直说底线——如果我们不把它做得安全可靠，就不会有 AI 技术行业。好吧，这对我们来说是一个生死攸关的问题。如果我们让它成为其他所有人的生死攸关的行业，那不会发生，因为它首先会是我们的生死攸关的问题，对吧？所以话虽如此，对我们来说赢的定义——这是我总得稍微改一改的部分——

Right. What the greatest risk is, of course—I'll just say baseline—there will be no AI technology industry if we don't build it safely and securely. Okay, this is an existential issue for us. And if we make it an existential industry for everybody else, well, that's not going to happen, because it'll be an existential issue for us first, right? So having said that, the definition of winning for us—and this is the part I always have to change a little bit—

对吧？

Right?

我想让我们转变一下。

I want us to pivot.

我喜欢这个说法。

I like that.

我想让我们转变一下。当我看到 AI 赢、美国 AI 赢时，我是从高处往下看——你知道，这些天我们花很多时间在空中，我们往下看。每个州都赢，每个县都赢，每个社区都赢，每个公司都赢。每个人都赢了，因为我们能负担得起在医疗上花更多钱。我们能负担得起在教育上花更多钱。我们能负担得起在州际商业、高速公路等等，这个国家所有基础设施上花更多钱。

I want us to pivot. And when I see AI winning, America AI winning, I'm looking down from—you know, these days we spend a lot of time in the air and we're looking down. Every state's winning, every county is winning, every community is winning, every company's winning. Everybody won because we could afford to spend more on healthcare. We could afford to spend more on education. We could afford to spend more on interstate commerce, highways, etc., all of the infrastructure of the country.

我听到你说的是，赢的定义就是当我们 25 年、30 年、40 年后回头看时，经济规模需要扩大 10 倍。每个人的生活都需要更好，而所有这一切的根本和首要前提是，必须以符合人类利益的方式来做。这是我在这个问题上的最后一个问题，我知道我们还有很多要聊的。大多数人并不了解我们国家的安全框架。我说过我们拥有世界上最健全的超级智能安全框架。我们显然是地球上唯一一个真正制裁过一个模型并在今年早些时候把 Mythos 下架的国家。但我把它描述成一个三条腿的凳子。第一条腿是 Casey 和商务部对每一个新的前沿模型进行前沿模型审查。他们做所有的工作。如果有新模型要发布，他们会审查它。他们会测试它，判断它是否会在市场上造成问题。第二条腿是我们现在有了关于超级智能的白宫协议，是你帮助架构的——内部和外部审计、董事会监督以及随之而来的所有责任。第三条腿是我们现在有了由 Jay Clayton 领导的 SI force，当然还有 Emil、Scott 和 Andrew，它正在做关键的行业协调、网络安全等等。所以这就是凳子上的三条腿。外面有人——我最近听到盖茨说，‘我们做得不够。我们需要制定新的法律等等。’在我看来，这第三条腿在很短的时间内就凝聚起来了，已经很多了。作为显然处于超级智能领域最重要公司之一的领导者，你对我们要引领的那个未来感到安心吗？这是向前迈出的好一步，是我们现在需要达到的位置吗？

What I hear you saying is the definition of winning is that when we look back 25, 30, 40 years from now, the economy needs to be 10 times as large. Everybody's lives need to be better, and existential and primary to all of that is it has to be done in a way that's aligned with human interest. Which—this is the last question I have on this and I know we have a lot to get to. Most people don't understand what our security framework is in the country. I've said we have the most robust super intelligent security framework in the world. We're clearly the only country on the planet that's actually sanctioned a model and pulled it off the shelf with Mythos earlier this year. But I describe it as a three-legged stool. We have the frontier model review of every new frontier model by Casey and the Department of Commerce. So they do all their work. If a new model's coming out, they review it. They test it for whether or not they think it's going to cause a problem in the marketplace. Now we have the White House accord on super intelligence that you helped to architect—internal and external audits, board oversight and all of the liability that comes with that. And then third we have the SI force now led by Jay Clayton, of course Emil and Scott and Andrew, and that's doing critical industry coordination, cyber, etc. So these three legs on the stool. There are people out there—I heard Gates recently saying, 'We're not doing enough. We need new laws on the books and everything else.' That three-legged stool seems like a lot to me in a very short period of time that's come together. As a leader of one of the most important companies obviously in super intelligence, do you feel comfortable for that future that we're trying to usher in, that that is a good step forward, where we need to be right now?

是的。通往未来的最好方式就是迅速采取行动。如果我们争论什么是完美的、什么是先做的，我们最终会一直在争论完美，永远到不了那里。所以我喜欢这个三条腿的凳子是可执行的。它们是实实在在的。它们有真正的约束力。

Yeah. The best way to get to the future is to take action quickly. And if we argue about what is perfect versus what's first, we'll end up arguing about perfection and we'll never get there. And so I love the fact that these three-legged stool are actionable. They're tangible. They have real teeth.

正在发生。

Happening now.

它正在发生。你可以——我们现在就在做。我们现在就在采取行动。我们现在采取行动的原因是，这对我们的公司有好处。对我们的行业有好处。对我们的国家有好处。对我们的社会有好处。方方面面都有好处。如果我们不做这些事，结果就不那么理想。我们就无法去追求我们所有人都梦想过的那个乐观的未来。

It's happening right now. You could—we're doing it now. We're taking action now. And the reason we're taking action now is because it's good for our company. It's good for our industry. It's good for our country. It's good for our society. It's good all the way around. And if we don't do these things, the outcome is less desirable. We don't get to pursue that optimistic future that we all dreamt about.

如果你回顾另一个行业，我不记得有哪个行业像现在这样处于当前状态，对吧？并且现在已经制定了所有这些跨公司、跨行业、政府与行业的协调，以确保我们以负责任和乐观的方式推进这件事。

And so if you go back and look at another industry, I don't remember another industry that is as it's currently at its current state, right? And have now formulated all of these cross-company, cross-industrial, government-to-industry coordination to make sure that we advance this in a responsible and optimistic way.

是的。我只是觉得让人们理解这一点非常重要，因为当我在电视上听到两边参议员谈论这件事时，说实话，他们并不——他们必须了解我们在这里做什么。他们甚至不知道我们在做什么，对吧？比如我们正在做大量非凡的工作，而且顺便说一句，这些不是自愿的事情。不会有一个前沿模型发布时 Casey 不审查、商务部不审查，而且他们已经证明根据出口管制他们有权将其下架。

Yeah. I just think it's so important for people to understand because when I hear senators talking about it on television on both sides to be perfectly honest, like they don't—they have to learn about what we're doing here. They don't even know what we're doing, right? Like we are doing an extraordinary amount, and by the way, these aren't voluntary things. There's not going to be a frontier model that gets released that Casey is not reviewing and the Commerce Department is not reviewing, and they've demonstrated they have the authority under export controls to pull them off the shelf.

做两件事：第一，来花些时间了解实际需要做什么。问题是什么，对吧？如果你想要解决方案，你至少得理解问题是什么。对吧。

Do two things: one, come and spend some time and learn what actually has to be done. What's the problem, right? If you want a solution, you got to at least understand what the problem is. Right.

对。所以第一，来了解一下——我们有很多时间。我对这类事情有的是时间。就，你知道，来和我们谈谈。来和我们的行业谈谈，学习一下。第二件事是了解我们制定的解决方案，然后试着看看我们能否让它——如果有想法让它更好。当然，我们全都完全持开放态度，对吧？但要建设性，对吧？

Right. And so one, come and just learn about—we got plenty of time. I got all the time in the world for this kind of stuff. And just, you know, come talk to us. Come talk to our industry and learn. The second thing is learn about the solution that we put in place and then try to see if we can make that—if there are ideas to make that better. Of course, we're all completely open-minded, right? But be constructive, right?

我认为那是超级智能协议第四点的一部分——这不是静态的，这是动态的。这个小组会定期聚在一起，当技术发展时我们会根据需要更新框架。但我想现在转向算力。当然，长时间运行的智能体以数千或数百万的团队协作解决复杂任务——这一切都消耗大量更多的算力。我们的朋友、SemiAnalysis 的 Dylan 说过，明年我们将新增 43 吉瓦的新 AI 算力。为了让大家有个概念，我们会在网上分享一张图表——这比当今世界上所有 AI 算力加起来还多。所以明年我们将新增这么多。后年再新增 40 吉瓦算力。你是那个知道每一块 GPU 分配在哪里的人，对吧？所以当你坐在这里，这听起来对吗？你觉得我们会新增那么多算力吗？你觉得我们听到的一些关于许可问题、供应链瓶颈或支持本地的抗议——根据你所看到的，你觉得这个数字会与此有实质性不同吗？

I thought that was part of the fourth point on the accord on superintelligence—was this is not static, this is dynamic. This group is going to get together routinely, we're going to update the framework as required when the technology evolves. But I want to move forward now to compute. Of course, long-running agents that work in teams of thousands or millions to solve complicated tasks—it all consumes massively more compute. Dylan, our friend at SemiAnalysis, has said we're going to bring on board 43 gigawatts of new AI compute next year. Just to contextualize it, we'll share a chart with people online—that's more than all of the AI compute that exists in the world today. So we're going to bring that on board next year. The year after that, another 40 gigawatts of compute. You are the one—you know where every single GPU is allocated, right? So as you sit here, does that sound right to you? Do you think we're going to bring on that much compute? And do you think that some of the things we're hearing out there about permitting issues or bottlenecks in the supply chain or pro-local protests—do you think the number is going to be materially different than that based upon what you're seeing?

我认为需求肯定存在，而且更多。

I think the demand is definitely there and more.

是的。

Yeah.

需求肯定存在，而且更多。当然，挑战在于当你从数字世界转向物理世界时，有很多障碍。那里有很多障碍。土地许可。你必须上线发电。当然，你需要建筑工人、装配工、水管工，你知道，各种各样的——仅仅建造这些东西就创造了上百万个新工作岗位。创造上百万个新工作岗位，而他们的技能不是可替代的。你知道，他们有很多技能，尤其是那些建造这些 500 亿美元数据中心的人。你会想要一些技能出色的人，而且每一个这类建设都有一个关键部分，你需要一些非常具体的技能。所以我认为——然后当然还有正在发生的任何暂停令，我们在全国各地看到的任何问题。所以我认为这是——我的猜测是，土地、电力、外壳、建设资源将是达到那个目标的最重要障碍之一。

The demand is definitely there and more. The challenge, of course, is that there are a lot of obstacles when you move from digital to the physical. There are a lot of obstacles there. Land permitting. There's power generation that you've got to bring online. Of course, you need construction workers and outfitters and plumbers and, you know, all kinds of—created a million new jobs just building these things. Create a million new jobs, and their skills are not fungible. You know, they have a lot of skills, and particularly the ones that are building these $50 billion data centers. You're going to want some people with great skills, and there's a critical part of every single one of the buildouts of these things where you need some very specific skills. And so I think that—and then there's, of course, whatever moratoriums that are happening, whatever issues that we're seeing around the country. And so I think this is—my guess is that land, power, shell, the buildout resources will be one of the most important obstacles to reaching that.

