#494 – Jensen Huang: NVIDIA – The $4 Trillion Company & the AI Revolution

The following is a conversation with Jensen Huang, CEO of NVIDIA, one of the most important and influential companies in the history of human civilization.

以下是与英伟达CEO黄仁勋的对话，英伟达是人类文明史上最重要、最具影响力的公司之一。

NVIDIA is the engine powering the AI revolution, and a lot of its success can be directly attributed to Jensen's sheer force of will and his many brilliant bets and decisions as a leader, engineer, and innovator.

英伟达是推动AI革命的引擎，其成功很大程度上可以直接归功于黄仁勋作为领导者、工程师和创新者的强大意志力以及他许多出色的赌注和决策。

And now, a quick few second mention of his sponsor.

现在，快速用几秒钟提一下他的赞助商。

Check them out in the description or at lexfriedman.com/sponsors.

请在描述中或访问lexfriedman.com/sponsors查看他们。

It is, in fact, the best way to support this podcast.

事实上，这是支持本播客的最佳方式。

We got Shopify for selling stuff online, Element for electrolytes, Fin for customer service AI agents, Quo for a phone system like calls, texts, contacts for your business, and Perplexity for curiosity-driven knowledge exploration.

我们得到了Shopify用于在线销售，Element用于电解质，Fin用于客户服务AI代理，Quo用于像电话系统一样的通话、短信、联系人业务，以及Perplexity用于好奇心驱动的知识探索。

Choose Wazzle, my friends.

选择Wazzle，我的朋友们。

And now, on to the full ad reads.

现在，进入完整的广告阅读。

I try to make them interesting, but if you skip, please still check out our sponsors.

我尽量让它们有趣，但如果你跳过，请仍然查看我们的赞助商。

I enjoy their stuff. Maybe you will too.

我喜欢他们的东西。也许你也会喜欢。

To get in touch with me for whatever reason, go to lexfriedman.com/contact.

如需以任何理由联系我，请访问lexfriedman.com/contact。

All right, let's go.

好的，我们开始吧。

This episode is brought to you by Shopify, a platform designed for anyone to sell anywhere with a great looking online store.

本集由Shopify赞助，这是一个为任何人设计的平台，可以在任何地方销售，并拥有一个外观精美的在线商店。

Now, I know it's an incredible platform for selling stuff.

现在，我知道这是一个非常棒的销售平台。

It's a mechanism by which you can buy stuff on the internet.

它是一种机制，通过它你可以在互联网上购买东西。

But the thing I like to celebrate is the engineering.

但我喜欢庆祝的是工程。

They just recently tweeted about SimGym, which runs simulated shopping sessions by the hundreds of thousands daily.

他们最近刚刚在推特上提到了SimGym，它每天运行数十万次模拟购物会话。

I personally love the idea that things at scale, especially now with the LLM models, can be simulated.

我个人很喜欢这个想法：大规模的事物，尤其是现在有了大语言模型，可以被模拟。

You basically want to be simulating human behavior, human decision making, human choice.

你基本上想要模拟人类行为、人类决策、人类选择。

In this particular context, of course, is shopping.

在这个特定的情境下，当然就是购物。

It's really fascinating.

这真的很吸引人。

And they describe in their blog posts how they're leveraging NVIDIA stack to accomplish this task.

他们在博客文章中描述了如何利用NVIDIA技术栈来完成这项任务。

But you should know, in general, that you can sign up for a $1 per month trial period at shopify.com slash lex.

但你应该知道，一般来说，你可以在shopify.com/lex注册每月1美元的试用期。

That's all lowercase.

全部小写。

Go to shopify.com slash lex to take your business to the next level today.

今天就访问shopify.com/lex，把你的业务提升到新的水平。

This episode is also brought to you by Element, my daily zero-sugar delicious electrolyte mix

本期节目还由Element赞助，这是我每天喝的无糖美味电解质冲剂。

that, as far as I know, has very little to do with artificial intelligence and GPUs and CPUs and the revolution that we're experiencing in the tech sector.

据我所知，它与人工智能、GPU、CPU以及我们在科技领域经历的变革几乎没什么关系。

And I think that's beautiful because I got a chance to train a bunch of world-class fighters, wrestlers, grapplers recently.

我觉得这很美，因为我最近有机会训练了一批世界级的格斗选手、摔跤手和擒拿手。

I'm going to be traveling to parts of the world that doesn't really have much.

我将要去世界上一些资源不太丰富的地方旅行。

And I think in those parts of the world is where the mind can reconnect with the things that are truly important, that are truly timeless.

我认为正是在世界的那些地方，心灵才能重新连接那些真正重要、真正永恒的事物。

Anyway, in those parts of the world, I often get pretty out there in terms of physical strain and diet and dehydration and so on as an element.

总之，在世界的那些地方，我经常在体力消耗、饮食、脱水等方面做得很极端，作为一个元素。

One of the crucial things in my bag, really water and salt and really nice, Delicious, well-balanced salt, meaning sodium, potassium, magnesium, electrolytes.

我包里最关键的东西之一，真的是水和盐，还有非常好的、美味的、均衡的盐，意思是钠、钾、镁、电解质。

Element is my go-to. Watermelon salt, my favorite flavor.

Element是我的首选。西瓜盐，我最喜欢的口味。

Get a free eight-count sample pack with any purchase. Try it at drinkelement.com slash lex.

任何购买都可获得免费八份样品包。请在drinkelement.com/lex试用。

This episode was also brought to you by FIN, a powerful AI system that focuses on customer service.

本期节目还由FIN赞助，这是一个专注于客户服务的强大AI系统。

It's trusted by over 6,000 companies. It has a 65 average resolution rate.

它受到超过6000家公司的信任。它的平均解决率为65%。

and is built to handle complex, multi-step queries like returns, exchanges, and disputes.

并且能够处理复杂的多步骤查询，如退货、换货和争议。

This is such a fascinating problem because customer problems, the bulk of them, fall into a very specific set of categories,

这是一个非常有趣的问题，因为客户问题，其中大部分，都属于非常特定的类别，

but there's nuanced details within those categories that make all the difference.

但在这些类别中有细微的细节，这些细节会带来很大的不同。

And it can be an incredibly frustrating thing for a human being like myself, I swear. I promise. Definitely not a robot. Wouldn't tell you if I was.

对于像我这样的人类来说，这可能是一件非常令人沮丧的事情，我发誓。我保证。绝对不是机器人。如果我是，我也不会告诉你。

But it's frustrating for a human to come to the customer service process and to know that your problem kind of is like this problem.

但令人沮丧的是，人类来到客户服务流程中，知道你的问题有点像这个问题。

There's all these details you can provide about the system you're operating on,

你可以提供关于你正在操作的系统的大量细节，

the specifics of the puzzle you're trying to solve.

你试图解决的谜题的具体细节。

But there's details that you just know in your gut that this is important,

但有些细节你凭直觉就知道这很重要，

especially if you kind of thought through the problem.

尤其是如果你已经思考过这个问题。

I've been through this quite a bit.

我经历过不少这种情况。

You want to have some level of personalization that can get to the tricky aspect,

你希望有一定程度的个性化，能够触及棘手的方面，

the perspective on the problem that really would lead you down the road to a solution.

对问题的视角，真正能引导你走向解决方案。

Anyway, love this problem.

总之，我喜欢这个问题。

Really glad Finn is focusing on it.

很高兴Finn正在关注它。

Go to fin.ai.lex to learn more about transforming your customer service and scaling your support team.

访问fin.ai.lex，了解更多关于转变客户服务和扩展支持团队的信息。

That's fin.ai.lex.

那就是fin.ai.lex。

This episode is also brought to you by Quo, spelled Q-U-O,

本期节目还由Quo赞助，拼写为Q-U-O，

also known as a company with just three letters will win you a game of Scrabble.

也被称为一家只有三个字母的公司，会在拼字游戏中为你赢得一局。

That is not a joke.

这不是开玩笑。

It feels like a joke I have made before,

感觉像是我以前开过的玩笑，

but let's run with it.

但我们就继续吧。

It's a dad joke.

这是个老爸笑话。

It's a bad dad joke.

这是个糟糕的老爸笑话。

The only thing worse than a dad joke is a bad dad joke.

唯一比老爸笑话更糟的是糟糕的老爸笑话。

But here we go.

但我们开始吧。

Spelled Q-U-O, a business phone platform for calling and messaging.

拼写为Q-U-O，一个用于通话和消息的商业电话平台。

Basically, you have a bunch of people trying to help a larger group of people

基本上，你有一群人试图帮助更大的一群人，

and you want to orchestrate how they communicate with each other.

而你想协调他们如何相互沟通。

And this is just the system that does it extremely well, period.

而这正是那个把它做得极其出色的系统，就是这样。

Quo integrates AI into the whole shebang, organizing everything, generating summaries, highlighting the next steps, all that kind of stuff.

Quo 把 AI 整合进整个体系，组织一切、生成摘要、突出下一步，诸如此类的事情。

It just does it well.

它只是把它做得很好。

The interface on top of the AI is also really strong.

AI 之上的界面也非常强大。

So try Quo for free, plus get 20% off your first six months when you go to quo.com slash lex.

所以免费试用 Quo，另外当你访问 quo.com/lex 时，前六个月可享 20% 折扣。

That's q-u-o dot com slash Lex.

就是 q-u-o 点 com 斜杠 Lex。

This is the Lex Friedman Podcast.

这里是莱克斯·弗里德曼播客。

And now, dear friends, here's Jensen Huang.

现在，亲爱的朋友们，有请黄仁勋。

You've propelled NVIDIA into a new era in AI, moving beyond its focus on chip scale design to now rack scale design.

你推动英伟达进入了 AI 的新时代，从专注于芯片级设计，转向如今的机架级设计。

And I think it's fair to say that winning for NVIDIA for a long time used to be about building the best GPU possible.

我认为可以公平地说，长期以来英伟达的制胜之道曾经是打造尽可能最好的 GPU。

And you still do.

而你们仍然如此。

But now you've expanded that extreme co-design of GPU, CPU, memory, networking, storage, power, cooling, software, the rack itself, the pod that you've announced, and even the data center.

但现在你们已经扩展了那种极致的协同设计，涵盖 GPU、CPU、内存、网络、存储、电源、冷却、软件、机架本身、你们宣布的 pod，甚至数据中心。

So let's talk about extreme co-design.

所以让我们谈谈极致协同设计。

What is the hardest part of co-designing a system with that many complex components and design variables?

协同设计一个拥有那么多复杂组件和设计变量的系统，最难的部分是什么？

Yeah, thanks for that question.

是的，谢谢你的提问。

So first of all, the reason why extreme co-design is necessary is because the problem no longer fits inside one computer to be accelerated by one GPU.

首先，之所以需要极致协同设计，是因为问题不再能装进一台计算机里由一块 GPU 来加速。

The problem that you're trying to solve is you would like to go faster than the number of computers that you add.

你要解决的问题是，你希望速度提升超过你添加的计算机数量。

So you added 10,000 computers, but you would like it to go a million times faster.

所以你添加了1万台计算机，但你希望它快一百万倍。

Then all of a sudden, you have to take the algorithm, you have to break up the algorithm, you have to refactor it, you have to shard the pipeline, you have to shard the data, you have to shard the model.

然后突然之间，你必须采用算法，你必须拆分算法，你必须重构它，你必须分片流水线，你必须分片数据，你必须分片模型。

Now, all of a sudden, when you distribute the problem this way, not just scaling up the problem, but you're distributing the problem, then everything gets in the way.

现在，突然之间，当你以这种方式分配问题时，不仅仅是扩大问题规模，而是在分配问题，那么一切都会成为阻碍。

This is the Amdahl's law problem, where the amount of speed up you have for something depends on how much of the total workload it is.

这就是阿姆达尔定律问题，即某事物的加速量取决于它占总工作量的多少。

And so if computation represents 50% of the problem, and I sped up computation infinitely, like a million times, you know, I only sped up the total workload by a factor of two.

所以，如果计算占问题的50%，而我将计算速度无限提升，比如一百万倍，你知道，我实际上只将总工作量加速了两倍。

Now, all of a sudden, not only do you have to distribute the computation, you have to, you know, shard the pipeline somehow, you also have to solve the networking problem.

现在，突然之间，你不仅必须分配计算，你还必须，你知道，以某种方式分片流水线，你还必须解决网络问题。

Because you've got all of these computers are all connected together.

因为你所有这些计算机都连接在一起。

And so distributed computing at the scale that we do,

所以，在我们所做的这种规模下的分布式计算，

the CPU is a problem,

CPU 是个问题，

the GPU is a problem,

GPU 是个问题，

the networking is a problem,

网络是个问题，

the switching is a problem,

交换是个问题，

and distributing the workload across all these computers are a problem.

而将工作负载分配到所有这些计算机上也是个问题。

It's just a massively complex computer science problem.

这只是一个极其复杂的计算机科学问题。

And so we just got to bring every technology to bear.

所以我们必须动用一切技术。

Otherwise, we scale up linearly

否则，我们只能线性扩展，

or we scale up based on the capabilities of Moore's Law,

或者基于摩尔定律的能力来扩展，

which has largely slowed because Denard's scaling has slowed.

而摩尔定律已大幅放缓，因为登纳德缩放已经放缓。

I'm sure there's trade-offs there.

我相信这里面有权衡。

Plus, you have a complete disparate disciplines here.

此外，这里还有完全不同的学科。

I'm sure you have specialists in each one of these.

我相信你们每个领域都有专家。

High bandwidth memory, the network and the NVLink, the NICs, the optics and the copper that you're doing,

高带宽内存、网络和 NVLink、网卡、光学器件以及你们正在做的铜缆，

the power delivery, the cooling, all of that.

电力输送、冷却，所有这些。

I mean, there's like world experts in each of those.

我的意思是，每个领域都有世界级专家。

How do you get them in a room together to figure out?

你怎么让他们聚在一起想办法？

That's why my staff is so large.

这就是为什么我的团队这么大。

What's the process?

流程是什么？

Can you take me through the process of the specialists and the journalists?

你能带我了解一下专家和记者的流程吗？

Like, how do you put together the rack when you know the set of things you have to shove into a rack together?

比如，当你知道必须把哪些东西塞进一个机架时，你怎么把它们组装起来？

Like, what does that process look like of designing it all together?

比如，把所有东西一起设计的过程是什么样的？

There's the first question, which is, what is extreme co-design?

第一个问题是，什么是极致协同设计？

you're optimizing across the entire stack of software, from architectures to chips to systems to system software to the algorithms to the applications.

你要在整个软件栈中进行优化，从架构到芯片，到系统，到系统软件，到算法，再到应用。

That's one layer.

这是一层。

The second thing that you and I just talked about is it goes beyond CPUs and GPUs and networking chips and scale-up switches and scale-out switches.

第二件事，你和我刚刚谈到的，是它超越了CPU、GPU、网络芯片、纵向扩展交换机和横向扩展交换机。

And then, of course, you've got to include power and cooling and all of that because, you know, all these computers are extremely, extremely power-hungry.

当然，然后你还必须包括电力和冷却以及所有这些，因为，你知道，所有这些计算机都极其、极其耗电。

They do a lot of work and they're very energy efficient, but they in aggregate still consume a lot of power.

它们做了很多工作，而且非常节能，但总体上它们仍然消耗大量电力。

And so that's one.

所以这是一点。

The first question is, what is it?

第一个问题是，它是什么？

The second question is, why is it?

第二个问题是，为什么是它？

And we just spoke about the reason, you know, you want to distribute the workload so that you can exceed the benefit of just increasing the number of computers.

我们刚刚谈到了原因，你知道，你想分配工作负载，这样你就能超越仅仅增加计算机数量的好处。

And then the third question is, how is it? How do you do it?

然后第三个问题是，它是如何实现的？你怎么做？

And that's kind of the miracle of this company.

而这正是这家公司的奇迹所在。

When you're designing a computer, you have to have operating system of computers.

当你设计一台计算机时，你必须拥有计算机的操作系统。

When you're designing a company, you should first think about what is it that you want the company to produce?

当你设计一家公司时，你应该首先思考你希望这家公司产出什么？

You know, I see a lot of companies' organization charts, and they all look the same.

你知道，我看过很多公司的组织架构图，它们看起来都一样。

Hamburger organization charts, software organization charts, and car company organization charts, they all look the same.

汉堡公司的组织架构图、软件公司的组织架构图，还有汽车公司的组织架构图，它们看起来都一样。

And it doesn't make any sense to me.

而这在我看来毫无道理。

You know, the goal of a company is to be the machinery, the mechanism, the system that produces the output.

你知道，一家公司的目标就是成为生产产出的机器、机制和系统。

And that output is the product that we like to create.

而这个产出就是我们想要创造的产品。

It is also designed, the architecture of the company should reflect the environment by which it exists.

它也是被设计出来的，公司的架构应该反映它所处的环境。

It almost directly says what you should do with the organization.

它几乎直接说明了你应该如何安排这个组织。

My direct staff is 60 people.

我的直接下属有60人。

You know, I don't have one-on-ones with them because it's impossible.

你知道，我不和他们进行一对一谈话，因为那是不可能的。

You can't have 60 people on your staff if you're, you know, going to get work done.

如果你想把工作做完，你知道，你不可能有60个直接下属。

So you still have 60 reports. You still have more.

所以你仍然有60个汇报对象。你还有更多。

Yeah. More. Yeah.

是的。更多。是的。

And most stars at least have a foot in engineering. Almost all of them.

而且大多数明星至少都涉足工程领域。几乎所有人都是。

There's experts in memory. There's experts in CPUs.

有内存方面的专家。有CPU方面的专家。

experts in optical, all, yeah, GPUs and architecture, algorithms, design.

光学方面的专家，所有，是的，GPU和架构、算法、设计方面的专家。

So you constantly have an eye on the entire stack and you're having like intense discussions about the design of the entire stack.

所以你会一直关注整个技术栈，并且会就整个技术栈的设计进行激烈的讨论。

And no conversation is ever one person.

而且任何对话都不只是一个人。

That's why I don't do one-on-ones.

这就是我不做一对一谈话的原因。

We present a problem and all of us attack it.

我们提出一个问题，然后我们所有人一起攻克它。

You know, because we're doing extreme co-design and literally the company is doing extreme co-design all the time.

你知道，因为我们在做极致的协同设计，而且这家公司实际上一直在做极致的协同设计。

So even if you're talking about a particular component, like cooling, networking, everybody's listening in.

所以即使你在谈论某个特定的组件，比如散热、网络，每个人都在听。

Yeah, exactly. And they can contribute, well, this doesn't work for the power distribution.

是的，没错。他们可以贡献意见，嗯，这对配电不起作用。

This doesn't work for the memory.

这对内存不起作用。

This doesn't work for this. Exactly.

这对这个不起作用。没错。

And whoever wants to tune out, tune out.

谁想走神就走神吧。

You know what I'm saying? Yeah.

你明白我的意思吗？是的。

And the reason for that is because the people who are on the staff, they know when to pay attention.

而原因在于，在职的员工们知道什么时候该注意。

They're supposed, you know, something they could have contributed to, they didn't contribute to, I'm going to call them out, you know, and say, hey, come on, let's get in here.

他们本该，你知道，有些事他们本可以贡献却没贡献，我会点他们的名，你知道，然后说，嘿，来吧，我们一起来参与。

So, as you mentioned, NVIDIA is this company that's adapting to the environment.

所以，正如你提到的，NVIDIA 是一家适应环境的公司。

So, at which point can you say, did the environment change and begin adapting sort of secretly in the early days from GPU for gaming, maybe the early deep learning revolution, to we're now going to start thinking of it as an AI factory?

那么，在哪个时间点你可以说，环境发生了变化，并开始悄悄地适应，从早期用于游戏的 GPU，也许是早期的深度学习革命，到我们现在开始把它看作一个 AI 工厂？

What does NVIDIA do is produce AI, let's build a factory that makes AI.

NVIDIA 做的是生产 AI，让我们建一个制造 AI 的工厂。

I could reason through it just systematically.

我可以系统地推理一下。

We started out as an accelerator company. But the problem with accelerators is that the application domain is too narrow.

我们最初是一家加速器公司。但加速器的问题在于应用领域太窄了。

It has the benefit of being incredibly optimized for the job.

它的好处是，针对这项工作做了极其出色的优化。

You know, any specialist has that benefit.

你知道，任何专才都有这个好处。

The problem with intense specialization is that, of course, your market reach is narrower.

高度专业化的问​​题在于，当然，你的市场覆盖面会更窄。

But that's even fine.

但这甚至也没关系。

The problem is the market size also dictates your R&D capacity.

问题在于，市场规模也决定了你的研发能力。

And your R&D capacity ultimately dictates the influence and impact that you can possibly have in computing.

而你的研发能力最终决定了你在计算领域可能拥有的影响力和作用。

And so when we first started out as an accelerator, a very specific accelerator,

所以当我们最初作为加速器起步时，一个非常特定的加速器，

we always knew that that was going to be our first step.

我们一直知道那将是我们的第一步。

We had to find a way to become accelerated computing.

我们必须找到一条路，成为加速计算。

But the problem is when you become a computing company, it's too general purpose

但问题是，当你成为一家计算公司时，它就太通用了，

and it takes away from your specialization.

这会削弱你的专业化。

I connected two words that actually have fundamental tension.

我把两个实际上存在根本性张力的词联系在了一起。

The better computing company we become, the worse we become as a specialist.

我们越是成为一家更好的计算公司，作为专才我们就越差。

The more of a specialist, the less capacity we have to do overall computing.

越是专才，我们做整体计算的能力就越少。

And so that, and I connected those two words together on purpose,

所以，我是故意把那两个词联系在一起的，

that the company has to find that really narrow path,

公司必须找到那条非常狭窄的路径，

step by step by step to expand our aperture of computing,

一步一步地扩大我们的计算视野，

but not give up on the most important specialization that we had.

但不能放弃我们最重要的专业化。

Okay.

好的。

So the first step that we took beyond acceleration was we invented the programmable pixel shader.

所以我们迈出加速之外的第一步，就是发明了可编程像素着色器。

So that was the first step towards programmability.

所以那是迈向可编程性的第一步。

You know, it was our first journey towards moving into the world of computing.

你知道，那是我们迈向计算世界的第一段旅程。

The second thing that we did was we created, we put FP32 into our shaders.

我们做的第二件事是，我们创建了，我们把FP32放进了我们的着色器。

That FP32 step, IEEE compatible FP32, was a huge step in the direction of computing.

那个FP32步骤，即IEEE兼容的FP32，是朝着计算方向迈出的一大步。

It was the reason why all of the people who were working on stream processors and, you know, other types of data flow processors discovered us.

这就是为什么所有研究流处理器以及你知道的其他类型数据流处理器的人发现了我们。

And they say, hey, all of a sudden, you know, we might be able to use this GPU that's incredibly computationally intensive.

他们说，嘿，突然之间，你知道，我们也许能够使用这个计算强度极高的GPU。

And it's now, you know, compliant with IEEE.

而且它现在，你知道，符合IEEE标准。

I can take my software that I was writing, you know, previously on CPUs and I can, you know, see about, you know, using the GPU for them.

我可以把我以前在CPU上编写的软件，你知道，拿来，并且我可以，你知道，考虑，你知道，为它们使用GPU。

And which led us to create, put C on top of FP32, what's called, we call CG.

这引导我们创建了，把C放在FP32之上，所谓的，我们称之为CG。

That CG path took us to eventually CUDA.

那条CG路径最终把我们带到了CUDA。

CUDA, step by step by step.

CUDA，一步一步地。

Well, putting CUDA on GeForce, that was a strategic decision

嗯，把CUDA放到GeForce上，那是一个战略决策

that was very, very hard to do because it cost the company enormous amounts of our profits and we couldn't afford it at the time.

那个决策非常非常难做，因为它耗费了公司大量的利润，而我们当时负担不起。

But we did it anyways because we wanted to be a computing company.

但我们还是做了，因为我们想成为一家计算公司。

A computing company has a computing architecture.

一家计算公司要有计算架构。

A computing architecture has to be compatible across all of the chips that we build.

计算架构必须在我们生产的所有芯片上兼容。

Can you take me to that decision?

你能带我回顾一下那个决策吗？

So putting CUDA on GeForce could not afford to do.

所以把CUDA放到GeForce上，我们当时负担不起。

Can you explain that decision?

你能解释一下那个决策吗？

Why boldly choose to do that anyway?

为什么还是大胆地选择那样做？

Can you explain that decision?

你能解释一下那个决策吗？

Excellent. That was the first... I would say that that was the first...