所以我们的公司，如你所知，现在具体说我们，我们正在做几件事来缓解这个问题。第一，你知道，因为我们是唯一一家服务所有 CSP 的公司，我们服务所有 NCP。我们通过 Dell、HP、Lenovo 等拥有伟大的企业业务。而且因为我们服务每一个行业，我们横跨 AI 的整个生命周期——从你知道的当然数据处理到训练到后训练到推理到智能体式处理，以及我们加速的所有工具使用——你知道，我们的市场覆盖是最大的。世界上每一家合同制造商都在为我工作。每一个市场渠道都连接到我们的架构。所以如果某个特定的州、某个特定的社区、某个特定的数据中心遇到困难，对吧，你知道，目前我们的需求远大于供应，我们只需相应调整我们的供应链。

And so our company, as you know, now speaking specifically for us, there are several things that we're doing to mitigate that. And one, you know, because we're the only company that serves every CSP, we serve every NCP. We have great enterprise businesses through Dell and HP and Lenovo and others. And because we serve every single industry and we're across the entire life cycle of AI—everything from, you know, of course data processing to training to post-training to inference to agentic processing, and all of the tool use that we accelerate—you know, our exposure to our go-to-market is the largest of all. Every contract manufacturer in the world is working on my behalf. Every go-to-market is connected to our architecture. And so if one particular state, one particular community, one particular data center has a hard time, right, you know, our demand is so much greater than our supply at the moment, we'll just adjust our supply chain accordingly.

我们有的另一件事——我不那么喜欢这个答案——但如你所知，Nvidia 也是这个星球的基础设施。所以无论是在欧洲还是——我们现在在非洲、中亚、东南亚和拉丁美洲、中东等地都有早期发展。所以我不喜欢美国跟不上。

The other thing that we have—and I don't love this answer as much—but as you know, Nvidia is also the infrastructure for the planet. And so whether it's in Europe or—and we've got early developments now in Africa, in Central Asia, in Southeast Asia and Latin America, in the Middle East, and so on and so forth. And so I don't love that the United States can't keep up.

是的。

Yeah.

但具体到我们的业务，我们拥有如此全球化的覆盖，全世界都有建设需求。如果我们不在这里建设，你知道，我想我们都必须认识到，世界上每个国家都希望在那里建设。

But for our business specifically, we have such a global reach, and there's buildout demand all around the world. If we don't build it out here, you know, I think we just all have to recognize that every country in the world would love to have it built out there.

是的。

Yeah.

所以我认为我们的业务具体来说可能是最具韧性的。原因就是我提到的所有事情——你知道，我们覆盖 AI 的整个谱系。我们覆盖计算的整个谱系。我们在每一朵云、每一个地区。所以我们公司的韧性要大得多。但我确实认为土地、电力、外壳明年将是一个具有挑战性的部分。

And so I think that our business specifically is probably the most resilient. And the reason for that is all the things that I mentioned—you know, we cover the entire spectrum of AI. We cover the entire spectrum of computing. We're in every single cloud, every single region. And so the resilience of our company is much greater. But I do think that land, power, shell is going to be a challenging part next year.

我猜我从市场上听到的是这样一种想法：也许有 50 吉瓦已宣布的芯片正在进行或讨论中，也许有 35 到 40 吉瓦的算力或土地、电力、外壳可能上线。所以什么算力能装进那些土地、电力和外壳，以及按什么时间线，是一个悬而未决的问题。

And I guess what I hear from the street is there's this idea that maybe there's 50 gigawatts of announced chips that are underway or talked about, and maybe 35 to 40 gigawatts of compute or land, power, shell that might come online. So what compute fits into that land, power, and shell, and on what timeline, is a question up in the air.

但我想，如果我们从整体来看——Clark 刚才打断你了——当然，如你所知，Nvidia 的架构是最高效的，对吧？它每瓦能生成最多的 token。

But I guess if we think of through the entire— and just Clark just interrupt you there— and of course, as you know, Nvidia's architecture is the most productive, right? It generates the most tokens per watt.

对。所以如果全球的瓦特电力能源减少，就会把他们的选择推向最高效的版本。原因在于他们仍然想赚钱，对吧？我所有的 CEO 同行们，他们对明年的营收都有很高的期望，人们需要兑现它。他们的股东指望他们兑现。他们所有的合作伙伴都需要他们兑现。所以当电力减少时，你投入的算力质量必须是最卓越的，或者说是当今世界上你能投入的最高效的计算资源。我们是最能赚钱的。我们有趣且耐用。所以由于资源短缺，全球的各种力量正在把兴趣转向我们。我们的市场份额本来就已经很高了。但令所有人惊讶的是，不仅我们是一家大公司、市场份额很高，由于这些原因，我们的市场份额实际上还在增长。

Right. And so if the world's amount of watt power energy reduces, it shifts their choices to the most productive version. And the reason for that is that they still want to make their revenues, right? And so all of my CEO colleagues, they all have high expectations for revenues next year and people need to deliver it. Their shareholders counting on them to deliver it. All their partners need them to deliver it. And so when the amount of power reduces, the quality of the compute you put into it has to be the most extraordinary or the most productive computing resource you could put in the world today. We're the most rentable. We're funible and we're durable. And so the forces around the world because of the shortage of resources is shifting interest towards us. Our market share was already very high. But surprisingly to everybody, not only are we a large company and our market share is high, our market share is actually increasing for these reasons.

嗯。

Mhm.

所以你会看到我们从——因为我们每吉瓦提供的价值比以前多得多。我们曾经是一家 GPU 公司，然后我们成为最大、增长最快的网络公司。我们现在将成为最大的 CPU 公司之一，肯定也是增长最快的。因为我们提供的东西越来越多，我们涉足存储计算，我们涉足知识基础设施计算。现在我们通过 OpenShell 和 Sentry 涉足安全领域。所以我们投入一吉瓦的报价从大约每吉瓦 180 亿美元左右，到每吉瓦 200 多亿美元，到现在每吉瓦 400 多亿美元。我的感觉是这还会继续上升。所以我们不仅在增长份额，因为人们在转向我们的计算平台，我们还在增长份额，因为我们为全球的资本支出提供了更多价值，对世界而言这是市场份额的美元占比。

And so you're going to see us go from— and because we offer so much more per gigawatt than we used to. We used to be a GPU company, then we became the largest and fastest growing networking company. We're now going to be one of the largest CPU companies, surely the fastest growing one. And so because we're offering more and more and we're in computing for storage, we're in computing for the knowledge infrastructure. And now we're in security with OpenShell and Sentry. So the offering that we offer into one gigawatt went from, I guess, about 18 billion or so per gigawatt to 20-some billion per gigawatt to now 40 plus per gigawatt. And my sense is that that's going to keep on going up. And so not only are we growing share because people are shifting towards our computing platform, we're also growing share because we offer more to the world's capex and to the world it's a dollar share of the market.

是的。对。如果你看看，就回到六年前。

Yes. Yeah. If you look at just go back six years ago.

是的。

Yes.

我们在全球资本支出中的占比是个位数。

Our percentage of the world's capex was in single digits.

是的。

Yes.

嗯，然后我想现在在全球的云服务提供商中，我们大概占到 40%。

Um then I think right now with the world's CSPs we're probably running about 40%.

嗯。但我们也在把市场扩展到云服务提供商之外。在云服务提供商之外，无论是本地部署还是新型云，我们一半的业务，我们的资本支出占比大概是 80% 吧。

Mhm. But we're also expanding the market beyond the CSPs. And outside of the CSPs, whether it's on-prem or neo clouds, half of our business, our percentage of capex is probably what, 80%.

对。

Yeah.

所以在 40%——一半业务是 40%，一半业务是 80%。而且两者都在增长，每年翻倍。所以你可以看到，我们在全球资本支出中的占比实际上还在增长。

And so between 40— half of our business being 40%, half of our business being 80%. And both of us growing, doubling every year. And so you could see that our percentage of the world's capex is in fact even growing.

是的。嗯，Nvidia 的买家之一就是我们两年前聊过的 Elon，当时他开始建 Colossus，两年前我们看到他在孟菲斯搭起几十万块 Hopper，真是惊呆了，对吧。我记得你说过没人能像他那样做。现在他又在筹集 400 亿美元，可能要买一百万块以上的 Vera Rubin。他从 K 级跨到了 M 级。他说他想在 2027 年底前拥有 10 吉瓦。那是数千亿美元，对吧，花在 Vera Rubin 上。这是前所未有的规模扩张。你昨天又和他在一起，当时他还被授予了一枚奖章。Elon 今天在构建算力方面比那时强那么多吗？现在对你来说已经很明显了吗？SpaceX 将成为世界上最大的 AI 超大规模厂商，这是板上钉钉的事吗？

Yes. Well, one of those buyers of Nvidia is somebody we talked about two years ago, Elon, when he started building Colossus, and we were blown away a couple years ago to see him stand up a couple hundred thousand Hoppers, right, in Memphis. And I remember you saying nobody can do it like him. Now he's raising another 40 billion, going to buy maybe a million plus Vera Rubin. He's moved from the K's to the M. He says he wants to have 10 gigawatts by the end of 2027. That's hundreds of billions, right, spent on Vera Rubin. It's an unprecedented ramp at scale. You were with him again yesterday when he was also presented with a medal. Is Elon that much better at building compute today as he was then? Is it now obvious to you? Is it a foregone conclusion that SpaceX will be the largest AI hyperscaler in the world?

嗯，我们以前，你知道，在我们图表的两个部分里，Elon 以前归在企业类。我们就是这样报告数字的。现在他当然是我们云服务提供商的一部分了。

Well, we used to, you know, in the two parts of our chart, Elon used to be under Enterprise. That's the way we reported our numbers. Now he's part of our CSPs, of course.

对。对。当然。

Yeah. Yeah. Of course.

所以他们规模巨大。所以我很高兴——我喜欢和 Elon 合作。他在用我们的技术做他自己的 AI，训练模型之类的，部署模型。当然，Grok Bot 超级酷，对 Grok Bot 的需求、Grok Bot 的消耗量真的在飙升。然后当然，他还把它租出去。

And so they're massive. So I'm delighted that— I love working with Elon. He's using our technology for his own AI in training models and such, deploying models. Of course, Grok Bot is super cool and the demand for Grok Bot, the consumption for Grok Bot is really skyrocketing. And then of course, he rents it.