太好了。那是第一个……我想说那是第一个……

The first strategic decision that is as close to an existential threat.

第一个战略决策，几乎接近生存威胁。

For people who don't know, it turned out to be, spoiler alert, one of the most incredibly brilliant decisions ever made by a company.

对于不知道的人来说，结果证明，剧透一下，这是公司有史以来最极其明智的决策之一。

So CUDA turned out to be an incredible foundation for computation in this AI infrastructure world.

所以CUDA结果成为了这个AI基础设施世界中令人难以置信的计算基础。

So just setting the context, it turned out to be a good decision.

所以只是设定一下背景，结果证明这是一个好决策。

Yeah, it turned out to have been a good decision.

是的，结果证明这是一个好决策。

So here's the way it went.

所以事情是这样的。

So we invented this thing called CUDA, and it expanded the aperture of applications that we can accelerate with our accelerator.

所以我们发明了这个叫CUDA的东西，它扩大了我们可以用加速器加速的应用范围。

The question is, how do we attract developers to CUDA?

问题是，我们如何吸引开发者使用CUDA？

Because a computing platform is all about developers.

因为一个计算平台的核心就是开发者。

And developers don't come to a computing platform just because it could perform something interesting.

开发者不会仅仅因为一个计算平台能执行一些有趣的事情就选择它。

They come to a computing platform because the install base is large.

他们选择某个计算平台是因为它的安装基数很大。

Because a developer like anybody else wants to develop software that reaches a lot of people.

因为开发者和其他人一样，希望开发的软件能触及很多人。

So the install base is in fact the single most important part of an architecture.

所以安装基数实际上是架构中最重要的部分。

The architecture could attract enormous amounts of criticism.

这个架构可能会招致大量的批评。

For example, no architecture has ever attracted more criticism than the x86.

例如，没有哪个架构比x86招致过更多的批评。

You know, as a less than elegant architecture.

你知道，作为一个不那么优雅的架构。

But yet, it is the defining architecture of today.

但尽管如此，它是当今具有决定性意义的架构。

It gives you an example that, in fact, so many risk architectures, which were beautifully architected, incredibly well designed by some of the brightest computer scientists in the world, largely failed.

它给你一个例子，事实上，许多RISC架构，由世界上最聪明的一些计算机科学家精心设计、架构优美，却大多失败了。

And so I've given you two examples where one is elegant, the other one's barely aesthetic.

所以我给了你两个例子，一个优雅，另一个几乎谈不上美感。

And so, yet x86 survived.

然而，x86却存活了下来。

Install base is everything.

安装基数就是一切。

Install base defines an architecture.

安装基数定义了一个架构。

Not everything else is secondary.

并非其他一切都是次要的。

Okay. And so there were other architectures at the time.

好的。所以当时还有其他架构。

CUDA came out. OpenCL was here.

CUDA出现了。OpenCL已经存在。

There were, you know, there's several other competing architectures.

你知道，还有几个其他竞争架构。

But the thing that the decision that we made that was good was we said, hey, look, ultimately, it's about install base.

但我们做出的那个好的决定是，我们说，嘿，看，归根结底，关键在于安装基数。

And what is the best way we could get a new computing architecture into the world?

那么，我们把一种新的计算架构推向世界的最佳方式是什么？

By that timeframe, GeForce had become successful.

到那个时候，GeForce 已经取得了成功。

We were already selling millions and millions of GeForce GPUs a year.

我们每年已经卖出数百万块 GeForce GPU。

And we said, you know, we ought to put CUDA on GeForce and put it into every single PC, whether customers use it or not.

我们说，你知道，我们应该把 CUDA 放到 GeForce 上，并把它装进每一台 PC，不管客户用不用。

And use it as a starting point of cultivating our install base.

并把它作为培养我们装机量的起点。

Meanwhile, we'll go and attract developers and went to universities and wrote books and taught classes and put CUDA everywhere.

与此同时，我们会去吸引开发者，去大学，写书，教课，把 CUDA 推广到各处。

And eventually people will discover, and at the time the PC was the primary computing vehicle.

最终人们会发现，当时 PC 是主要的计算工具。

There was no cloud and we could put a supercomputer in the hands of every researcher in school,

没有云，我们可以把超级计算机交到学校里每一位研究人员手中，

every scientist, every engineer in school or every student in school.

每一位科学家，学校里的每一位工程师或每一位学生。

And eventually something amazing will happen.

最终，一些惊人的事情将会发生。

Well, the problem was CUDA increased our cost of that GPU, which is a consumer product, so tremendously, it completely consumed all of the company's gross profit dollars.

嗯，问题是 CUDA 极大地增加了我们那块 GPU 的成本，而 GPU 是消费产品，所以它完全消耗了公司所有的毛利润。

And so at the time, the company was probably worth, I don't know, at the time, $8 billion or something, $6-7 billion or something like that.

所以当时，公司大概值，我不知道，当时，80 亿美元左右，67 亿美元或类似这样。

After we launched CUDA, I recognized that it was going to add so much cost, but it was something we believed in.

在我们推出 CUDA 之后，我意识到它会增加很多成本，但这是我们相信的事情。

You know, our market cap went down to like one and a half billion dollars.

你知道，我们的市值跌到了大约十五亿美元。

And so we were down there for a while and we clawed our way back slowly.

所以我们在那里待了一段时间，然后我们慢慢地爬了回来。

But we carried CUDA on GeForce.

但我们在GeForce上搭载了CUDA。

I always say that NVIDIA is the house that GeForce built because it was GeForce that took CUDA out to everybody.

我总是说，NVIDIA是GeForce建造的房子，因为是GeForce把CUDA带给了所有人。

Researchers, scientists, they discovered CUDA on GeForce because they were all, you know, many of them were gamers.

研究人员、科学家们，他们在GeForce上发现了CUDA，因为他们都是，你知道，其中很多是游戏玩家。

Many of them built their own PCs anyways.

他们中的许多人反正都自己组装电脑。

In a university lab, many of them built clusters themselves, you know, using PC components.

在大学实验室里，他们中的许多人自己用PC组件搭建了集群。

And so that, you know, that's kind of how we got going.

所以，你知道，我们就是这样起步的。

And then that became the platform, the foundation for the deep learning revolution.

然后这就成了平台，深度学习革命的基础。

That was also another great, great observation.

那也是另一个非常、非常棒的观察。

Yeah. That existential moment, do you remember what were those meetings like?

是的。那个生死攸关的时刻，你还记得那些会议是什么样的吗？

What were those discussions like deciding as a company, risking everything?

那些讨论是什么样的，作为一家公司决定冒一切风险？

Well, I had to make it clear to the board what we were trying to do.

嗯，我必须向董事会明确我们正在努力做什么。

And the management team knew our gross margins were to get crushed.

管理团队知道我们的毛利率会被压垮。

So you could imagine a world where GeForce would carry the burden of CUDA,

所以你可以想象一个GeForce承担CUDA负担的世界，

And none of the gamers would appreciate it.

而游戏玩家们都不会感激它。

And none of the gamers would pay for it.

而游戏玩家们也不会为它付钱。

You know, they only pay certain price and it doesn't matter what your cost is.

你知道，他们只支付一定的价格，而你的成本是多少并不重要。

And so, you know, we increased our cost by 50%.

所以，你知道，我们把成本增加了50%。

And that consumed, and we were a 35% gross margin company.

而那消耗了，我们是一家35%毛利率的公司。

And so it was quite a difficult decision to make.

所以这是一个相当艰难的决定。

But you could imagine that someday this would go into workstations and it would go into supercomputers.

但你可以想象，有一天这会进入工作站，也会进入超级计算机。

And in those segments, maybe we can capture more margin.

而在那些细分市场中，也许我们可以获得更多利润。

So you could reason your way into being able to afford this, but it still took a decade.

所以你可以通过推理来让自己负担得起这个，但它仍然花了十年。

But that's more of a conversation with the board convincing them, but you psychologically.

但这更多是与董事会交谈说服他们，但你在心理上。

Because NVIDIA has continued to make bold bets that predict the future and in part, especially now, define the future.

因为NVIDIA一直在进行大胆的押注，预测未来，并且部分地，尤其是现在，定义未来。

So I'm almost looking for wisdom about how you're able to make those decisions to make leaps like that as a company.

所以我几乎在寻找智慧，关于你如何能够做出那些决定，像那样作为一家公司实现飞跃。

Well, first of all, I'm informed by a lot of curiosity.

嗯，首先，我受到很多好奇心的启发。

At some point, there's a reasoning system that convinces me so clearly this outcome will happen.

在某个时刻，有一个推理系统如此清晰地说服我，这个结果将会发生。

That this will happen.

这将会发生。

And so I believe it in my mind.

所以我在心里相信它。

And when I believe it in my mind, you know how it is.

当我在心里相信它时，你知道那是怎样的。

you manifest a future and that future is so convincing.

你显化一个未来，而那个未来如此令人信服。

There's no way it won't happen.

它不可能不发生。

There's a lot of suffering in between, but you've got to believe what you believe.

中间会有很多痛苦，但你必须相信你所相信的。

So you envision the future. Yeah.

所以你设想未来。是的。

And you essentially, from a sort of engineering perspective, manifest it. Yeah.

而你本质上，从某种工程角度，显化它。是的。

And you reason about how to get there.

而你推理如何到达那里。

You reason about why it must exist.

你会推理它为什么必须存在。

And, you know, we all reason it here.

而且，你知道，我们都在这里推理它。

The management team will reason about it.

管理团队会推理它。

All the people that we spent a lot of time reasoning about it.

我们花了大量时间推理它的所有人。

The thing that the next part of it is probably a skill thing,

它的下一部分可能是一种技能问题，

which is, you know, oftentimes in leadership,

也就是说，你知道，在领导力中常常是，

the leadership stays quiet or they learn about something and then they do some manifesto.

领导者保持沉默，或者他们了解到某些事情，然后他们做一些宣言。

And it's a brand new year.

这是一个全新的一年。

And somehow at the end of the year, next year, we're going to have a brand new plan,

不知怎么的，在年底，明年，我们会有一个全新的计划，

big, huge layoff this way, big, huge organization change this way,

大规模裁员这边，大规模组织变革那边，

new mission statement, brand new logos, you know, that kind of stuff.

新的使命宣言，全新的标志，你知道，那种东西。

We've just never, I never do things that way.

我们从来没有，我从不那样做事。

When I learn about something and it's starting to influence how I think,

当我了解到某些事情，并且它开始影响我的思考方式时，

I'll make it very clear to everybody near me that, you know, this is interesting.

我会向我身边的每个人明确表示，你知道，这很有趣。

This is going to make a difference.

这将会带来改变。

This is going to impact that.

这将会影响那个。

And I reason about things step by step by step.

我会一步一步地推理事情。

Oftentimes, I've already made up my mind, but I'll take every possible opportunity,

通常，我已经下定决心，但我会抓住每一个可能的机会，

external information, new insights, new discoveries, new engineering revelations, new milestones developed.

外部信息、新见解、新发现、新的工程启示、新的里程碑。

I'll take those opportunities and I'll use it to shape everybody else's belief system.

我会抓住这些机会，并用它来塑造其他所有人的信念体系。

And I'm doing that literally every single day.

我几乎每一天都在这样做。

I'm doing that with my board.

我在和我的董事会这样做。

I'm doing that with my management team.

我在和我的管理团队这样做。

I'm doing that with my employees.

我在和我的员工这样做。

I'm trying to shape their belief system such that when I come the day I say, hey, let's buy Mellanox.

我试图塑造他们的信念体系，以至于当我有一天说，嘿，让我们收购Mellanox。

It's completely obvious to everybody that we absolutely should.

对每个人来说都完全显而易见，我们绝对应该这样做。

On the day that I said, hey guys, let's go all in on deep learning.

在我说的那一天，嘿伙计们，让我们全力投入深度学习。

And let me tell you why.

让我告诉你为什么。

I've already been laying down the bricks to different organizations inside the company.

我已经在为公司内部的不同组织铺路。

Every organization and everybody, many of the people might have heard everything.

每个组织和每个人，许多人可能已经听说了一切。

Most of the company heard, of course, pieces of it.

当然，公司里大多数人听到了一些片段。

And on the day that I announce it, everybody's kind of bought into many pieces of it.

而在我宣布的那天，大家已经接受了其中的许多部分。

And in a lot of ways, I like to announce these things.

在很多方面，我喜欢宣布这些事情。

And I imagine that the employees are kind of saying, you know, Jensen, what took you so long?

我想员工们大概会说，你知道，Jensen，怎么花了这么长时间？

And in fact, I've been shaping their belief system for some time.

事实上，我一直在塑造他们的信念体系有一段时间了。

And therefore, leadership, sometimes it looks like you're leading from behind.

因此，领导力，有时看起来你是在从后面领导。

But you've been shaping their, you know, to the point where on the day that I declared it, 100% buy-in.

但你一直在塑造他们的，你知道，以至于在我宣布的那天，100%的认同。

But that's what you want. You want to bring everybody along.

但这就是你想要的。你想让每个人都跟上。

You know, otherwise, we announce something about deep learning and everybody goes, what are you talking about?

你知道，否则，我们宣布一些关于深度学习的事情，每个人都会说，你在说什么？

You know, you announce something about let's go all in on this thing.

你知道，你宣布一些关于让我们全力投入这件事的事情。

And your management team, your board, your employees, your customers, they're kind of like, where's this coming from?

而你的管理团队、你的董事会、你的员工、你的客户，他们有点像是，这是从哪来的？

You know, this is insane.

你知道，这太疯狂了。

And so GTC, in fact, if you go back in time, you look at the keynotes.

所以GTC，事实上，如果你回顾过去，看看那些主题演讲。

I'm also shaping the belief system of my partners in the industry.

我也在塑造行业内合作伙伴的信念体系。

And I'm using that to shape, you know, the belief system of my own employees.

我用这一点来塑造，你知道，我自己员工的信念体系。

And so by the time that I announce something, like for example, we just announced Grok.

所以等到我宣布某件事的时候，比如，我们刚刚宣布了Grok。

I've been talking about the stepping stones for two and a half years.

我已经谈论这些垫脚石两年半了。

You guys just go back and go, oh my gosh, they've been talking about it for two and a half years.

你们回过头去看，然后说，天哪，他们已经谈论这个两年半了。

And so I've been laying the foundation step by step by step.

所以我一直在一步一步地打下基础。

So when the time comes, you announce it, everybody's like, what took you so long?

所以当时候到了，你宣布它，每个人都像，你怎么花了这么长时间？

But it's not just inside the company.

但这不仅仅是在公司内部。

You're shaping the landscape, the broader global landscape of innovation.

你在塑造格局，更广泛的全球创新格局。

Like putting those ideas out there, you really are manifesting reality.

就像把那些想法公之于众，你真的在显化现实。

We don't build computers. We actually don't build clouds.

我们不制造电脑。我们实际上也不制造云。

We don't, as it turns out, we're a computing platform company.

我们不，事实证明，我们是一家计算平台公司。

And so nobody can buy anything from us.

所以没有人能从我们这里买到任何东西。

That's the weird thing.

这就是奇怪的地方。

You know, we vertically design, vertically integrate to design and optimize.

你知道，我们垂直设计，垂直整合来设计和优化。

But then we open up the entire platform at every single layer to be integrated into other companies' products and services and clouds and supercomputers and OEM computers.

但然后我们在每一层都开放整个平台，以便整合到其他公司的产品和服务、云、超级计算机和OEM电脑中。

And so the amazing thing is I can't do what I do without having convinced them first.

所以神奇的是，如果我没有先说服他们，我就无法做我所做的事。

And so most of GTC is about manifesting a future that by the time that my product is ready, they're going, what took you so long?

所以GTC的大部分内容都是在展现一个未来：等到我的产品准备好时，他们会说，你怎么花了这么长时间？

Yeah. So one of the things you've been a believer for a long time is scaling laws broadly defined.

是的。所以你长期以来一直相信的一件事就是广义定义的缩放定律。

So are you still a believer in the scaling laws?

那么你现在还相信缩放定律吗？

Yeah, we have more scaling laws now.

是的，我们现在有更多的缩放定律。

So I think you've outlined four of them with pre-training, post-training, test time, and agentic scaling.

所以我认为你概述了其中四个：预训练、后训练、测试时和智能体缩放。

What do you think, when you think about the future, deep future and the near-term future, what are the blockers that you're most concerned about to keep you up at night, that you have to overcome in order to keep scaling?

当你思考未来，远期未来和近期未来时，你认为最让你夜不能寐、为了继续扩展必须克服的障碍是什么？

Well, we can go back and reflect on what people thought were blockers.

嗯，我们可以回顾一下人们过去认为的障碍是什么。

So in the beginning, we were the first, the pre-training scaling law.

所以一开始，我们是第一个，预训练缩放定律。

You know, people thought, well, rightfully so, that the amount of data that we have, high quality data that we have, will limit the intelligence that we achieve.

你知道，人们认为，嗯，理所当然地，我们拥有的数据量，我们拥有的高质量数据，将限制我们实现的智能。

And that scaling law was an important, very important scaling law.

而那个缩放定律是一个重要的、非常重要的缩放定律。

The larger the model, the correspondingly more data results in a better, results in a smarter AI.

模型越大，相应地更多数据会带来更好的、更智能的AI。

And so that was pre-training.

所以那就是预训练。

And Ilya, Suscover, Ilya said, we're out of data or something like that. Pre-training is over or something like that.

而Ilya，Suscover，Ilya说，我们没数据了或类似的话。预训练结束了或类似的话。

The industry panicked, you know, that this is the end of AI.

业界恐慌了，你知道，认为这就是AI的终结。

And of course, of course, that's obviously not true.

当然，当然，那显然不是真的。

We're going to keep on scaling the amount of data that we have to train with.

我们将继续扩大我们必须用来训练的数据量。

A lot of that data is probably going to be synthetic.

很多数据很可能是合成数据。

And that also confused people, you know.

这也让人们感到困惑，你知道。

And what people don't realize is they've kind of forgotten that most of the data that we are training, that we teach each other with, inform each other with, is synthetic.

而人们没有意识到的是，他们有点忘记了，我们用来训练、用来互相教导、互相告知的大部分数据都是合成的。

You know, it's synthetic because it didn't come out of nature.

你知道，它是合成的，因为它不是来自自然。

You created it. I'm consuming it.

你创造了它。我在消费它。

I modify it, augment it. I regenerate it.

我修改它，增强它。我重新生成它。

Somebody else consumes it.

其他人消费它。

And so we've now reached a level where AI is able to take ground truth, augment it, enhance it, synthetically generate an enormous amount of data, and that part of post-training continues to scale.

所以我们现在已经达到了一个水平，AI能够获取真实数据，增强它，改进它，合成生成大量数据，而训练后阶段的那部分继续扩展。

And so the amount of data that we could use that is human-generated will be smaller and smaller and smaller.

因此，我们可以使用的、由人类生成的数据量将越来越小。

The amount of data that we use to train model is going to continue to scale to the point where we're no longer limited.

我们用来训练模型的数据量将继续扩展，直到我们不再受限。

Training is no longer limited by data.

训练不再受数据限制。

It's now limited by compute.

现在它受计算能力限制。

And the reason for that is most of the data is synthetic.

原因是大部分数据是合成的。

Then the next phase is test time.

然后下一个阶段是测试时间。

And I still remember people telling me that inference, oh yeah, that's easy.

我仍然记得人们告诉我，推理，哦是的，那很容易。

Pre-training, that's hard.

预训练，那很难。

These are giant systems that people are talking about.

这些是人们正在谈论的巨型系统。

Inference must be easy.

推理一定很容易。

And so inference chips are going to be little tiny chips.

所以推理芯片将会是小小的芯片。

And they're not like NVIDIA's chips.

而且它们不像英伟达的芯片。

Oh, those are going to be complicated and expensive.

哦，那些会变得又复杂又昂贵。

And we could make, and this is, in the future, inference is going to be the biggest market.

而且我们可以制造，而且这就是，在未来，推理将成为最大的市场。

And it's going to be easy.

而且它会变得容易。

And we're going to commoditize.

而且我们会把它商品化。

And everybody can build their own chips.

而且每个人都可以制造自己的芯片。

And that was always illogical to me

而这一直让我觉得不合逻辑，

because inference is thinking.

因为推理就是思考。

And I think thinking is hard.

而我认为思考是困难的。

Thinking is way harder than reading.

思考比阅读难得多。

You know, pre-training is just memorization and generalization, you know, and looking for patterns in relationships.

你知道，预训练只是记忆和泛化，你知道，以及寻找关系中的模式。

You're reading versus thinking, reasoning, solving problems, taking unexplored experiences, new experiences, and breaking it down into decomposing it into solvable pieces that we then go off either through first principle reasoning or through previous examples, prior experiences, or just exploration.

你是在阅读，而思考、推理、解决问题、接受未探索的体验、新体验，并把它分解成可解决的碎片，然后我们通过第一性原理推理，或者通过以前的例子、先前的经验，或者只是探索。

and search and, you know, trying different things.

以及搜索，你知道，尝试不同的东西。

And that whole process of test time scaling, inference is really about thinking.

而整个测试时扩展的过程，推理，实际上就是关于思考。

And it's about reasoning.

它是关于推理的。

It's about planning.

它是关于规划的。

It's about search.

它是关于搜索的。

It's about, and so how could that possibly be compute light?

它是关于，那么这怎么可能计算量小呢？

And we were absolutely right about that.

而我们对此绝对正确。

You know, so test time scaling is intensely compute intensive.

你知道，所以测试时扩展是极其计算密集型的。

Then the question is, okay, now we're at inference and we're at test time scaling.

那么问题就是，好吧，现在我们处于推理阶段，我们处于测试时扩展阶段。

What's beyond that?

那之后是什么？

Well, obviously, we have now created one agentic person.

嗯，显然，我们现在创造了一个代理型的人。

And that one agentic person has a large language model that we've now developed.

而那一个代理型的人拥有一个我们现在已经开发出的大型语言模型。

But during test time, that agentic system goes off and does research and bangs on databases and uses tools.

但在测试时，那个智能体系统会出去做研究，敲打数据库并使用工具。

And one of the most important things it does is spins off and spawns off a whole bunch of sub-agents, which means we're now creating large teams.

它做的最重要的事情之一就是衍生并生成一大堆子智能体，这意味着我们现在正在创建大型团队。

It's so much easier to scale NVIDIA by hiring more employees than it is to scale myself.

通过雇佣更多员工来扩展英伟达，比扩展我自己要容易得多。

And so the next scaling law is the agentic scaling law.

所以下一个扩展定律就是智能体扩展定律。

It's kind of like multiplying AI.

这有点像在倍增AI。

Multiplying AI, we could spin off agents as fast as you want to spin off agents.

倍增AI，我们可以以你想要的速度快速衍生出智能体。

And so, you know, you and I have four scaling laws.

所以，你知道，你和我有四个扩展定律。

And as we use the agentic systems, they're going to create a lot more data.

随着我们使用智能体系统，它们将创造更多的数据。

They're going to create a lot of experiences.

它们将创造大量的经验。

Some of it we're going to say, wow, this is really good. We ought to memorize this.

其中一些我们会说，哇，这真的很好。我们应该记住这个。

That data set then comes all the way back to pre-training.

那个数据集然后会一路回到预训练阶段。

We memorize and generalize it.

我们记住并泛化它。

We then refine it and fine tune it back into post-training.

然后我们改进它，并微调回后训练阶段。

Then we enhance it even more with test time, you know, in the agentic systems, you know, put it out into the industry.

然后我们通过测试时进一步增强它，你知道，在智能体系统中，你知道，把它投入产业。

And so this loop, the cycle is going to go on and on and on.

所以这个循环，这个周期将会一直持续下去。

It kind of comes down to basically intelligence is going to scale by one thing and it's compute.

这归根结底基本上就是智能将靠一件事来扩展，那就是算力。

But there's a tricky thing there that you have to anticipate and predict,

但那里有个棘手的地方，你必须预判和预测，

which is some of these components, it requires different kind of hardware to really do it optimally.

也就是其中一些组件，需要不同种类的硬件才能真正做到最优。

So you have to anticipate where the AI innovation is going to lead.

所以你必须预判AI创新会走向哪里。

For example, a mixture of experts with sparsity.

比如，带稀疏性的专家混合模型。

With hardware, you can't just pivot on a week's notice.

在硬件方面，你不能只提前一周就转向。

You have to anticipate what that's going to look like.

你必须预判那会是什么样子。

So good. That's so scary and difficult to do, right?

太好了。这做起来太可怕、太难了，对吧？

For example, these AI model architectures are being invented about once every six months.