对。

Yeah.

所以这三件事结合起来，他在这方面相当独特。我想 Google 也这么做，对吧？那会是另一个例子。所以我认为他还有一个优势，就是他热爱计算和实体建设之间的交叉点。没错。你知道，他热爱整个这个领域，没人比他更懂能源。我们要一起探索的一个特别的新市场是把数据中心放到太空。所以我们两个工程团队正在非常紧密地合作，创造一种新型服务器、新型架构，可能在太空和地球上都非常出色。

And so the combination of these three things, he's fairly unique in doing that. I guess Google does that, right? That would be another yet another example. And so I think he's also has the advantage that he loves the intersection between computing and the physical buildouts. That's right. You know, he loves that whole area and nobody knows more about energy than he does. And one particular new market that we're going to go explore together is putting data centers out in space. And so our two engineering teams are working super closely together to create a new type of server, new type of architecture that might be super good, you know, in space and on Earth.

嗯，我们上周看到 Starship 发射了，你知道，它进行了首次轨道会合，还发射了那些 V3 卫星，真的让人浮想联翩，一度把 Twitter 都搞崩了。人们就那样着迷地看着它们发射升空。我迫不及待想成为载荷。我是说，不是人。你觉得——那会——你越来越有信心我们会在太空拥有大量算力吗，你知道，通过 Starship，我们姑且说 2028、2029 年？我是说，Elon 放出了一些大数字，但你觉得我们能在那个时间框架内实现太空中的实质性推理吗？

Well, we saw that Starship launched last week, you know, had its first orbital rendezvous and was launching those V3 sats that literally captured the imagination and kind of broke Twitter for a moment. Just people mesmerized by watching these launch into space. I can't wait to be a payload. I mean, not person. Do you think— does that get— do you have increasing confidence that we will have a material amount of compute in space, you know, via Starship by let's call it 2028, 2029? I mean, Elon's put some big numbers out there, but do you think that we can get to material inference in space in that time frame?

只有 Nvidia 的算力。让我告诉你为什么。这不仅仅是技术问题。Nvidia 的单位能耗性能是世界最好的。所以我们称之为每瓦性能。但最重要的是这一点：Nvidia 算力的生产率太好了，意思是投入资本的回报太好了，以至于你可以把它放到太空，而发射成本只占其中很小一部分。

Only Nvidia compute. Let me tell you why. It's not just about the technology. Nvidia's performance per unit energy is the world's best. And so we call it performance per watt. But the most important thing is this: the productivity of Nvidia's compute is so good, meaning the return on investment invested capital was so good, that you could put it out in space and the launch cost is a small percentage of it.

好。

Okay.

发射成本并非无关紧要。所以你不会把低成本的计算放上去。

The launch cost is not inconsequential. And so you're not going to put low-cost computing out there.

没错。你会把高价值的计算放上去。

Correct. You're going to put high value computing out there.

所以我们就是你能想象到的最高价值的计算系统，计算史上最高的计算系统。

And so we're the highest value computing system you could imagine, the highest computing system in the history of computing.

所以我们现在其实非常适合发射到太空。不管发射成本如何，回报率依然会非常出色。那么到 29 年，我们能在太空获得 500 兆瓦吗？

So we're the perfect fit really to launch into space right now. Irrespective of the launch cost, it'll still be excellent ROI. So do we get 500 megawatts in space by '29?

这个我会让他来告诉你，不过我觉得——

Well, I'll let him tell you, but I think—

不可能。

Impossible.

你知道 Elon 的——现在一年发射 100 次？我不确定具体数字，一天好几次。

You know what's Elon's—what, launching 100 launches a year right now? I don't know what exactly it is, multiple a day.

对。没错。

Yeah. Right.

所以你可以把 Vera Rubin 放上去，这就说得通了。需求就在那里。他的投资回报非常好。

And so you could put Vera Rubin in there and it'll justify it. The demand is there. His return on investment is so good on it.

所以生产率非常出色。

And so the productivity is excellent.

它是可替代的，意思是说你把一颗卫星放到太空。如果没有客户来租它，它就闲置着，你肯定不想要那样。所以 Nvidia 的算力将完全可替代，会有非常非常多的租户。你知道，我们和世界上每一家闭源模型公司合作。我们运行世界上每一个开源模型。我们有数千家初创公司在和我们合作。还有遍布各个行业的企业。所以这些算力完全可替代，而且很耐用。

It's fungible, meaning you put a satellite out in space. If you don't have a customer to go rent it, it sits idle and you certainly don't want that. And so Nvidia compute is going to be fully fungible and there'll be lots and lots of renters. You know, we work with every single closed model company in the world. We run every single open model in the world. We have thousands of startups working with us. And enterprises around in every single industry. And so the computing is completely fungible and it's durable.

对。

Yeah.

你知道，它会待在上面，我们会持续跟进，原因是我们一直在升级 CUDA 软件。你知道，我们是世界上唯一一家你八年前付了钱，我们至今还在为你更新软件、更新算法的计算机硬件公司，

You know, it'll sit up there and we'll keep up and the reason for that is because we keep CUDA upgrading the software. You know, we're the only computer hardware company in the world that you pay us eight years ago and we're still updating your software, new algorithms,

对吧？

Right?

你知道，所以只要卫星还在上面，我们就会更新软件，所以它会非常耐用。所以我认为这三种特性的结合，

You know, and so we'll update the software for as long as the satellites up there and so it's going to be really durable. And so I think the combination of these three traits,

是的。

Yes.

再加上这项技术，使它成为 SpaceX 的完美载荷。

along with the technology makes it the perfect payload for SpaceX.

建设所有这些资本开支要花上万亿美元。你们动用了自己的资产负债表，帮助 Neoclouds 这样的公司把它建起来。要让金融界的人算得过账，他们需要看到回报，对吧？你提到有些公司看到了一年回本，对吧？第一年 500 亿美元，也许还有 SpaceX 的一些现货定价。但如果我看总的增量采购量，也就是所有实验室今年的营收，把两家前沿实验室加起来——哦对，你先吃点东西——再加上一些推理实验室，大约是 2000 亿美元。对吧？我们需要看到明年翻倍到 4000 亿。后年再翻倍到 8000 亿，对吧？这样才能接近上万亿美元资本开支的回报区间。今年 Claude Code 带来了巨大的解锁，让 Anthropic 的营收增长了 10 倍。你有多大信心，这些实验室能继续每年把营收增长 2 到 3 倍，从而让我们获得建设上万亿美元资本开支所需的回报？我的意思是，你在投资这些公司。你在这些公司有大笔投资，对吧？你觉得这些公司能持续这种增长水平吗？

It cost trillions of dollars to build all this capex. You guys have leveraged your own balance sheet to help folks like the Neoclouds build that out. In order for finance people to get it to pencil out, they need to see payback, right? You talked about some companies seeing one-year payback, right? $50 billion in the first year, some spot pricing out of SpaceX perhaps. But if I look at the total incremental offtake, right, the revenue of all the labs this year, so if I add up the two frontier labs—oh yeah, get yourself a snack—and we add up some of the inference labs, it's about $200 billion. Okay? We need to see that double next year to 400 billion. We need to see it double the year after that to 800 billion, right? In order to get in the zip code of the payback on trillions of dollars of capex. We had this massive unlock this year with Claude Code that led to a 10x increase in Anthropic's revenues. What's your confidence that the labs can continue to 2 to 3x their revenues every year in order that we get the payback that's going to be required to build out trillions of dollars of capex? I mean, you're investing in these companies. You have big investments in these companies, right? Do you see that level of growth continuing in these companies?

哦，是的。人们误解了正在发生的事情。首先，特别是——我的意思是，Nvidia 占今天全球 AI 基础设施营收份额的大约 80%。所以我对计算这个宇宙以及 Nvidia 计算的描述在这里很重要。基本上真的——可以说是三个板块，三个大板块，可能略多于三个，但让我就分成三大板块。我们计算的第一个板块就是基础计算。你用它做数据处理、分子动力学、物理、模拟、聚变模拟、裂变模拟，你能想到的，对吧？量子化学、量子物理，你知道，所有这些东西，

Oh, yeah. People misunderstand what's going on. So first of all, particularly—I mean Nvidia represents call it 80% of the revenue share of AI infrastructure in the world today. And so my description of the universe of computing and Nvidia computing matters here. There really basically—call it three segments, three large segments, is probably slightly more than that, but let me just categorize in three large segments. The first segment of our computing is just baseline computing. Just you use it for data processing, molecular dynamics, physics, simulation, fusion simulation, fission simulation, you name it, right? Quantum chemistry, quantum physics, you know, and so everything,

对吧？

Right?

好。模拟和数据处理的世界，更不用说计算机图形，所有这些处理都在我们的 GPU 上完成。如果你假设完全没有 AI，只有经典机器学习，Visa、Mastercard 上的 XGBoost 用于欺诈检测，推荐系统来决定给你推荐什么新闻，所有这些加起来，

Okay. The world of simulation and data processing, not to mention computer graphics, and so all of that processing done on our GPUs. If you just kind of said there was no AI at all, just classical machine learning, XGBoost at Visa, Mastercard, for fraud detection, recommender systems to figure out what news to recommend to you, just all of that stuff combined,

经典计算，不是 AI，不是深度学习，不是智能体系统，你知道，大约占我们业务的三分之一。

Classical computing, not AI, not deep learning, not agentic systems, you know, call it a third of our business.

对。

Yeah.

而且那块业务在增长，你知道，大约每年 30%、40%。

And that business is growing, you know, call it 30%, 40% per year.

对。

Yeah.

好。就算没有任何 AI 实验室，Nvidia 也不是零，对吧？我们是一家相当大的公司，你知道。所以这是第一点。

Okay. In the absence of any AI labs, Nvidia is not zero, right? We're quite a large company, you know. So that's number one.

第二部分是 AI 实验室。人们看到两家，但我们和每一家都合作，

The second part is the AI labs. People see two, but we work with every one of them,

对吧？

Right?

MetaMuse 做得非常棒，对吧？

And MetaMuse is doing fantastic, right?

不可思议。

Incredible.

而且它几乎全是 Nvidia，你知道，去年还是零，

And it's vastly Nvidia and you know, there were zero last year,

对吧？

Right?

你知道，或者再往前一年只有推荐系统，现在他们靠 MetaMuse 大杀四方，当然还有 Grok，

You know, or a year before that it was just recommender systems and now they're killing it with MetaMuse and, of course, Grok

三年前还不存在。

Didn't exist three years ago.