比如，这些AI模型架构大约每六个月就会被发明出一种。

Yeah. Right?

是啊。对吧？

And system architectures and hardware architectures kind of every three years.

而系统架构和硬件架构差不多每三年一次。

And so you need to anticipate what likely is going to happen, you know, two, three years from now.

所以你需要预判，你知道，从现在起两三年后可能会发生什么。

And there's a couple of ways that you could do that.

有几种方法可以做到这一点。

First of all, we could do research internally ourselves.

首先，我们可以在内部自己做研究。

And that's one of the reasons why we have basic research.

这就是我们拥有基础研究的原因之一。

We have applied research. We create our own models.

我们有应用研究。我们创建自己的模型。

And so we have hands-on life experience right here.

所以我们在第一线有亲身的实践经验。

This is part of the co-design that I'm talking about.

这就是我所说的协同设计的一部分。

We're also the only AI company in the world that works with literally every AI company in the world.

我们也是世界上唯一一家与世界上几乎所有AI公司都有合作的AI公司。

And to the extent that we can, we try to get a sense of what are the challenges that people are experiencing.

在力所能及的范围内，我们努力了解人们正在经历哪些挑战。

So you're listening to the whispers across the industry, the AI labs?

所以你在倾听整个行业、各个AI实验室的私语？

That's right. You've got to listen and learn from everybody.

没错。你必须倾听并向每个人学习。

And then the last part is to have an architecture that's flexible, that can adapt and move with the wind.

然后最后一部分是拥有一个灵活的架构，能够适应并随风而动。

And one of the benefits of CUDA is that it's, you know, on the one hand, an incredible accelerator.

CUDA 的好处之一是，你知道，一方面，它是一个令人难以置信的加速器。

On the other hand, it's really flexible.

另一方面，它非常灵活。

And so that balance, incredible balance between specialization, otherwise we can't accelerate the CPU,

所以这种平衡，在专业化之间的惊人平衡，否则我们无法加速 CPU，

versus generalization so that we can adapt with changing algorithms, that's really, really important.

与通用化之间的平衡，以便我们能够适应不断变化的算法，这真的非常重要。

That's the reason why CUDA has been so resilient on the one hand,

这就是为什么 CUDA 一方面如此有韧性，

and yet we continue to enhance it. We're at CUDA 13.2, and so we're evolving the architecture so fast that we can stay with the modern algorithms.

同时我们继续增强它。我们现在是 CUDA 13.2，所以我们如此快速地演进架构，以便能够跟上现代算法。

For example, when Mixture Experts came out,

例如，当混合专家模型出现时，

That's the reason why we had NVLink 72 instead of NVLink 8.

这就是为什么我们有 NVLink 72 而不是 NVLink 8。

We could now take an entire 4 trillion, 10 trillion parameter model and put it in one computing domain as if it's running on one GPU.

我们现在可以把一个完整的 4 万亿、10 万亿参数的模型放在一个计算域中，就像它在一个 GPU 上运行一样。

People probably didn't notice, I said it, but if you look at the architecture of the Grace Blackwell racks, it was completely focused on doing one thing, processing the LLM.

人们可能没有注意到，我说过，但如果你看看 Grace Blackwell 机架的架构，它完全专注于做一件事，处理 LLM。

All of a sudden, one year later, you're looking at a Vera Rubin rack.

突然之间，一年后，你看到的是一个 Vera Rubin 机架。

It has storage accelerators. It has this incredible new CPU called Vera.

它有存储加速器。它有一个令人难以置信的新 CPU，叫做 Vera。

It has Vera Rubin and MVLink 72 to run the LLMs.

它有 Vera Rubin 和 MVLink 72 来运行 LLM。

It also has this new additional rack called GROC.

它还有一个新的附加机架，叫做 GROC。

And so this entire rack system is completely different than the previous one.

所以这整个机架系统与之前那个完全不同。

And it's got all these new components in it.

而且它里面装了所有这些新组件。

And the reason for that is because the last one was designed to run MOE, large language models, inference.

原因在于，上一个机架是为了运行MOE、大语言模型的推理而设计的。

And this one is to run agents.

而这一个是为了运行智能体。

And agents bang on tools.

而智能体会频繁使用工具。

Obviously, the design of the system had to have been done before Clawed Code, Codex, Open Claw.

显然，这个系统的设计必须在Clawed Code、Codex、Open Claw之前就完成。

You were anticipating the future, essentially.

你基本上是在预见未来。

And that comes from what?

那这来自什么呢？

From the whispers, from the understanding of what all the state of the art is doing?

来自那些低语，来自对最先进技术都在做什么的理解？

No, it's easier than that.

不，比那更简单。

You just reason about it.

你只需要推理一下。

First of all, you just reason.

首先，你只需要推理。

No matter what happens at some point, in order for that large language model to be a digital worker,

无论某个时候发生什么，为了让那个大语言模型成为数字工作者，

let's just use that metaphor.

我们就用这个比喻吧。

Let's say that we want the LLM to be a digital worker.

假设我们想让LLM成为一个数字工作者。

What does it have to do?

它必须做什么？

It has to access ground truth.

它必须访问真实数据。

That's our file system.

那就是我们的文件系统。

It has to be able to do research.

它必须能够做研究。

It doesn't know everything.

它并非无所不知。

And I don't want to wait until this AI becomes universally smart about everything past, present, and future before I make it useful.

我不想等到这个AI对过去、现在和未来的一切都变得无所不知之后，才让它变得有用。

And so therefore, I might as well let it go do research.

所以因此，我不如让它去做研究。

It's obviously, if it wants to help me, it's got to use my tools.

显然，如果它想帮我，它就得使用我的工具。

A lot of people would say, AI is going to completely destroy software.

很多人会说，AI将彻底摧毁软件。

We don't need software anymore. We don't even need tools anymore.

我们不再需要软件了。我们甚至不再需要工具了。

That's ridiculous.

这太荒谬了。

Let's use a thought experiment.

让我们做个思想实验。

And you could just sit there, enjoy a glass of whiskey and think about all these things and it would become completely obvious.

你可以就坐在那里，享受一杯威士忌，思考所有这些事情，然后一切就会变得完全显而易见。

Like if I were to create the most amazing agent that we can imagine in the next 10 years, let's say it'd be a human or robot.

就像如果我要创造出我们在未来10年里能想象到的最惊人的智能体，比如说它会是个人类或机器人。

If that human or robot were to be created, is it more likely that the human or robot comes into my house and uses the tools that I have to do the work that it needs to do?

如果那个人类或机器人被创造出来，更有可能的是那个人类或机器人来到我家，使用我拥有的工具来完成它需要做的工作吗？

Or does this hand turns into a 10-pound hammer in one instance, turns into a scalpel in another instance, and in order to boil water, it beams, you know, microwaves out of its fingers?

还是说这只手在一个瞬间变成10磅重的锤子，在另一个瞬间变成手术刀，而为了烧开水，它从手指里射出，你知道，微波？

You know, or is it more likely just to use a microwave, you know?

你知道，还是说更有可能只是用微波炉，你知道吗？

And the first time it goes up to the microwave, it probably doesn't know how to use it.

而它第一次走到微波炉前时，它可能不知道怎么用它。

But that's okay.

但没关系。

It's connected to the internet.

它连接着互联网。

It reads the manual of this microwave, reads it, instantly becomes an expert.

它阅读这台微波炉的说明书，读完后，立刻成为专家。

And so it uses it.

于是它就使用它。

And so I think the, I just described, in fact, almost all of the properties of OpenClaw.

所以我认为，我刚才描述的，实际上几乎是OpenClaw的所有特性。

You know, that it's going to use tools, that it's going to access files.

你知道，它会使用工具，会访问文件。

It's going to be able to do research.

它能够做研究。

It has IO subsystem.

它有IO子系统。

And when you're done reasoning about it through it in that way,

当你以那种方式通过它推理完之后，

then you say, oh my gosh, the impact to the future computing is deeply profound.

然后你说，天哪，对未来计算的影响是极其深远的。

And the reason for that is I think we've just reinvented the computer.

而原因在于，我认为我们刚刚重新发明了计算机。

And then now you say, okay, when did we reason about that?

然后现在你说，好吧，我们什么时候推理过那个？

When did we reason about OpenClaw?

我们什么时候推理过OpenClaw？

If you take the OpenClaw schematic that I used at GTC, you will find it two years ago.

如果你拿我在GTC上使用的OpenClaw示意图，你会发现它是两年前的。

Literally two years ago at GTC, I was talking about agentic systems that exactly reflect OpenClaw today.

真的，两年前在GTC上，我就在谈论完全反映今天OpenClaw的智能体系统。

And of course, the confluence of many things had to happen.

当然，许多事情的汇聚必须发生。

First of all, we needed Claw and GPT and all of these models to reach a level of capability.

首先，我们需要Claw和GPT以及所有这些模型达到一定的能力水平。

So their innovation and their breakthroughs and their continuing advances was really important.

所以他们的创新、突破和持续进步非常重要。

And then, of course, somebody had to create an open source project

然后，当然，必须有人创建一个开源项目，

that was sufficiently robust, you know, and sufficiently complete, and that we can all put to work.

这个项目足够健壮，你知道，足够完整，而且我们都能投入使用。

And I think OpenClaw did for agentic systems what ChatGPT did for generative systems.

我认为OpenClaw对智能体系统所做的，就像ChatGPT对生成式系统所做的一样。

And I just think it's a very big deal.

我只是觉得这是一件非常重要的事。

Yeah, it's a really special moment.

是的，这真是一个特别的时刻。

I'm not exactly sure why it captured so much of the world's attention,

我不太确定为什么它吸引了全世界如此多的关注，

but it did more than cloud code and codex and so on.

但它做的远不止云代码和Codex等等。

Because consumers could reach it.

因为消费者能够接触到它。

Sure. Yeah. But there's also so much of this is vibes.

当然。是的。但这里面也有很多是氛围。

And Peter, I had a podcast with him, a wonderful human being.

还有Peter，我和他做过一期播客，一个很棒的人。

So part of it is also the humans that represent the thing.

所以部分原因也在于代表这件事的人。

Yeah, no doubt. And part of it is memes.

是的，毫无疑问。还有部分原因是梗图。

Because we're all trying to figure it out.

因为我们都在试图弄明白。

There's really serious and complicated security concerns about when you have such powerful technology,

当你拥有如此强大的技术时，确实存在严重而复杂的安全担忧，

how do you hand over your data so they can do useful stuff?

你如何交出你的数据，以便它们能做有用的事情？

But then there are scary things associated with that.

但随之而来的是与之相关的可怕事情。

And we as a civilization, as individual people and as a civilization, figuring out how to find that right balance.

而我们作为一个文明，作为个人，也作为一个文明，正在弄清楚如何找到那个正确的平衡。

Yeah, we jumped on it right away and we sent a bunch of security experts this way.

是的，我们立刻投入其中，派了一批安全专家过来。

And we did this thing called OpenShell.

我们做了这个叫OpenShell的东西。

It's already been integrated into OpenClaw.

它已经被整合到OpenClaw里了。

And NVIDIA put forward NemoClaw.

NVIDIA推出了NemoClaw。

Yeah, exactly. The install is super easy.

是的，没错。安装非常简单。

It makes sure that it's secure.

它确保它是安全的。

We give you two out of three rights.

我们给你三项权利中的两项。

Agentic systems can access sensitive information,

智能体系统可以访问敏感信息，

it can execute code, and it can communicate externally.

它可以执行代码，还可以对外通信。

We could keep things safe if we gave you two out of those three capabilities at any time, but not all three.

我们可以保证安全，如果我们随时给你三项能力中的两项，但不是全部三项。

And out of those two out of three capabilities, we also give you access control based on whatever rights that you're given by enterprise.

在这三项能力中的两项里，我们还根据企业赋予你的权限，给你访问控制。

And then we connect it to a policy engine that all these enterprises already have.

然后我们把它连接到所有这些企业已经拥有的策略引擎。

And so we're going to try to do our best to help OpenClaw become a better claw.

所以我们会尽力帮助OpenClaw成为更好的爪子。

So you eloquently explained how we have a long history of blockers that we thought were going to be blockers and we overcame them.

所以你雄辩地解释了，我们有一段很长的历史，那些我们以为会成为阻碍的障碍，我们都克服了。

But now looking into the future, what do you think might be the blockers now that it's clear that agents will be everywhere?

但现在展望未来，既然智能体将无处不在已经很明显，你认为现在可能是什么阻碍？

So it's obviously we're going to need compute.

所以很明显，我们将需要算力。

So what is going to be the blocker for that scaling?

那么，这种扩展的阻碍会是什么？

Power is a concern, but it's not the only concern.

电力是一个问题，但不是唯一的问题。

But that's the reason why we're pushing so hard on extreme co-design so that we can improve the tokens per second per watt orders of magnitude every single year.

但这就是为什么我们如此大力推动极端协同设计，以便我们能够每年将每瓦每秒的token数提高几个数量级。

And so in the last 10 years, Moore's Law would have progressed computing about 100 times in the last 10 years.

所以在过去10年里，摩尔定律本应使计算能力在过去10年里进步约100倍。

We progressed and scaled up computing by a million times in the last 10 years.

我们在过去10年里将计算能力提升并扩展了一百万倍。

And so we're going to keep on doing that through extreme co-design.

所以我们将通过极端协同设计继续这样做。

So energy efficiency per watt completely affects the revenues of a company.

所以每瓦的能效完全影响一家公司的收入。

It affects the revenues of a factory and

它影响工厂的收入，而且

we're just, we're just going to push that to the limit

我们只是，我们只是要把这一点推到极限，

so that we keep on driving token costs down as fast as we can.

这样我们就能尽可能快地持续降低token成本。

You know, the our computer price is going up,

你知道，我们的电脑价格在上涨，

but our token generation effectiveness is going up so much faster.

但我们的token生成效率提升得快得多。

The token cost is coming down.

token成本正在下降。

It's coming down an order of magnitude every year.

它每年下降一个数量级。

So power, that's an interesting one.

所以电力，那是个有趣的问题。

So the way to try to get around the power blocker

所以绕过电力瓶颈的方法是

is to try to, with the tokens per second per watt, try to make it more and more efficient.

是尝试，在每瓦每秒token数上，努力让它越来越高效。

Of course, there's the question of how do we get more power.

当然，问题是我们如何获得更多电力。

We should also get more power.

我们也应该获得更多电力。

That's a really complicated one.

那是个非常复杂的问题。

You've talked about small modular nuclear power plants.

你谈到过小型模块化核电站。

There's all kinds of ideas for energy.

有各种各样的能源想法。

How much does it keep you up at night, the bottlenecks in the supply chain of AI?

AI供应链中的瓶颈在多大程度上让你夜不能寐？

like ASML with EUV lithography machines,

比如ASML的EUV光刻机，

GSMC with advanced packaging,

台积电的先进封装，

like Coos, and SK Hennix with high bandwidth memory.

比如Coos，以及SK海力士的高带宽内存。

Well, all the time.

嗯，一直如此。

And we're working on it all the time.

我们一直在努力解决。

No company in history has ever grown at a scale that we're growing while accelerating that growth.

历史上没有任何公司以我们这样的规模增长，同时还在加速增长。

It's incredible.

这太不可思议了。

And it's hard for people to even understand this.

人们甚至很难理解这一点。

In the overall world of AI computing, we're increasing share.

在整个人工智能计算领域，我们的份额正在增加。

And so supply chain upstream and downstream are really important to us.

所以供应链的上游和下游对我们来说非常重要。

I spent a lot of time informing all the CEOs that I work with,

我花了很多时间告诉我合作的所有CEO，

what are the dynamics that's going to cause the growth to continue or even accelerate.

哪些动态会导致增长持续甚至加速。

It's part of the reasons why to the entire right-hand side of me were CEOs of practically the entire IT industry upstream and practically the entire infrastructure industry downstream.

这也是为什么我右边几乎整个IT行业上游和几乎整个基础设施行业下游的CEO们都在这里的原因之一。

And there were several hundred CEOs.

有几百位CEO。

And I don't think there's ever been keynotes where several hundred CEOs show up.

我不认为曾经有过几百位CEO出席的主题演讲。

And part of it is I'm telling them about our business condition now.

部分原因是我在告诉他们我们目前的业务状况。

I'm telling them about the growth drivers in the very near future and what's happening.

我在告诉他们近期的增长驱动力以及正在发生的事情。

And I'm also describing where are we going to go next so that they could use all of this information and all of the dynamics that are here to inform how they want to invest.

我还描述我们接下来要去哪里，这样他们就可以利用所有这些信息和这里的动态来指导他们想要如何投资。

And so I inform them that way like I inform my own employees.

所以我以这种方式告知他们，就像我告知自己的员工一样。

And then, of course, then I make trips out to them and make sure that, hey, listen, I want you to know this quarter, this coming year, this next year, these things are going to happen.

然后，当然，我会去拜访他们，确保，嘿，听着，我想让你知道这个季度、明年、后年，这些事情将会发生。

And if you look at the CEOs of the DRAM industry, the number one DRAM in the world was DDR memory for CPUs and data centers.

如果你看看DRAM行业的CEO们，世界上排名第一的DRAM是用于CPU和数据中心的DDR内存。

About three years ago, I was able to convince several of the CEOs that even though at the time HBM memory was used quite scarcely, you know, and barely by supercomputers, that this was going to be a mainstream memory for data centers in the future.

大约三年前，我成功说服了几位CEO，尽管当时HBM内存使用相当稀少，你知道，几乎只有超级计算机在用，但这将成为未来数据中心的主流内存。

And at first it sounded ridiculous, but several of the CEOs believed me and decided to invest in building HBM memories.

起初这听起来很荒谬，但几位CEO相信了我，并决定投资建设HBM内存。

Another memory was rather odd to put into a data center is the low power memories that we use for cell phones.

另一种放入数据中心的奇怪内存是我们用于手机的低功耗内存。

And we wanted them to adapt them for supercomputers in the data center.

我们想让他们为数据中心的超级计算机适配这些内存。

And they go, cell phone memory for supercomputers?

然后他们说，手机内存用于超级计算机？

And I explained to them why. Well, look at these two memories, LPDDR5, HBM4.

我向他们解释了原因。嗯，看看这两种内存，LPDDR5、HBM4。

The volumes are so incredible. All three of them had record years in history. And these are 45-year companies.

产量如此惊人。这三家公司都创下了历史记录。而这些都是有45年历史的公司。

And so, you know, that's part of my job is to inform and shape, inspire, you know.

所以，你知道，我的部分工作就是告知和塑造、激励，你知道。

So you're not just manifesting the future and maybe inspiring NVIDIA, the different engineers of the company.

所以你不仅仅是在显化未来，也许还在激励NVIDIA，公司里不同的工程师们。

You're manifesting the supply chain of the future.

你在显化未来的供应链。

So you're having conversations with TSMC, with ASML. Upstream, downstream. Upstream, downstream.

所以你在与台积电、与ASML对话。上游、下游。上游、下游。

So that's the thing. GEV, Caterpillar.

所以就是这样。GEV、卡特彼勒。

Yeah, that's downstream from us.

是的，那是在我们的下游。

Yeah, the whole thing.

是的，整个事情。

I mean, but that's so, there's so much incredibly difficult engineering that happens in the entire semiconductor industry.

我的意思是，但那就是如此，整个半导体行业中有太多极其困难的工程。

And it just feels scary how intricate the supply chain is, how many components there are.

而且供应链如此复杂，有多少组件，这让人感到害怕。

But it works somehow.

但不知怎么的，它还是运转起来了。

Exactly. the deep science, the deep engineering, the incredible manufacturing, and so much of the manufacturing is already robotics.

确实如此。深层的科学、深层的工程、令人难以置信的制造，而且很多制造已经是机器人化了。

But we have a couple of hundred suppliers that contribute the technology that goes into our 1.3 million component rack.

但我们有几百家供应商，他们提供的技术进入我们130万个组件的机架。

Each rack is 1.3, 1.5 million components.

每个机架有130万、150万个组件。

There are 200 suppliers across the Verorubin rack.

Verorubin机架有200家供应商。

So it's interesting that you don't list that as the thing that keeps you up at night in the list the blockers.

所以有趣的是，你没有把它列为让你夜不能寐的事情，在阻碍清单上。

But I'm doing all the things necessary to see.

但我正在做所有必要的事情来确保。

I can go to sleep because I checked it off.

我可以去睡觉，因为我把它勾掉了。

I said, okay. I can go to sleep and I go, well, let's see.

我说，好的。我可以去睡觉，然后我说，嗯，让我们看看。

Let's reason about this. What's important for us?

让我们推理一下。对我们来说什么重要？

Okay, let's reason about this.

好的，让我们推理一下。

Because we changed the system architecture from the original DGX-1 that you remembered to MVLink 72 rack scale computing.

因为我们改变了系统架构，从你记得的原始DGX-1到MVLink 72机架规模计算。

What does that mean?

那意味着什么？

What does that mean to software?

那对软件意味着什么？

What does that mean to engineering?

那对工程意味着什么？

What does that mean to how we design and test?

那对我们如何设计和测试意味着什么？

And what does that mean to the supply chain?

那对供应链意味着什么？

Well, one of the things that it meant was we moved supercomputer integration at the data center into supercomputer manufacturing in the supply chain.

嗯，它意味着的一件事是，我们把数据中心的超级计算机集成转移到了供应链中的超级计算机制造环节。

If you're doing that, you also have to recognize you're going to move one.

如果你这么做，你还必须认识到，你会移动一个环节。

And if, if, if you're, if you're, you know, total footprint of whatever data center you're going to build,

而且如果，如果，如果你要建的数据中心的总占地，你知道，

let's say you would like to have, you know, 50 gigawatts of supercomputers that are running simultaneously.

假设你想要拥有，你知道，50吉瓦同时运行的超级计算机。

And it takes one week to manufacture that 50 gigawatts of supercomputers,

而制造那50吉瓦的超级计算机需要一周时间，

then each week in the supply chain, the supercomputers are going to need a gigawatt of power.

那么在供应链中，每周这些超级计算机将需要一吉瓦的电力。

And so we're going to need the supply chain to increase the amount of power it has to build,

所以我们需要供应链增加它必须建设的电力，

test, to build and test the supercomputers in the supply chain before I ship it.

测试，在供应链中建造和测试超级计算机，然后我才发货。

Well, NVLink 72 literally built supercomputers in the supply chain and ships some two, three tons at a time per rack.

嗯，NVLink 72实际上是在供应链中建造超级计算机，每次每个机架运送两三吨。

It used to be, they used to come in parts and we used to assemble them inside the data center.

过去，它们是以零件形式运来，我们过去在数据中心内部组装它们。

But that's impossible now because MVLink 72 is so dense.

但现在这不可能了，因为MVLink 72太密集了。

And so that's an example.

所以这就是一个例子。

And I would have to go into, you know, I fly into the supply chain,

而我必须进入，你知道，我飞进供应链，

go meet my partners and hey, I said, guess what?

去见我的合作伙伴，嘿，我说，你猜怎么着？

So here's what I'm going to do with,

所以这就是我要做的，

this is the way we used to build our DGXs.

这就是我们过去建造DGX的方式。

We're going to build them this way.

我们将以这种方式建造它们。

This is going to be so much better because we're going to need them for inference.

这会好得多，因为我们将需要它们来进行推理。

The market for inference is, you know, coming.

推理市场，你知道，正在到来。

The inflection point for inference is coming.

推理的拐点正在到来。

It's going to be a big market.

这将是一个巨大的市场。

And so I first explain to them what's going on, why it's going to happen.

所以我首先向他们解释正在发生什么，为什么会发生。

And then I ask them to make several billion dollars of capital investments each.

然后我要求他们各自进行数十亿美元的资本投资。

And because they, you know, they trust me and I'm very respectful of them.

因为他们，你知道，他们信任我，而我也非常尊重他们。

And I give them every opportunity to question me.

我给他们每一个机会来质疑我。

And I spend time to explain things to people.

我花时间向人们解释事情。

And I reason about it.

我对此进行推理。

I draw them pictures and I reason about it in first principles.

我给他们画图，并从第一性原理进行推理。

And by the time I'm done with them, there's no what to do.

等我和他们谈完，就没有什么可做的了。

So a lot of it is about relationships and building a shared view of the future.

所以很多都是关于关系和建立对未来的共同愿景。

But do you worry about certain bottlenecks?

但你担心某些瓶颈吗？

I mean, what are the biggest bottlenecks in the supply chain?

我是说，供应链中最大的瓶颈是什么？

Are you worried about ASMLs, EV tooling?