呃，我们以前不跑 Gemini，现在跑 Gemini 了。从今年开始，我们开始和 Anthropic 合作。所以如果你看我们的算力，人们看到几家实验室，但我看到的更多。我看到五六家实验室。对吧？所以这些实验室都做得很好。

Uh, we didn't run Gemini before, we run Gemini now. And starting this year, we started working with Anthropic. And so if you look at our compute, people see a couple of labs, but I see more than that. I see five, six labs. Okay? And so these labs are all doing great.

没错。

Correct.

如果你说那代表我们业务不到三分之一，就算四分之一吧，

And if you told me that represents not quite one-third of our business and let's say a quarter of our compute,

太好了。

Fantastic.

好。那就说那是四分之一。所以当人们看着 Nvidia 的 80 分，我们现在运行着世界的基础设施，而我刚说其中 25%、也许 30% 是前沿实验室，

Okay. So let's say that's a quarter. So when people are looking at 80 Nvidia and we now run the infrastructure of the world and I just said 25% of it, maybe 30% of it is the frontier labs,

对吧？

Right?

你知道，大约三分之一就是一般处理，

You know, call it a third is just processing in general,

对。

Right.

还有另外三分之一，这部分才是真正精彩的。

And another third, this is the part that's really great.

因为我们正处在一个新的计算平台迁移的开端，大量新的初创公司正在涌现。

Because we're at the beginning of a new computing platform shift, just a vast number of new startups are coming out.

好多伟大的公司，其中一些当然在做消费级的东西，你知道，有一家做得特别好。Noah 做得很好。

So many great companies and some of them are of course doing consumer stuff, you know, one of them really doing great. Noah's doing great.

我正想提这个。

I was just going to mention that.

到各种物理 AI、生物 AI、材料科学 AI，Jeff Bezos 的新公司 Prometheus，他们做得非常棒。所以横跨各个行业的公司数量，

To all kinds of physical AI, biological AI, material sciences AI, Jeff Bezos's new company, Prometheus, they're doing fantastic. And so the number of companies across the spectrum of industries

已经创立的，有数千家。它们做什么？它们租 Nvidia 的算力。

That have been started, thousands of them. And what do they do? They rent Nvidia compute.

所以它们是租户。是的。所以我们业务的三分之一，有点像下一代的云计算，

So they're renters. Yes. And so what is a third of our business is kind of cloud computing for the next generation,

对吧？

Right?

有点像今天的 AWS，你知道，对吧？

It's kind of like AWS for today, you know, right?

所以它不只是未来的 AWS，

And so it's just not AWS for the future,

对吧？

Right?

而这总共大约占我们业务的三分之一，你知道。

And that represents about a third of our business altogether, you know.

我的感觉是，我们正处在这场新计算架构建设的早期阶段。只看到几家公司，就以为基础设施全靠它们出资，这是一个很大的错误。

My sense is that we're in the early innings of the buildout of this new computing architecture. And it's a real big mistake to recognize just a couple of companies and think that they were funding the infrastructure.

有一个观察：开放模型——我一直是开放模型的坚定倡导者。原因在于开放模型打开了新的机会。如果你看看所有这些初创公司，80% 都在用开放模型。

One observation: open models — I've been a great advocate of open models. The reason is because open models open up new opportunities. If you look at all these startup companies, 80% of them use open models.

对。

Right.

他们不是用它去和前沿实验室竞争，而是用它去别处解锁新的机会。可能是物理 AI，可能是制造业，可能是医疗健康。他们在做完全不同的事情。

And they're not using it to compete with the frontier labs. They're using it to unlock a new opportunity somewhere else. It could be in physical AI, could be manufacturing, it could be healthcare. They're doing something completely different.

他们拿这些开放模型——其中很多人用的是 Neotron——然后把这些模型定制成超级智能。

And they're taking these open models — and many of them taking Neotron — and these models are then customized into super intelligence.

嗯。

Yeah.

所以可以说，我认为那些通用的闭源实验室会走向 AGI（通用人工智能），而且会变得超级好。

So if you will, I kind of see the general closed labs will be artificial general intelligence and it gets super good.

嗯。

Yeah.

然后你会有所有这些垂直公司，专注于某个特定领域，拥有非常专有的数据。他们教模型做非常具体的事，从而在各自领域实现超级智能。

And then you have all these vertical companies that are focused on a particular domain and they have very proprietary data. They teach it to do something very specific and they achieve super intelligence for their domain.

对吧？所以你不认为——再回到你那套算力数学，我想很多人用它来预测实验室未来的营收。比如你说，嘿，如果你建起 1 吉瓦的这种算力，另一端就能产出 400 亿美元。但如果我们用之前聊到的数字——明年将有 40 吉瓦算力上线——如果另一端按 400 亿算，那从这座工厂另一端出来的营收就是 1.6 万亿。当然，事情不是这么运作的，对吧？归根结底，你得看需求曲线和扩散速度，才能决定工厂另一端会产出多少营收。我听到很多说法是，人们很紧张。他们看到这一年里实验室的营收可能略有放缓，就担心某个时点会出现所谓的泡沫破裂——某一时刻上线的供给超过需求——AI 更像是阶跃函数，而不是一条直直向上的抛物线。我想问你的问题是：你是否担心会出现某个时刻，所有人对为这轮建设融资都变得紧张，然后我们就建不成了？

Right? So you don't see — so again, if we go to your compute math, which I think a lot of people are using to forecast the future revenues of the labs, for example, you said, hey, if you stand up a gigawatt of this compute you can get $40 billion out the other end. But if we took the number we talked about earlier — 40 gigawatts of compute coming online next year — and if you took 40 billion out the other end, that would be 1.6 trillion of revenues coming out the other end of that factory. But of course, it doesn't work that way, right? At the end of the day, you've got to look at the demand curve and the rate of diffusion to determine the revenue that's going to come out the other end of the factory. And what I hear a lot is that people are nervous. They've seen over the course of the year maybe the labs' revenue slowing down a little bit, and they're nervous that at some point you're going to have the proverbial bubble popping, right, where we have more supply coming online than we have demand at a moment in time — that AI is more like these step functions rather than just this parabolic line straight up. I guess the question to you is: are you worried about a moment in time where everybody gets nervous about financing the buildout, and then we don't get the buildout?

你知道，所有这些不同的事情，你都得把它拆开。拆回到第一性原理。

Well, you know, all these different things, you just got to break it down. Break it down back down to first principles.

人们试图外推的那部分——1000 亿美元的 AR，或者 1250 亿，他们所在的位置——他们试图外推——

The part that people are trying to extrapolate — the hundred billion dollars of AR, or 125 billion, where they are — they're trying to extrapolate —

是一小部分——不是一小部分，它只是一部分——

is a small part — not a small part, it's just a part —

对——是整个基础设施建设的一部分，毫无疑问。

Right — of the overall infrastructure buildout, for sure.

所以问题是：人们还会继续在网上购物吗？

And so the question is: are people going to continue doing shopping online?

你知道，那是——

You know, that's —

所以那三分之一建设，我们就叫它三分之一。你说它是你业务的三分之一。那部分稳拿。

So that third of the buildout, let's call that a third. You say it's a third of your business. Take that to the bank.

嗯。

Yeah.

对。这甚至没什么可争的。所以三分之一——

Right. Like that's not even a debate. So a third of the —

你知道，我几年前说过，等我们走出这个十年，基础设施建设大概会达到每年 3 到 4 万亿。

Say, you know, I said a couple years ago that as we exit this decade, infrastructure buildout will probably be somewhere between 3 to 4 trillion a year.

显然更接近 4 万亿而不是 3 万亿。而且是每年。

Obviously it's closer to four than it is closer to three. And per year.

到这个十年末。

By the end of the decade.

对。如果你说其中 1.5 万亿只是算力——

Yeah. And if you said that one and a half trillion of that was just compute —

对吧？对。

Right? Right.

只是算力。那为什么它会这么大？因为在 AI 开始之前，它是 5000 亿。

Just compute. Now, why is it going to be so big? Well, it was half a trillion before AI started.

对。

Right.

如果我告诉你，在大约六、七、八年里，它变成了 1.5 万亿，因为世界只是在消耗更多算力——我的意思是，这并不违背逻辑，对吧？

And if I told you that over the course of some six, seven, eight years, it became one and a half trillion because the world's just consuming more compute — I mean, that's not illogical, right?

除此之外，记住，摩尔定律已经结束了。

On top of this, remember, Moore's law is over.

对吧？

Right?

曾经有一段时间，你想要两倍的晶体管，只要等一年。再想要两倍，再等一年。价格不变。但现在你想要两倍的晶体管，不管等多久，你都得付两倍的钱。

There was a time if you want to get twice as many transistors, you just wait one year. And if you want twice as many transistors, wait another year. Same price. But if you want twice as many transistors now, doesn't matter how long you wait, you have to pay twice as much.

明白吗？因为摩尔定律结束了。所以无论世界想要多少计算容量，每一年，如果你想要两倍，你就得付两倍的钱。

Okay? Because Moore's law is over. And so whatever computing capacity that the world wants, every single year, if you want twice as much, you're going to pay twice as much.

嗯。

Yeah.

所以在那 5000 亿美元自然增长的基础上——因为世界只是用得越来越数字化——你实际上还得给它乘上一个系数。

And so on top of that half a trillion dollars that was kind of growing just naturally — because the world just used more and more digital — you're going to have to effectively give it an X factor.

嗯。

Yeah.

就是为了计入摩尔定律已经结束这个事实。

Just to account for the fact that Moore's law is over.

对吧？

Right?

所以即便没有 AI，半导体产业本来也在走向成为全球最大产业。

And so without AI, the semiconductor industry was on its path to become the largest industry in the world.

对。

Right.

我很久以前就提过，只是人们忘了。这就是我说过的事情之一。

And I've mentioned that long ago, just people forgot. It's just one of those things that I said.

对。对。对。对。我两年前刚说过。嗯。

Yes. Yes. Yes. Yes. I just said that two years ago. Yeah.

嗯。所以那只是一部分。

Yeah. And so that's just one part.

对吧？

Right?

然后别忘了，因为我们正处在一次新的计算平台迁移中，会有大量初创公司涌现。

And then don't forget, because we're in a new computing platform shift, there are going to be a whole lot of startups.

创新是经济的一部分，对吧？新公司——

Innovation is part of the economy, right? New companies —

而且我们说过它在加速。

And we said it's accelerating.

嗯。你知道，所以如果我们花——回到摩尔定律，通用计算，每吉瓦 50 亿美元。

Yeah. You know, so if we spent — if you go back to Moore's law, general purpose compute, $5 billion per gigawatt.

嗯。

Yeah.

今天是 500 亿。

Today it's 50.

嗯。

Yeah.

对。对于 Vera Rubin 来说。所以问题是：Vera Rubin 是否在世界上创造了 10 倍的价值？

Right. For a Vera Rubin. So the question is: is Vera Rubin delivering 10x more value in the world?