你担心ASML、EV工具吗？

Are you worried about the packaging, co-host packaging of TSMC, about how fast it could scale?

你担心台积电的封装、共封装，担心它能多快扩展吗？

Like you said, you're not only growing incredibly fast, you're accelerating your growth.

就像你说的，你不仅增长得极快，还在加速增长。

So it feels like everybody in the supply chain, and those are certainly bottlenecks, would have to scale up.

所以感觉供应链上的每个人，而这些确实是瓶颈，都必须扩大规模。

Are you having conversations with them? Like, how can you scale this faster?

你在和他们沟通吗？比如，怎样才能更快地扩大规模？

Do you worry about it?

你担心这件事吗？

No.

不担心。

Okay. Because I told them what I needed.

好的。因为我告诉了他们我需要什么。

They understood what I need.

他们明白我需要什么。

They told me what they're going to go do, and I believe in what they're going to do.

他们告诉了我他们打算做什么，我相信他们要做的事。

Interesting.

有意思。

Yeah. That's great to hear.

是啊。很高兴听到这个。

So maybe if we can just linger in the power for a little bit, what are your hopes for how to solve the energy problem?

所以也许我们可以再在电力这个话题上多停留一会儿，你对如何解决能源问题有什么期望？

One of the areas that I would love us to talk about and just get the message out, you know, our power grid is designed for the worst case condition with some margin.

我希望我们能讨论并传达出去的一个领域是，你知道，我们的电网是按最坏情况并留有一定余量来设计的。

Well, 99% of the time, we're nowhere near the worst case condition because the worst case condition is a few days in the winter, a few days in the summer, and extreme weather.

嗯，99%的时间里，我们远未接近最坏情况，因为最坏情况是冬天几天、夏天几天，以及极端天气。

Most of the time, we're nowhere near the worst case condition, and we're probably running around, call it 60% of peak.

大多数时候，我们远未接近最坏情况，我们大概运行在峰值的60%左右。

And so 99% of the time, our power grid has excess power, and they're just sitting idle.

所以99%的时间里，我们的电网都有多余的电力，它们只是闲置着。

But they have to be there sitting idle because just in case, when the time comes,

但它们必须闲置待命，以防万一，当需要的时候，

hospitals have to be powered and infrastructure has to be powered and airports have to run and so on and so forth.

医院必须供电，基础设施必须供电，机场必须运转，等等等等。

And so the question that I have is whether we could go and help them understand

所以我的问题是，我们能否去帮助他们理解，

and create contractual agreements and design computer architecture systems, data centers, such that when they need the maximum power for infrastructure in society,

并制定合同协议，设计计算机架构系统、数据中心，使得当他们需要社会基础设施的最大电力时，

that the data centers would get less.

数据中心获得的电力会减少。

But that's in a very rare instance anyways.

但无论如何，那都是非常罕见的情况。

And during that time, we either have a backup generator for that little part of it, or we just have our computers shift the workload somewhere else,

在那段时间里，我们要么为那一小部分配备备用发电机，要么就让我们的计算机把工作负载转移到别处，

or we have the computers just run slower.

要么让计算机运行得慢一些。

You know, we could degrade our performance, reduce our power consumption, and provide for, you know, a slightly longer latency response, you know, when somebody asks for, you know, asks for an answer.

你知道，我们可以降低性能，减少功耗，并提供，你知道，稍微长一点的延迟响应，你知道，当有人请求，你知道，请求一个答案时。

And so I think that that way of using computers, of building data centers, instead of expecting 100% uptime and these contracts that are really, really quite rigorous, it's putting a lot of pressure on the grid to be able to, now they're going to have to increase from their maximum.

所以我认为，这种使用计算机、建设数据中心的方式，而不是期望100%的正常运行时间和这些非常非常严格的合同，给电网带来了很大压力，现在他们不得不从最大值开始增加。

I just want to use their excess. It's just sitting there.

我只是想用他们多余的（电力）。它就闲置在那里。

Yeah, that's not talked about enough.

是啊，这一点讨论得还不够。

So what's stopping there? Is it regulation? Is it bureaucracy?

那么问题出在哪里？是监管吗？是官僚体制吗？

I think it's a three-way problem.

我认为这是一个三方问题。

It starts with the end customer. The end customer puts requirements on the data centers

它从终端客户开始。终端客户对数据中心提出要求，

that they can never not be available.

要求它们永远不能不可用。

Okay, so that the end customer expects perfection.

好，所以终端客户期望完美。

Now, in order to deliver that perfection, you need a combination of backup generators and your grid power supplier to deliver on perfection.

现在，为了实现那种完美，你需要备用发电机和电网供电方结合起来，共同实现完美。

And so everybody's got to have six nines.

所以每个人都必须达到六个九（99.9999%的可用性）。

Well, I think, first of all, right now, we ought to have everybody understand that when the customer asks for these things,

嗯，我认为，首先，现在，我们应该让每个人都明白，当客户提出这些要求时，

you have somebody in your data center operations team disconnected from the CEO.

你的数据中心运营团队里有人与CEO是脱节的。

I bet the CEO doesn't know this.

我敢打赌CEO不知道这件事。

I'm going to talk to all the CEOs.

我要去跟所有的CEO谈谈。

The CEOs are probably not paying any attention to the contracts that are being signed.

CEO们可能根本没有关注正在签署的合同。

And so everybody wants to sign the best contract, of course.

所以当然，每个人都想签最好的合同。

And they go down to the cloud service providers and the contract, the two contract negotiators that are,

然后他们去找云服务提供商，而合同，那两位合同谈判代表，

I could just see them now, you know, negotiating these multi-year contracts,

我现在就能想象到他们，你知道，在谈判这些多年期合同，

both sides want, you know, the best contract as a result.

双方都想要，你知道，最终得到最好的合同。

The CSPs then have to go down to the utilities

然后云服务提供商必须去找公用事业公司，

and they expect the six nines.

而他们期望六个九（99.9999%的可用性）。

And so I think the first thing is just make sure that all of the customers, the CEOs of the customers realize what they're asking for.

所以我认为第一件事就是确保所有客户，客户的CEO们，意识到他们要求的是什么。

Now, the second thing is we have to build data centers that gracefully degrade.

现在，第二件事是我们必须构建能够优雅降级的数据中心。

And so if the power, if the utility of the grid tells us, listen, we're going to have to back you down to about 80%, we're going to say that's no problem at all.

所以如果电力，如果电网的公用事业公司告诉我们，听着，我们得把你降到大约80%，我们会说这完全没问题。

We're just going to move our workload around. We're going to make sure that data is never lost, but we can reduce the computing rate and use less energy.

我们只是要转移我们的工作负载。我们要确保数据永远不会丢失，但我们可以降低计算速率并使用更少的能源。

The quality of service degrades a little bit. For the critical workloads, I shift that somewhere else right away.

服务质量会稍微下降。对于关键工作负载，我会立即将其转移到其他地方。

So I don't have that problem. And so, you know, whichever data center still has 100% uptime.

所以我没有那个问题。所以，你知道，无论哪个数据中心仍然保持100%的正常运行时间。

How difficult of an engineering problem is that the smart dynamic allocation of power in the data center?

数据中心中电力的智能动态分配，这有多难解决？

As soon as you could specify, you could engineer it.

一旦你能明确指定，你就能设计出来。

So long as it obeys the laws of physics on first principles,

只要它遵循物理学定律和第一性原理，

I think we're good.

我觉得就没问题。

What was the third thing you were mentioning? Um, so the second thing is the data centers,

你提到的第三件事是什么？嗯，所以第二件事是数据中心，

and the third thing is we need utilities to also recognize that this is an opportunity,

第三件事是，我们需要公用事业公司也认识到这是一个机会，

and and instead of instead of saying, look, um, it's going to take me five years to increase my grid capability.

而不是说，看，嗯，我要花五年时间才能提高我的电网能力。

If you're willing to take power of this level of guarantee,

如果你愿意接受这种级别的保证，

I can make them available for you next month and at this price.

我可以在下个月以这个价格为你提供。

And so if utilities also offered more segments of power delivery promises,

所以，如果公用事业公司也提供更多分段的电力交付承诺，

then I think everybody will figure out what to do with it.

那么我想大家都会想出该怎么利用它。

But there's just way too much waste in the grid right now.

但现在电网中的浪费实在太多了。

We should go after it.

我们应该着手解决它。

You've highly lauded Elon and XAI's accomplishment in Memphis in building Colossus supercomputer, probably in record time in just four months.

你高度赞扬了埃隆和xAI在孟菲斯建造Colossus超级计算机的成就，可能仅用了四个月就创下了纪录。

It's now at 200,000 GPUs and growing very quickly.

现在它已有20万块GPU，并且增长非常迅速。

Is there something that you could speak to, understand about his approach that's instructive broadly to all the data center creators that enable that kind of accomplishment?

关于他的方法，有没有什么你可以谈谈、让我们理解的，能对所有促成这种成就的数据中心创建者有广泛启发？

His approach to engineering, his approach to the whole management of construction, everything?

他的工程方法，他对整个施工管理的方法，一切？

First of all, Elon is deep in so many different topics, yet he's also a really good systems thinker.

首先，埃隆深入涉猎许多不同领域，但他也是一个非常出色的系统思考者。

And so he's able to think through multiple disciplines.

因此他能够跨多个学科进行思考。

And he obviously pushes things, questions everything, whether number one is it necessary?

他显然会推动事情，质疑一切，第一，这有必要吗？

Number two, does it have to be done this way?

第二，必须这样做吗？

And in other words, you know, does it have to take this long?

换句话说，你知道，它必须花这么长时间吗？

And so he has the ability to question everything to the point where everything is down to its minimal amount that's necessary.

所以他能够质疑一切，直到一切都降到必要的最低限度。

You can't take anything else out.

你不能再拿走任何东西了。

And yet the necessary capabilities of the product retains, you know.

然而产品必要的功能得以保留，你知道。

And so he is as minimalist as you could possibly imagine.

所以他极简到超乎你的想象。

And he does it at a system scale.

而且他是在系统规模上做到这一点。

I also love the fact that he is represented, he is present at the point of action.

我也喜欢他亲临现场、出现在行动第一线这一点。

You know, he'll just go there. If there's a problem, he'll just go there and show me the problem.

你知道，他会直接去那里。如果有问题，他会直接去那里，把问题指给我看。

You know, when you do all of this in combination, you overcome a lot of previous, this is just the way we do it.

你知道，当你把所有这些结合起来做时，你就能克服很多以前的，这就是我们做事的方式。

Um, you know, I'm, I'm waiting for them.

嗯，你知道，我在等他们。

Uh, you know, I mean, it's just, everybody has a lot of excuses.

呃，你知道，我的意思是，只是，每个人都有很多借口。

And so, and then the last thing is when, when you act personally with so much urgency, it causes everybody else to act with urgency, you know, and, and every supplier has a lot of customers going on.

所以，然后最后一点是，当你亲自以如此紧迫感行动时，它会让其他人也紧迫地行动，你知道，而且，每个供应商都有很多客户在进行中。

Every supplier has a lot of projects going on and he, he make it, he made it, he makes it his business

每个供应商都有很多项目在进行，而他，他把它，他把它，他把它当成自己的事

that he's the top priority of everybody else's, you know, projects.

就是他是其他所有人项目的最高优先级。

And so he does that by demonstrating it.

所以他通过亲身示范来做到这一点。

Yeah, I've been in a bunch of those meetings.

是啊，我参加过很多这种会议。

It's fun to watch because really not enough people ask the question like,

看着挺有意思的，因为真的没有足够多的人问这样的问题，

okay, so can this be done a lot faster and how?

好吧，那这个能不能快很多完成，怎么做到？

Why does it have to take this long?

为什么非得花这么长时间？

Yeah, right.

是啊，没错。

And then that becomes an engineering question often.

然后这往往就变成了一个工程问题。

And yes, I think when you get the ground truth of actually,

是的，我觉得当你真正获得实际情况的真相时，

I remember one of the times I was hanging out with him,

我记得有一次我和他待在一起，

he literally is going through the entire process of how to plug in cables into a rack.

他真的在走一遍把线缆插进机架的整个流程。

He's working with an engineer on the ground that's doing that task,

他在和现场做那个任务的工程师一起工作，

and he's just trying to understand what does that process look like so it can be less error-prone.