所以结论是：我认为没必要把任何一家实验室外推到数万亿美元。

So the punch line is: I don't think it's necessary to extrapolate any of the labs into multiple trillions of dollars.

嗯。

Yeah.

没必要那么做。

Unnecessary to do that.

嗯。我觉得这一点很重要，因为我认为人们——我们昨天甚至看到市场上有一些议论，说 OpenAI 的营收到底是 700 亿还是 500 亿。我想你和我——

Yeah. I think that's an important one, because I think people — we just saw yesterday even in the market some chatter about whether OpenAI's revenue was 70 billion or 50 billion. I think you and I —

它们会成为极其成功的公司。

They're going to be insanely successful companies.

没错。

Correct.

嗯。

Yeah.

但你看到很多基础设施公司回调了，对吧，就是因为担心：天哪，OpenAI 没钱，将来也不会有钱来支付这些基础设施。

But you saw the pullback in a lot of the infrastructure companies, right, on the fear that oh my gosh, OpenAI doesn't have the money, won't have the money in order to pay for the infrastructure.

我想你反复强调的一点是，当我们看明年 40 吉瓦的基础设施建设时，至少三分之一只是用于通用算力的替换，也就是超大规模云厂商那些最基础的东西。三分之一会流向新云、主权云、企业，以及所有其他在用这些算力的玩家，包括所有初创公司。还有三分之一不是流向两家实验室，而是流向许多家实验室，对吧，包括 Groq 和 Thinking Machines 以及其他那些。

I think the point that you're hammering home is that when we look at the 40 gigawatts of infrastructure build-out next year, at least a third of it is just going to right general purpose compute replacement, the basic stuff out of the hyperscalers. A third of it is going to neoclouds, sovereigns, enterprises, everybody else who's using this, all the startups. And a third of it is going to not two labs but to many labs, right, including Groq and Thinking Machines and all the other yet.

我们给两家实验室施加了太大的压力。

We're putting way too much pressure on two labs.

我很高兴能与他们合作并支持他们。它们将成为历史上最重要的两家科技公司。它们会取得极其辉煌的成功。

I'm delighted to partner with them and support them. They're going to be two of the most consequential technology companies ever built in history. They're going to be magnificently successful.

是的。

Yeah.

但它们不必做到接近一万亿美元的营收才算成功。对吧。

But they don't have to be anywhere near a trillion dollars in revenues to be successful. Right.

是的。我认为我们在这上面施加了太大的压力。而至于行业里的其他部分，我们仍然需要建设几万亿美元的算力基础设施。

Yeah. And I think we're putting way too much pressure on that. And then the rest of the industry, we still do have to build out a few trillion dollars of computing infrastructure.

嗯，继续吧，嗯，我们稍微换个话题。各位，这个好吃吗？顺便问一下，这个叫什么？

Um, just to keep going, um, let's let's shift a little. How good is this, you guys? By the way, what are what are these called?

这是话梅。

These are dried plums.

话梅。老式话梅。

Dried plums. The old dried plums.

是的。台湾有一家特别的店，是我最喜欢的之一。

Yeah. This this one particular store in Taiwan is one of my favorites.

有点甘草味。

Kind of a licoricey.

是有甘草。是的。甘草。有点甘草味。嗯，盐、糖、肉桂。

There is licorice. Yeah. Licorice. A little licoricey. Uh salt, sugar, cinnamon.

是的。有点冲劲。

Yeah. Has a little twang to it.

是的。是的。你说对了。你知道的。

Yeah. Yeah. You got it. You know.

是的。

Yeah.

好。

Okay.

你说到明年，对英伟达芯片的需求远超 100%。你给出的营收增长指引是 70%。你做了一个一年的前瞻，因为你说供应受限。

You said as to next year, demand for Nvidia's chips well over 100%. You guided to 70% revenue growth. You gave a year forward look because you said supplies constrained.

那是我当时能给出的全部。

That's all I had at the time.

没错。是的。你说我有 18 个月的时间来做这件事。

That's right. Yeah. You said I got 18 months to work here.

对。

Right.

你说你非常有信心。

You said you were very confident.

不是每天来上班就行。

Don't come to work every day.

你说非常有信心，而且有充足的时间与供应链合作，做得更好。你期望做得更好，不辜负客户。那么事情是不是进展得更好了？订单簿是否还在持续增长？因为另一部分是这样的，对吧，你在某个时间点拍了一张快照，当然人们还在继续下单。所以你认为供应链里是否还有足够的余量让你挤出更多产能，而随着这些能力被不断解锁，订单簿是否还在持续增长？

You said very confident and plenty of time to work with the supply chain to do even better. You expected to do even better and to not disappoint customers. So are are things going even better? And has the order book continued to grow? Because the other part of that is right, you took a snapshot at a moment in time and of course people are continuing you know to place orders. So it do you think there's enough slack in the supply chain you can squeeze more out and is the order book continuing to grow as more of these capabilities are unlocked?

首先让我告诉你我为什么告诉所有人。嗯，我们面临的挑战是我们有巨大的订单簿。每一家 CSP、每一家 NCP、每一家 OEM 都已经向我们下了订单。初创公司、大公司、小公司，他们都已经向我们下了订单。原因在于我们和他们合作时，会根据客户的规模，至少提前一年，嗯，对最大的客户是两年的预测。所以我们已经有一个非常大的订单簿，而因为我当时正在做分配，我分配给所有客户的量远少于

The first let me tell you why I told everybody. Um the challenge that we have is we had giant order books. Every single CSP, every single NCP, every single OEM had already placed orders on us. Startups, big companies, small company, they already placed orders on us. And and the reason for that is because we work with them on a on a depending on on the size of the customer at least a year um for the the largest ones two years forecast. And so so we already have a a very large order book and because I was in the process of allocating and what I was allocating to all of our customers is a lot less than

他们想要的

they wanted

他们想要的。

they wanted.

是的。如果我，如果我没有公开说

Yeah. If I if I was if I didn't publicly say

对

right

我们有什么

what we had

对

right

每个人都会猜疑，是不是被克扣了，我们是不是被，你知道，所以

everybody would be second-guessing are these short changed are we being you know and so

所以这是我处理每一次让人失望的对话的方式

and so it was it was my way of of essentially handling um every single conversation where I'm disappointing people

对

right

嗯，而且

um and

所以基本上每个客户，只要你能交付更多，都会从你这里买更多

so pretty much every customer would buy more from you if you could deliver more

远远如此，然后从那以后，当然我们几乎每周、每天都有新订单。所以，嗯，此外，我们也在获得更多供应，而限制我们的一些供应是土地、电力、厂房，下游，对吧

by far and then and then since then of course we have new orders almost almost weekly and daily and and so so um uh in addition however we're getting more supply and some of the supply that that is limiting us is land power shown downstream correct

所以几周前我们宣布了在澳大利亚的 2 吉瓦

and so a couple weeks ago we announced 2 gigawatts in Australia

所以这是件大事，2 吉瓦对我们来说大约是 800 亿美元

and so that's a very big deal 2 gigawatts to us is like $80 billion

巨大

huge

然后，嗯，我希望当我和 UTC 一起去欧洲时，你知道，我们能在那里宣布一些新的合作，并解锁更多的土地、电力、厂房。所以土地、电力、厂房显然是障碍之一，而这是英伟达拥有巨大优势的一个领域。

and and then and then um um I'm hoping when I go to Europe with UTC you know we'll be able to announce some some uh new partnerships there and and unlock more more uh land power shell and so so this land power shell is is clearly one of the obstacles and and this is this is a an area where Nvidia has a huge advantage.

是的。你知道，因为基于我们所构建的东西，如果，嗯，有一家新云，嗯，像 CoreWeave 或 EBS 这样的新云，他们想要扩展业务，英伟达拥有整个全栈完整的 AI 基础设施，所以我们可以支持他们扩展。嗯，所以我们可以在别人无法建设的地方建设，与别人无法合作的伙伴和客户一起建设。嗯，所以我认为这给了我们更多的韧性。但尽管如此，嗯，随着我们继续在上游和下游增加供应，你知道，嗯，一个关键领域是基板。基板变得越来越大，芯片越来越多，所以它在两个方面对他们不利。每一颗出货的芯片都使用更多的材料重量。所以，

Yeah. you know, because based on what we build, if if uh there's a a Neo cloud um uh and uh this Neo cloud like Core Weave or EBS um and they would like to to scale out their business, Nvidia has the entire full stack complete AI infrastructure and so so we could support them as their scale out. Um and so we we could build in places and we can build with partners and customers that nobody else can. Um and and so I I think this this gives us a lot more resilience. But none nonetheless uh as we as we continue to increase supply upstream and downstream uh you know uh one of the critical areas is substrate. Substrates are getting larger and there more chips and so so it's working against them in two ways. every single chip that's being shipped uses more of the weight of the material. And so,

所以这些，嗯，有机材料真的很难以我们需要的精度制造，因为随着封装变大。嗯，热膨胀系数，嗯，以及我们堆叠的方式，使得处理翘曲之类的问题变得非常困难，你知道，还有各种制造缺陷。所以在我们这样的规模下，我们必须非常精确，我们需要非常优秀的合作伙伴，而且

so these these um organic materials are are are really hard to make at the precision that we need because the as the packages get larger. um uh the the the coefficient the thermal coefficients um and the way we stack things make it really hard to handle things like warpage and you know and various various manufacturing defects and and so at the scale that we do we have to be very precise and we need really great partners and

所以这就是供应受限，这些只是一些例子。嗯，硅光子也供应受限，所以一切都是供应受限的。所以我们会继续努力，我有 18 个月来做这件事，你知道，每过一周我们都在增加，但尽管如此，我们不断增加更多供应，而订单簿持续上升

and so that's that's supply constraint and just that those are just some examples uh silicon photonics is is supply constrained and so so everything is supply constrained and so we're going to keep working on I have 18 months to do so when you know every single week that goes by we increase but nonetheless every we keep increasing more supply and the order books keeps going up

对

right

所以这个缺口在几年内都很难弥合

and so it's going to be hard to close that gap for a few years

嗯，看看这张图，嗯，你知道

um take a look at this chart um you know

这就像一场庭外取证

this is like a deposition

不，对吧

no right

不好意思，通常是这样，Monk 先生，请，请，现在

excuse me now usually goes like this Mr. Monk ex please please now

受询人

examinate

请现在看，受询

please now look at exam

好的庭外取证。那么灰线，这是 2025 年初的共识，这是 2025 年初对英伟达的共识。蓝线是当时 Alimter 的预测，所以我们当时比共识线高得多，而现在新的共识是基于你的指引。所以让我震惊的一件事是，那 25 个拿钱专门预测这家公司会怎样的人。他们所有想法的共识，对吧，偏差了几百个百分点。对吧，就在 20 年初，呃，哦不。去年的这个时候，

good deposition okay so the gray line this is was the consensus at the start of 2025 this was the consensus for Nvidia at the start of 2025 the blue line was alimter's forecast way back then so we were a lot higher than the consensus line is now where the new consensus is based upon your guidance and so one of the things that's blows my mind about this is the 25 people who get paid to just forecast what this company is going to do. Their consensus of all their ideas, right, was off by a couple hundred%. Right, just at the beginning of 20 uh Oh, no. This time last year,

去年的这个时候，连 25 那条线都不到。

at this time last year, not even 25 line.

是的，那是去年这个时候的共识。一年前。

Yes, that was the consensus at this time last year. One year ago.

上周。

Last week.

好吧。所以，一年之内，这家公司的共识预测上调了 100% 到 200%。

Okay. So, in a year, the consensus forecast has shifted up by 100 to 200% for this company.

嗯。

Mhm.

太惊人了。在我 20 多年的从业经历中，很少见到共识预测偏差这么大。然而股价只涨了 20%。你们实际上把公司的营收预测和自由现金流预测都翻了一番。现在市盈率只有 13 到 14 倍，取决于你看谁的数据。那么你认为市场对英伟达或对市场预期的信心在说什么，导致这种情况？因为通常如果预测上调这么多，股价应该涨得更多。

Extraordinary. In my 20 plus years of doing this, you don't often see consensus being off by that much. And yet the stock's only up 20%. You literally doubled what the revenue forecast was and the free cash flow forecast was for the company. You're now trading at 13 to 14 times earnings depending upon who you're looking at. So what do you think the market is saying about Nvidia or about the confidence in the expectations of the market right to cause that to be the case? Because usually if you see it shift up by that much the stock would have shifted up by a lot more.

首先，像我们这样规模的公司还能获得市场份额，这说明了韧性。

First of all, for a company to be as large as us gaining share tells you something about the resilience.

是的。

Yes.

以及我们优势的规模和范围，

And the size and scale of our advantage,

对吧？

Right?

而且

And

这个优势是在增加还是减少？你的护城河是在加深还是变浅？

Is that advantage increasing or decreasing? Is your moat increasing or decreasing?

在增加。

Increasing.

你的竞争护城河随着规模在加深。

Your competitive moat is increasing at scale.

是的。让我举个例子。我们现在就做个思想实验。如果你现在必须给某人一张 1000 亿美元的支票，为下一代产品下采购订单

Yeah. Let me give you an example. Let's just do a thought experiment right now. If you had to hand somebody a hundred billion dollar check to place a purchase order right now for a next generation

是的。

Yeah.

如果你想排队购买 Vera Rubin Ultra，现在快到下单时间了。

If you want to get in line for Vera Rubin Ultra, it's close to placing order time.

哇。

Wow.

连看都没看过。

Sight unseen.

是的。

Yeah.

对。但他们有信心。所以假设你必须把 1000 亿美元押在某人身上。是的。一个

Right. But they have the confidence. And so let's say there's a you got to place a hundred billion dollars on somebody. Yeah. One

你怎么知道它会好呢，

How do you know it's going to be any good,

对。但因为英伟达制造它，你知道它会比 Vera Rubin 更好，

Right. But because Nvidia builds it, you know it's going to be better than Vera Rubin,

对吧？

Right?

而且 Vera Rubin 很棒，

And Vera Rubin's awesome,

对吧？

Right?

Vera Rubin 之所以——大家对 Vera Rubin 有信心的原因是 Grace Black 很棒。

And the reason why Vera Rubin you the reason why everybody had confidence in Vera Rubin is because Grace Black was awesome.

是的。

Yeah.

所以，在计算世界里，这些计算机执行的工作负载非常复杂。你不能只用一个基准来衡量。

And so and the world of computing the the workloads that these computers do are so complex. You can't just have like measure one benchmark.

对。对。

Right. Right.

你知道，这是 500 亿美元。是的。要交付这样的性能，需要整合的技术数量多得惊人，所以问题是，在连看都没看过的情况下，你能把 500 亿美元的订单下给谁？

You know, it's $50 billion. Yes. And the number of technologies have to come together to deliver the performance is insane, you know, and so so the question is sight unseen, who can you place a $50 billion order on?

是的。

Yeah.

我们 literally 是唯一的一家。

We're literally the only one.

是的。

Yeah.

以前，下 10 亿美元的采购订单已经够有挑战性了。但如果你下 1 亿美元的订单，而且系统就在你面前，你直接就可以对芯片进行基准测试。

Now, back when it was place a PO, a billion dollar PO, that was challenging enough. But if you had place a $100 million PO and the the the the system is sitting right in front of you, you could just bench you could benchmark the chips.

对。

Right.

对。你怎么对这些系统进行基准测试？

Right. How do you benchmark these systems?

光是建造它就是个奇迹。

Just building it is a miracle.

对。你知道，如果有人想做评估，我很乐意支持，但你需要给我一个巨大的数据中心，我们需要几年时间来改造它，所以现在你甚至无法对这些东西进行基准测试了。

Right. You know, so if somebody wants to do an evaluation, I'm happy to support it, but you're going to need to give me a giant data center and we're going to need a couple years to retrofit it and you know, and so you can't even benchmark these things anymore.

有趣。所以这就像台积电的情况一样。

Interesting. And so it's exactly what happened with TSMC.

我们把 100% 的业务交给台积电的原因是，他们是世界上唯一一家我可以提前五年规划的公司，

The reason why we give 100% of our business to TSMC is because they're the only company in the world that I can plan five years out,

对吧？并且知道

Right? And know

而且在连看都没看过的情况下，知道那个晶体管会非常好，因为它比那个好。比那个好。

and sight unseen know that that transistor is going to be really good because it's better than that one. Better than that one.

是的。

Yeah.

而且他们会有产能给我，他们知道如何与我的业务团队合作，他们有敏捷性来处理我所有的起伏，比如应对我从灰色到蓝色到绿色的转变。

and that they will have capacity for me and that they know how to work with my business team that they have the agility to handle all of my ups and downs and you know for example to deal with with the fact that I went from gray to blue to green.

是的。没错。好的。

Yeah. Exactly. Okay.

没错。他们必须足够大，才能帮助我从灰色到蓝色到绿色。

Exactly. They have to be big enough to go help me go from gray to blue to green.

没错。

Exactly.

世界上没有其他公司能帮我做到这一点。

There's no other company in the world that could help me do that.

没错。并且拥有所有这些不同的工艺技术，从硅光子学到逻辑到内存，再到推拉式封装。我的意思是，我们做的事情数量非常动态

That's right. and to have and and to have all these different process technologies from silicon photonics to logic to memory to and push pull push pull packaging. I mean the number of things that we do is so dynamic

所以，你怎么可能用一家只有他们十分之一规模的公司来做到这一点？这不可能。所以我认为核心观点是：因为我们正在为地球建设基础设施，而英伟达已经成为一家基础设施公司。

and so how do you possibly do that with a company that is one-tenth their size it's impossible and so I think the big idea is this because we're building infrastructure for the planet and Nvidia has become an infrastructure company.

嗯。

Mhm.

而且因为这些基础设施耗资巨大，它们在变化，技术发展如此之快。

And because these infrastructures cost so much money and they're moving and technology is moving so fast.

是的。

Yeah.

我们实际上，你知道，信任的结合，我们运行一切的事实，你知道，所以它始于一些 GPT 模型。嗯，但看看被发明的 flash attention 机制。我们将模型并行化并在 MVLink 72 上分片的方式，以及所有这些如何改变，现在人们在谈论 Transformer 扩散模型、混合 SSM、Transformer 混合体，所有这些不同的技术都是对的

We're practically, you know, the combination of the trust, the the fact that we run everything, you know, so it started out with some some GPT models. Um, but look at the flash attention mechanisms that were were re were invented. the way that we parallelized the models and sharded across MVLink 72 and how all of that changed changed and now people are talking about transformer diffusion models that are hybrid SSMs that are transformers hybrids and all these different technologies are all right

所有这些的演变，我们知道它们都运行在 CUDA 上

the evolution of all that we know that it all runs on CUDA

对

right

这是唯一的，我们呃，我们不知道你是否需要做更多的预训练、后训练还是只是智能体推理，好吧，事实证明我们无论如何都运行它，对吧

it is the only we uh we don't know whether you need to do more pre-training post-raining or just agentic inference well it turns out we run it all anyways right

所以你可以建造一个 500 亿美元的数据中心，知道工作的可替代性，如果发生了什么事，你知道，无论出于什么未知原因，你想转租一半的容量给某人

and so you could build a $50 billion ion dollar data center and know that the fungeability of the work and if something were to happen you you know for whatever reason unknown you wanted to if you will sublease half of your capacity to somebody

英伟达是完全可替代的

Nvidia is completely funible

对

right

所以我们在工作负载上可替代，在模型上可替代，在客户之间可替代

so we're funible in workload we're funible in models we're fungeible across customers

而且我们可以在世界任何地方部署。你在欧洲创办公司，你在美国创办公司，但猜猜怎么着，你想服务欧洲的客户，你想服务日本的客户，好吧，事实证明我们已经在那里了。英伟达的基础设施已经在那里了。所以，OpenAI 想去欧洲。不，我们会立刻带他们去。如果他们想去印度，立刻，如果他们想，对吧？如果 Anthropic 想去日本、韩国，立刻，我们已经在那里了，

and we're deployable everywhere in the world you start a company in Europe you start a company in United States but guess what you would like your customer you want to serve customers in Europe you want to serve customers in Japan well Turns out we're already there. Nvidia's Nvidia's infrastructure is already there. So, OpenAI wants to go to Europe. No, we'll take them there in a heartbeat. If they want to go to India in a heartbeat, if they want, isn't that right? If Anthropic wants to go to Japan, Korea, in a heartbeat, we're already there,

对吧？

right?

所有这些架构都已经在那里了。所以，我认为，如果你把英伟达作为一个整体来看，然后你回来说，好吧，

All of these architectures are already there. And so, I think the the if you take Nvidia in a whole and then you come back and you say, okay,

我想做这个思想实验。我想把 1000 亿美元押在某人身上。嗯

I want to do the thought experiment. I want to place a hundred billion dollars on somebody. M

你怎么做？

how do you do it?

一年前，我们坐在这里，共识预测是你们在 2027 年将持平或萎缩。现在共识预测你们将增长 74%。

A year ago, we were sitting here and the consensus forecast was that you were going to be flat or shrink in 2027. Now the consensus forecast you're going to grow 74%.

好，我想让大家先消化一下，因为很多人都在追随这些共识预测，对吧？结果错得离谱。他们当时预测……

Okay, so I just want to let people let that sit in, because a lot of people follow these consensus forecasts, right? And it was really wrong. They were forecasting...