他只是在努力理解那个流程是什么样的，这样就能减少出错。

And just building up that intuition from every single task involved in putting together the data center,

通过参与搭建数据中心的每一项任务来积累这种直觉，

you start to immediately get a sense at the detailed scale and at the broad system scale of where the inefficiencies are.

你就能立刻从细节层面和整个系统层面感觉到低效之处在哪里。

And so you can make it more and more and more efficient.

这样你就能让它变得越来越高效。

Plus you have the big hammer of being able to say, let's do it totally different and remove all possible blockers.

另外，你有一个大锤，就是能够说：让我们完全换个做法，消除所有可能的障碍。

That's right.

没错。

Is there parallels in the NVIDIA Extreme Systems co-design approach that you see in the way Elon approaches systems engineering?

在NVIDIA Extreme Systems的协同设计方法中，是否有与埃隆处理系统工程方式相似的地方？

Well, first of all, co-design is an ultimate systems engineering problem.

嗯，首先，协同设计是一个终极的系统工程问题。

And so we approach the work that we do from that principle.

所以我们从那个原则出发来处理我们所做的工作。

The other thing that we do, and this is a philosophy that I thought, a state of mind, I guess, a method that I started 30 years ago, and it's called the speed of light.

我们做的另一件事，这是一种哲学，我认为，一种心态，我猜，一种我30年前开始的方法，叫做光速。

A speed of light is not just about the speed.

光速不仅仅是关于速度。

A speed of light is my shorthand for what's the limit of what physics can do.

光速是我对物理学极限的简写。

And so everything that we do is compared against the speed of light.

所以我们所做的一切都是与光速进行比较。

Memory speed, math speed, power, cost, time, effort, number of people, manufacturing cycle time.

内存速度、数学速度、功耗、成本、时间、精力、人数、制造周期时间。

And when you think about latency versus throughput, when you think about cost versus throughput, cost versus capacity,

当你考虑延迟与吞吐量、成本与吞吐量、成本与容量时，

All of these things, you test against the speed of light to achieve all of these different constraints separately.

所有这些东西，你都要对照光速来测试，以分别实现所有这些不同的约束。

And then when you consider it together, you know you have to make compromises

然后当你把它们放在一起考虑时，你知道你必须做出妥协

because a system that achieves extremely low latency versus a system that achieves very high throughput are architected fundamentally differently.

因为一个实现极低延迟的系统与一个实现极高吞吐量的系统在架构上根本不同。

But you want to know what's the speed of light of a system that achieves high throughput?

但你想知道一个实现高吞吐量的系统的光速是多少吗？

What's the speed of light of a system that achieves low latency?

一个实现低延迟的系统的光速是多少？

And then when you think about the total system, you could make trade-offs.

然后当你考虑整个系统时，你可以做出权衡。

And so I force everybody to think about what's the first principles, the limits, the physical limits for everything before we do anything.

所以我强迫每个人思考什么是第一性原理，极限，一切事物的物理极限，在我们做任何事情之前。

And we test everything against that.

我们针对那个来测试一切。

And so that's a good frame of mind.

所以这是一个好的心态。

I don't love the other methods, which is continuous improvement.

我不喜欢其他方法，也就是持续改进。

The problem with continuous improvement, first of all, you should engineer something from first principles with speed of light thinking, limited only by physical limits and physics limits.

持续改进的问题，首先，你应该从第一性原理出发，以光速思维来设计东西，只受物理极限和物理定律的限制。

And after that, of course, you would improve it over time.

在那之后，当然，你会随着时间的推移改进它。

But I don't like going into a problem and somebody says, hey, you know, it takes 74 days to do this today, right now.

但我不喜欢进入一个问题时有人说，嘿，你知道，现在做这个需要74天，就是现在。

And we can do it for you in 72 days.

我们可以为你72天完成。

You know, I'd rather strip it all back to zero.

你知道，我宁愿把它全部归零。

And so, first of all, explain to me why it's 74 days in the first place.

所以，首先，向我解释为什么一开始是74天。

And let's think about what's possible today.

让我们想想今天什么是可能的。

And if I were to build it completely from scratch, you know, how long would it take?

如果我完全从零开始建造它，你知道，需要多长时间？

Oftentimes you'd be surprised and might come to six days.

很多时候你会感到惊讶，可能会缩短到六天。

Now the rest of the six days to 74 could be very well reasoned and compromises and, you know, cost reductions and all kinds of different things.

现在，从六天到七十四天的其余部分，很可能是经过充分论证的，还有各种妥协，你知道，成本削减以及各种不同的事情。

But at least you know what they are.

但至少你知道它们是什么。

And now that you know that six days is possible, then the conversation from 74 to 6 is surprisingly much more effective.

既然你知道六天是可能的，那么从七十四天到六天的对话就会出乎意料地有效得多。

In such incredibly complex systems that you're working with, is simplicity sometimes a good heuristic to reach for?

在你所处理的如此极其复杂的系统中，简单性有时是不是一个值得追求的好启发式方法？

I mean, if I can just... I mean, the pod, the Vera Rubin pod that you announced is just incredible.

我的意思是，如果我能只是……我的意思是，那个 pod，你宣布的 Vera Rubin pod，简直太不可思议了。

We're talking about seven chips, seven chip types, five purpose built rack types, 40 racks, 1.2 quadrillion transistors,

我们说的是七种芯片，七种芯片类型，五种专门打造的机架类型，四十个机架，一点二千万亿个晶体管，

nearly 20,000 NVIDIA dyes, over 1,100 Rubin GPUs, 60 exaflops, 10 petabytes per second of scale of bandwidth.

近两万个 NVIDIA 裸片，超过一千一百个 Rubin GPU，六十百亿亿次浮点运算，每秒十拍字节的带宽规模。

That's all just one.

那全都只是一个。

That's just one pod.

那只是一个 pod。

That's just one pod.

那只是一个 pod。

Yeah, that's just one pod. I mean, so you have the, and then even the NVL72 rack alone is 1.3 million components, 1,300 chips, 4,000 pounds crammed into a single 19-inch wide rack.

是的，那只是一个 pod。我的意思是，所以你有那个，然后甚至仅 NVL72 机架就有 130 万个组件，1300 个芯片，4000 磅塞进一个 19 英寸宽的机架里。

And Lex, we'll probably kind of crank out about 200 of these pods a week, just to put in perspective.

Lex，我们每周大概能生产200个这样的pod，只是让你有个概念。

The amount of different components, I suppose simplicity is impossible, but is that a metric that you kind of reach for in trying to design things?

不同组件的数量，我想简单是不可能的，但这是你在设计东西时追求的一个指标吗？

You know, the phrase that I use most often is we need things to be as complex as necessary, but as simple as possible.

你知道，我最常用的说法是，我们需要东西尽可能复杂，但也要尽可能简单。

And so the question is, is all that complexity there necessary?

所以问题是，所有那些复杂性都是必要的吗？

And we ought to test for that. And we ought to challenge that.

我们应该测试这一点。我们应该挑战这一点。

And then after that, everything else above it, you know, is gratuitous.

然后在那之后，其他一切在上面，你知道，都是多余的。

But it's some of the most incredible. Semiconductor industry broadly, but what NVIDIA is doing, some of the greatest engineering in history.

但这是最令人难以置信的。整个半导体行业，但NVIDIA所做的，是历史上最伟大的工程之一。

So these systems are just truly, truly marvels of engineering.

所以这些系统真的是，真的是工程奇迹。

It is the most complex computer the world has ever made.

这是世界上最复杂的计算机。

Yeah, the engineering teams, I mean, I don't know. It's not a competition, but I don't know.

是的，工程团队，我的意思是，我不知道。这不是比赛，但我不知道。

If it was like an Olympics of engineering teams, I mean, TSMC does incredible engineering.

如果这像是工程团队的奥运会，我的意思是，台积电做得很出色。

Like I said, ASML at every scale. But NVIDIA is going to give them a run for their money.

就像我说的，ASML在各个层面。但NVIDIA会给他们带来激烈的竞争。

Just incredible, incredible teams. Gold medalists in every single sport all assembled right here.

只是令人难以置信的，令人难以置信的团队。每个项目的金牌得主都聚集在这里。

And have to work together and report directly to you.

并且必须一起工作，直接向你汇报。

This is wonderful.

这太棒了。

You've recently traveled to China.

你最近去了中国。

So it's interesting to ask you.

所以问您很有意思。

China has been incredibly successful in building up its technology sector.

中国在建立其科技领域方面非常成功。

What do you understand about how China is able to, over the past 10 years, build so many incredible world-class companies, world-class engineering teams, and just this technology ecosystem that produces so many incredible products?

你如何理解中国在过去10年里能够建立这么多令人难以置信的世界级公司、世界级工程团队，以及这个能产出这么多令人惊叹的产品的技术生态系统？

A whole bunch of reasons.

有很多原因。

Well, first of all, let's start with some facts.

嗯，首先，让我们从一些事实开始。

50% of the world's AI researchers are Chinese, plus or minus.

全球50%左右的人工智能研究人员是中国人。

And they're mostly in China.

而且他们大多在中国。

Still, we have many of them here, but there's amazing researchers still in China.

不过，我们这里也有许多这样的人，但中国仍然有出色的研究人员。

Their tech industry showed up at precisely the right time.

他们的科技产业恰好在正确的时间出现。

At the time of the mobile cloud era, their way of contributing was software.

在移动云时代，他们的贡献方式是软件。

And so this is a country's incredible science and math, really well-educated kids.

所以这是一个拥有出色科学和数学教育、真正受过良好教育的孩子的国家。

Their tech industry was created during the era of software.

他们的科技产业是在软件时代创建的。

They're very comfortable with modern software.

他们对现代软件非常熟悉。

China is not one giant economic country.

中国不是一个单一的巨大经济体。

It's got many provinces and cities with mayors all competing with each other.

它有许多省份和城市，市长们都在相互竞争。

That's the reason why there's so many EV companies.

这就是为什么有这么多电动汽车公司。

That's the reason why there's so many AI companies.

这就是为什么有这么多人工智能公司。

That's the reason why there's so many, every company you could imagine, they all create some of them.

这就是为什么有这么多，你能想象到的每家公司，它们都创建了一些。

And as a result, they have insane competition internally.

因此，它们内部有着疯狂的竞争。

And, you know, what remains is an incredible company.

而且，你知道，剩下的是一个了不起的公司。

They also have a social culture where it's family first, friends second, and company third.

他们还有一种社会文化，即家庭第一，朋友第二，公司第三。

And so, you know,

所以，你知道，

The amount of conversation that goes back and forth between, they're essentially open source all the time.

他们之间来回的对话量，基本上一直是开源的。

So the fact that they contribute more to open source is so sensible because they're probably, what are we protecting?

所以他们更多地为开源做贡献，这很合理，因为他们可能，我们在保护什么？

You know, my engineers, their brothers are in that company, their friends are in that company, and they're all schoolmates.

你知道，我的工程师们，他们的兄弟在那家公司，他们的朋友在那家公司，而且他们都是校友。

You know, the schoolmate concept is, you know, one schoolmate, your brother for life.

你知道，校友这个概念就是，你知道，一个校友，终身的兄弟。

And so they share knowledge very, very quickly.

所以他们非常非常快地分享知识。

And so there's no sense keeping technology hidden.

所以把技术藏起来没有意义。

You might as well put it on open source.

你不如把它开源。

And so the open source community then amplifies, accelerates the innovation process.

所以开源社区就会放大、加速创新过程。

So you get this rapid, incredibly great talent, rapid innovation because of open source and just, you know, the nature of friends.

所以你得到了这种快速的、极其优秀的人才，快速的创新，因为开源，还有，你知道，朋友的本质。

And insane competition among the company.

以及公司内部的疯狂竞争。

What emerges is incredible stuff.

涌现出来的是不可思议的东西。

And so this is the fastest innovating country in the world today.

所以这是当今世界上创新最快的国家。

And this is something that has everything that I've just said is fundamental to just how the kids were grown.

而这件事包含了所有我刚才说的，对孩子们的成长方式至关重要。

The fact that they have excellent education.

他们拥有优秀的教育。

The fact that parents want them to do well in school.

父母希望他们在学校表现好。

The fact that their culture is that way.

他们的文化就是那样。

These are just the thing about their country.

这些正是他们国家的特点。

And they showed up at precisely the time when technology is going through that exponential.

而他们正好出现在技术经历指数级增长的时候。

Plus, culturally, it's pretty cool to be an engineer.