他们有没有回去听你上次的播客？

Did they go back and listen to your podcast from last time?

对。你早就告诉过他们了。

Right. You told them.

是的。我全都解释过了。各位，听好了，我确实全都解释过了。

Yeah. I explained everything. Listen, ladies and gentlemen, I did explain everything.

所以我把目光投向 2029 年。

And so I look out at 2029.

当时有个 ROI 的问题，对吧？

There was an ROI question, wasn't it?

是的。

Yes.

然后还有这个定制与通用之争。

And then there was this custom versus general purpose.

英伟达之所以走通用路线，就是为了拥有这种可替换性。

The reason why Nvidia is general purpose is so I can have all this fungibility.

没错。

That's right.

可替换性的好处就是低风险。

The benefit of fungibility is low risk.

对吧？而低风险在你投入大笔资金时才真正发挥作用。当你建设价值 500 亿、1000 亿美元的基础设施时，你需要知道这套设施在很长一段时间内都会非常出色，对吧？

Right? And low risk really comes into play when you're placing big dollars. When you're building out 50, 100 billion dollars worth of infrastructure, you need to know that that infrastructure is going to be really good for a long time, right?

无论如何。

No matter what.

所以专业化是可替换性的敌人。

So specialization is the enemy of fungibility.

专业化是耐久性的敌人。

Specialization is the enemy of durability.

对吧？只要我能证明我们的性能超凡，而且我们每天都在证明这一点，

Right? And so long as I can prove that our performance is extraordinary and we prove that every single day,

一旦我们能证明这一点，可替换性和耐久性就会转化为可租赁性，进而转化为生产力和可融资性，因此可融资性带来最低的总拥有成本（TCO）。

the moment we can prove that, then fungibility and durability translates to rentability, which translates to productivity and financeability, and therefore financeability, which leads to the lowest TCO.

对。对。正是如此。

Right. Right. Exactly.

是的。

Yeah.

现在你们正在进行史上最大规模的回购。而且据我计算，你们可能在未来两年内回购高达 10% 的公司股份。

Well, now you're buying back the largest amount in history. And you know, on my math, you could probably buy upwards of 10% of the company back within the next two years.

在高盛，你称英伟达是第一家成长价值型公司。是的。

At Goldman, you called Nvidia the first growth value company. Yeah.

听起来很像苹果当年的说辞。

Sounds a lot like Apple's pitch to back.

好吧。是这样吗？

Okay. Is that right?

是的。或者更多。是的。或者说他们能够返还资本。

Yeah. Or more. Yeah. Or that they could return the capital.

那确实奏效了。

It worked for sure.

奏效了。从当时的 300 美元、30 美元涨到了 300 美元。

It worked. Went from 300 at the time $30 to 300.

是的。

Yeah.

300 多。

300 plus.

在很多方面，我们正在经历类似的转型。比如，我忘了提的一点是——布拉德，我想你正要问我这个——我们大部分的投资和融资合作策略，都是在尝试建立两种商业模式，

And in a lot of ways, we're going through a similar transition. And for example, the thing I forgot to mention is—and Brad, I think you were about to ask me about that—most of our investment and financing partnership strategies is trying to set up two business models,

对吧？

Right?

两种新商业模式。其中一种我们已经做了一段时间，并将在不久的将来加速推进，那就是软件授权。

Two new business models. One business model we've been working on for some time and will be accelerated in the near future is software licensing.

你知道，我们公司拥有的软件体系非常庞大，那个数据中心规模巨大。想象一下里面运行的所有软件，

You know, the body of software that we have in our company is just extraordinary, and that data center is so large. Imagine all the software that runs in it,

而我们提供整个技术栈。

and we provide the entire stack.

嗯，我们的整个技术栈对云服务提供商（CSP）来说完全是工业级强度的。他们想运行自己的技术栈，对吧？

Well, and our entire stack is completely industrial strength for the CSPs. They want to run their own stack, right?

所以我们授权或开源我们的软件给他们，然后他们拿去修改、集成到自己的系统里，再运营他们的服务。但对世界上很多公司来说，他们没有 IT 部门，没有软件部门来做这件事。所以他们需要我来维护。

And so we license or we give them our software open source, and then they take it and they modify it and they integrate into theirs, and then they operate their services. But for a lot of the world's companies, they don't have the IT department, they don't have the software department to do it. So they need me to maintain it.

因为过去几年 AI 主要聚焦在训练上，但现在他们正转向部署，

And so because AI in the last several years was largely focused on training, but now they're moving into deployment,

对吧？

right?

他们需要把这些当作工厂来运营，他们需要知道我会负责。我签了服务等级协议（SLA）来支持、维护它，保障安全，只要我们在，就一直保障。所以面向企业业务的软件授权将会非常、非常庞大。

They need to operate these as factories, and they need to know that I'm on the hook. I'm on the SLA to support it and maintain it, keep it secure, keep it safe for as long as we shall live. And so software license for enterprise business will be very, very large.

所以，你知道，如果我们的软件授权费，简单算一下，大约每块 GPU 每年几千美元，而我们有数百万块 GPU，

And so, you know, if our software license is, just roughly say for simple math, a couple of thousand dollars per GPU per year, and we have millions of GPUs,

对吧？

right?

我们一年出货，比如说 2000 万块 GPU。还在增长，其中一半进入全球企业。加起来就很可观了。

We ship, you know, 20 million GPUs say a year. I'm growing, and so half of it goes into world's enterprise. It adds up.

是的。

Yeah.

好。所以我相当确定，这将成为我们业务中非常大的一部分——软件授权。然后当然，有些云厂商依赖英伟达提供全栈。

Okay. And so I'm fairly certain this is going to be a very big part of our business—software license. And then of course there are some clouds that rely on Nvidia for our full stack.

他们依赖我们的承购、我们的生态系统。他们也依赖我们，也许作为投资伙伴，帮助他们获得融资和资本。

They rely on us for our offtake, our ecosystem. They also rely on us for maybe as an investment partner in helping them secure their financing and capital.

他们搞定了 Lamp、Power、Shell。他们是区域冠军。他们有各种优势。但在这三个领域，他们需要我们的支持。所以在那种关系中，我们很可能会转向收入分成（RevShare）。我们会投资。我们会先获得销售的好处，但未来也会获得收入分成的好处。所以，你知道，我们不是云服务提供商。嗯，但我们会为他们提供承购和软件支持，持续……

They've secured Lamp Power Shell. They're regional champion. They got all kinds of stuff working in their advantage. But these three areas they need our support on. And so in those kind of relationships, it's very likely that we'll move to RevShare. And so we'll make an investment. We'll get the benefit of the sale up front, but we'll also have the benefit of RevShare going forward. And so, you know, we're not a cloud service provider. Um, but we're going to be providing offtake and providing software support for them for, you know,

四、五、十年。

four, five, 10 years.

所以，我们应该从中获得一些经济回报。所以，收入分成——

And so, we ought to get some economic return on that. And so, RevShare—

注意，我刚才描述了两个新的增长向量，独立于我们公司的硬件销售，而且更与我们平台的装机量相关。如果没有装机量，软件授权有什么好处？如果没有承购，怎么会有收入分成？

notice I just described two new growth vectors independent of the hardware sale of our company, and it's much more related to the install base of our platform. What's the benefit of software licensing if you have no install base? And how could you have RevShare if there's no offtake?

好的。

Okay.

所以装机量要素、我们软件栈的丰富性、承购、我们承购的丰富性以及我们在全球的生态系统——这些将推动它。你知道，其中一种方式，我们谈到这个 13 倍的训练。当蒂姆和卢卡去见沃伦·巴菲特时，巴菲特说过他不喜欢科技公司的原因之一是，他们会有这种爆款，产生大量自由现金流，却从不返还给股东，而是拿去花在别的东西上。而往往那个别的东西并不如最初产生现金流的那个东西好。

And so the install base element, the richness of our software stack, the offtake, the richness of our offtake and the ecosystem that we have around the world—that's going to fuel that. One of the ways, you know, so we're talking about this training at 13 times. You know, when Tim and Luca went to see Warren Buffett, one of the reasons that Buffett has said he doesn't like technology companies is they have this hit, they produce an enormous amount of free cash flow, and rather than ever giving any of it back to the shareholders, they go and spend it on something else. And oftentimes the something else isn't as good as the thing that caused them to generate the cash flow in the first place.

你们现在把超过 50% 的自由现金流用于回购股票，实际上是在返还给股东，同时你们还拥有这种指数级增长。正如我们已经讨论过的，

You guys are now spending over 50% of your free cash flow buying back your stock, effectively returning it to shareholders while you have this exponential growth at the same time. As we've already discussed,

我们说过 50，

we said 50,

对吧？

right?

我想上个季度我们做了 60 多。

I think this last quarter we did 60 some odd.

没错。

Correct.

是的。我们会增加。你知道，我们投资非常高质量的公司，正如你们所知，而且，嗯，过去几年，英伟达从谁那里买，黄仁勋从谁那里买，你就得投资那家公司。没错。

Yeah. We're going to increase. You know, we invest in really high quality companies as you guys know, and, um, and you know, the last couple of years, whoever Nvidia buys from, whoever Jensen buys from, you got to invest in that company. That's right.

现在街上的新说法是，黄仁勋投资谁，你就投资那家公司。而且这是真的。

And now the new word on the street is whoever Jensen invests in, you invest in that company. And it's true.

原因在于我们身处这个生态之中。我们处在信息流里。我们共同了解那些投资。

And the reason for that is because we're in the ecosystem. We're in the flow of information. We know those investments together.

对。

Yeah.

我们认识很多聪明人，当他们想做一番大事时，我们愿意支持他们。所以我们算是身处内部。我们掌握大量信息，因此我们的投资是信息充分的。

And we know a lot of smart people, and we want to back these people when they want to do something great. So we're kind of on the inside. We have a lot of information, and so our investments are well-informed.

没错。

Right.

更不用说英伟达也在这个生态里，所以我们可以帮他们成功。

And not to mention Nvidia is also in the ecosystem, so we could help them succeed.

说得对。

That's right.

所以这些投资之所以表现优异，有很多不同的原因。但作为我们总现金生成的比例，它会迅速下降，部分原因是 OpenAI 和 Anthropic 这样的机会不会再有——坦白说，那个机会不会再来了。其次，我们的现金流生成增长相当快。

And so there are a lot of different reasons why the investments turn out great. But as a percentage of our total cash generation, it's going to decline very rapidly, partly because OpenAI and Anthropic isn't going to come around again — that opportunity is not going to come around again, frankly. And then secondarily, our cash flow generation is growing quite fast.

对。

Right.

说到这一点，在我看来，如果你回购那么多，而收入的持久性——正如你今天在这里描述的——比人们预期的更长。我认为这正是人们仍然过度纠结的一点：这只会持续这么久。前几天 Ezra 在播客上问过你：泡沫还有多久破裂？我记得你说过大概两到三年。你对需求的存在有很好的可见性，之后谁知道呢。我听到有人说，哦，可能只有两年就会破裂。但当我审视这件事时，这从根本上是对全球所有算力的永久性重新架构。

And to that point, it seems to me that if you're buying back that much and the durability of the revenues — I think as you've described here today — is longer than people expect. And I think that's the one place people are still hanging too much on: this is only going to last so long. Ezra asked you this on the pod the other day: how long until the bubble burst? And I think you said something about 2 to 3 years. You have good visibility that the demand is there, and who knows after that. And I heard some people saying, oh, it may only be two years before this bursts. But when I look at this, this is fundamentally a rearchitecture of all global compute forever.

永久。

Forever.

所以除非发生全球大萧条。

So unless there's a global depression.

对。

Yeah.

没错。那么每一年人们都将不得不消耗巨量的英伟达算力。

Right. Then every single year people are going to have to consume an enormous amount of NVIDIA compute.

对。所以这样想：云服务商大约运行着五千亿。整个行业、整个世界每年大约运行着——姑且说——一万亿美元的扩建。

Yeah. And so the way to think about that is, the CSPs were running about half a trillion. The whole industry, the whole world is running, call it a trillion dollars of it, build-out each year.

对。

Right.

好。每年都会建更多，大约每五年就有一部分折旧或关闭、下线，等等，有些则持续更久。旧世界是检索式的计算方式；而新的计算方式，我预测会持续数十年。所以三、四，姑且说四，姑且说五。我的感觉是，我们会从那里继续增长。如果你告诉我，每年的产能建设、基础设施建设、更新——新旧更新的总和——大约运行在 10 万亿美元，我一点也不惊讶。

Okay. And every year more gets built, and every 5 years or so some gets depreciated or turned off and taken offline, and so on and so forth, and some of it lasts longer. And so that old world was in the retrieval way of doing computation; the new way of doing computation, I've predicted, is going out of the decades. So three, four, let's say four, let's say five. And my sense is that we're going to continue to grow from there. If you told me that the annual capacity building, infrastructure building, refreshing — combination of new and old refreshes — runs around 10 trillion dollars, I'm not even a little bit surprised.

对。没错。我的意思是，再为听众说明一下，你认为到本十年末每年会有 4 万亿美元的支出，因为每次你消耗 token，每次你使用智能，你都得运行机器——工厂必须运转。我记得你描述过，那大约是 50%——如果你假设全球 GDP 中有 50 万亿美元与知识工作相关，如果我们在 5 年内从中获得 20% 的生产力提升——就像你在英伟达内部看到的，你在英伟达内部看到过更大的生产力提升。但如果你看到那种生产力提升，那么我们就会获得 10 万亿美元的增量经济活动。

Yeah. Right. I mean, just again for people listening, you think $4 trillion of annual spend by the end of the decade, because every time you consume tokens, every time you use intelligence, you have to run the machine — the factory has to run. And I think you've described that that's about 50% — if you assume that $50 trillion of global GDP is related to knowledge work, if we get a 20% productivity gain over 5 years on that — like you've seen inside Nvidia, you've seen much bigger productivity gains inside Nvidia. But if you see that sort of productivity gain, then we get $10 trillion of incremental economic activity.

对。没错。

Yeah. Right.

到 2029、2030 年，这看起来非常合理。而你只是说，我们会把其中的一半花在资本支出上，花在用于生产它的智能上。

By 2029, 2030, like that seems very reasonable. And you're just saying we'll spend half of that on capex on the intelligence to produce it.

所以也许这是我们稍微下匝道的好地方。你非常慷慨。我可以一直聊下去。但我想回到我们开始的地方，好吗？当我们回顾——如果这是那种边吃边喝的播客，我们可以继续聊。我们可以继续聊。

And so maybe that's a good place for us to take a bit of an off-ramp. You've been incredibly generous. I could go on forever. But I kind of want to wrap where we started, right? When we look back — if this was one of those eating, drinking podcasts, we could keep going. We could keep going.

对。

Yeah.

但不久之前，人类的境况还被描述为肮脏、野蛮而短暂，对吧？

But it wasn't so long ago that the human condition was described as nasty, brutish, and short, right?

那是上周还是——

Was that last week or —

创新真正驱动了人类所有的进步，主要是在过去 200 年里。更多与家人相处的时间，更长的寿命，更少做人类真的不想做的苦差事。更安全。

Innovation has really driven all human progress, mostly in the last 200 years. More time with family, longer lives, less time doing drudgery that humans really didn't want to do. Safer.

我们是第一代把 63 岁视为年轻的人。

We're the first generation where 63 years old is considered young.

正是。正是。

Exactly. Exactly.

当人们还拥有，你知道——

When people still got, you know —

正是。

Exactly.

人们叫我年轻人。

People call me young man.

更安全、更好的生活。所有这一切，尽管有你我在外面听到的所有这些担忧。顺便说一句，同样的担忧在工业革命初期也被表达过，对吧？你有多大信心，5 年后我们会做出正确的决定，世界会因为我们所做的投资和一切而变成我们描述的那种更美好的地方？今天坐在这里，你比 12 个月前更有信心吗？

Safer, better lives. All of this despite all of these concerns that you and I hear out there. By the way, those same concerns were expressed at the beginning of the industrial revolution, right? What's your level of confidence that in 5 years from now, we will have made the right decisions and the world will be that much better place that we describe, because of the investment and all that we're doing? Are you more confident as we sit here today than you were 12 months ago?

我 99.99%——只是为了致敬一位好朋友，不用 0 和 100。我 99.99% 有信心，我们中有太多人深切关心结果会变好。

I'm 99.99% — just in service, in honor of a good friend, not using zero and 100. I'm 99.99% confident that too many of us care deeply that the outcome turns out well.

对。

Yeah.

我们会投入我们的力量、我们的智慧、我们的资源，确保我们以负责任和乐观的方式到达那里。所以有太多人在做正确的事。现在，常常有人对这些事情夸夸其谈，他们受若干因素影响，却没有意识到，在他们的信息圈之外，有很多人、很多行业、了不起的研究人员和工程师，正独立于他们，靠自己做着惊人的事情。仅仅因为他们没说，并不意味着很多聪明人没有在做聪明的事。

That we will dedicate our might, our intellect, our resources to ensure we get there responsibly and optimistically. And so there are just so many people doing the right things. Now, oftentimes there are people who pontificate about these things, and they're informed by several things, and they don't realize that there are a lot of people, industries, incredible researchers and engineers outside of their sphere of information that are doing amazing things all by themselves, independent of them. Just because they didn't say it doesn't mean a lot of smart people are doing smart things.

对。

Right.

我在公司内部一直这么说：即使没有我的领导，没有战略，公司自己也在做聪明的事。聪明人会做聪明的事。当你把他们组织起来时，他们会做更有影响力的事，但了不起的人靠自己就在做聪明的事。所以当有人告诉我你为此担忧——我们担忧是好事，对吧？我们应该表达我们的担忧。我们应该表达我们的关切。但我们不应该预言世界末日。

And I say this all the time inside the company: even without my leadership, without strategies, the company is doing smart things all by itself. Smart people do smart things. They do more impactful things when you organize them, but amazing people are doing smart things all by themselves. And so when somebody tells me that you're worried about it — it's good that we worry, right? And we should express our worries. We should express our concerns. But we shouldn't predict the end of the world.

对。

Yeah.

因为他们在低估人类的境况。他们在低估人类的能力。而且他们出于某种原因——也许只是自负，也许是傲慢——认为除非他们来做，否则事情就不会完成。

Because they're underestimating the human condition. They're underestimating the human capacity. And they think, for some reason — maybe it's just hubris, maybe it's arrogance — that unless they do it, it won't get done.

对。对。

Right. Right.

除非他们来做，否则事情就不会完成。

Unless they do it, it won't get done.

对。

Right.

事实是，无论我说不说，这家公司都在做着了不起的事情。

The fact of the matter is this company is doing amazing things whether I say it or not.

好的。

Okay.

我得到了所有的赞誉和指责，但在这两种情况下，我都很公平，尤其是赞誉。原因就在于——当你身边都是了不起的人，并且你创造了条件，让每个人都想要同样的结果。

And I get all the accolades and the blame, but in both cases, I'm fair, especially the accolades. And the reason for that is just these—when you surround yourself with amazing people and you create the conditions by which everybody wants the same outcome.

是的。

Yeah.

好的结果，好的版本。我认为大多数人都这样。这就是为什么我说我们需要对自己负责。

The good outcome, the good version of it. And I think most people do. That's the reason why I say we need to hold ourselves accountable and responsible.

但我们应该表达一个乐观的未来。

But we should express a future that's optimistic.

当然。

For sure.

我 99.99% 确定，

And I am 99.99% certain,

对吧？

Right?

在世界上所有了不起的人的包围下，我们会一起到达那里。

That surrounded by all the amazing people in the world, we're going to get there together.

是的，同意。我认为你是国宝。我们很幸运有你帮助领导这项工作。国家感谢你把大家聚在一起，让我们达成共识，建立框架。还会有更多工作要做。但作为两个孩子的父亲，我知道你有孩子，你会有孙子孙女，我们最关心的就是为他们建设一个更美好的世界，对吧？我们很容易停下来。但如果我们今天停下来，我们就输了，对吧？我们失去了那个更美好世界的机会。所以，我们就在这里结束。谢谢你，Jensen。

Yeah, agreed. And I think you're a national treasure. We're lucky to have you helping to lead the charge. The nation's indebted to you for getting everybody together, getting us all on the same page, building the framework. There's going to be more work to do. But as a father of two, and I know you have kids, you're going to have grandkids, we don't care about anything more than actually building a better world for them, right? We could easily stop. But if we stop today, we lose, right? We lose the opportunity of that better world. So, we'll wrap it there. Thank you, Jensen.

谢谢大家。还有 Brad，谢谢你——我从未见过更伟大的爱国者，一个如此关心这个行业的人，但更关心这个国家，以及你的天赋、你的热情和不知疲倦的能量，在协调、团结很多人去做像你所做的这样伟大的事情。太不可思议了。提醒大家，这只是我们的观点，不是投资建议。

Thank you, guys. And Brad, thank you for—I've never met a greater patriot, somebody who cares so much about the industry, but even more so the country and your gift, your enthusiasm and just tireless energy in coordinating, pulling together a lot of people in order to do something great like what you did. Incredible. As a reminder to everybody, just our opinions, not investment advice.