第494期 – 黄仁勋：英伟达——价值四万亿美元的公司与人工智能革命

但这就是你想要的。

But that's what you want.

你希望把每个人都带上。

You wanna bring everybody along.

你知道的。

You know?

否则，当我们宣布关于深度学习的事情时，大家都会问：你在说什么？

Otherwise, we announce something about deep learning, and everybody goes, what are you talking about?

你知道，当你宣布我们要全力投入这件事时。

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

你的管理团队、董事会、员工、客户都会觉得：这主意哪来的？

And and your your management team, your board, your employees, your customers, they're kinda like, where is this coming from?

你知道的。

You know?

这太疯狂了。

This is insane.

所以，GTC实际上，如果你回溯过去，看看那些主题演讲，我不仅在塑造合作伙伴和整个行业的信念体系，也在用这种方式塑造我自己的员工的信念。

And so so GTC, in fact, if you go back in time, you look at look at the keynotes, I'm also shaping the belief system of my partners and the industry, and and I'm using that to shape, you know, the belief system of my own employees.

因此，当我宣布一些事情时，比如我们刚刚发布了Grok。

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

其实我已经提前两年半一直在谈论这些铺垫步骤了。

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

你们回头一看，简直惊呆了。

You guys just go back and go, oh my gosh.

他们已经谈论了两年半了。

They've been talking about it for two and a half years.

所以我一直在一步一个脚印地打基础。

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

所以当真正宣布的时候，大家都会说：‘这怎么花了这么久？’

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

但这不仅仅局限于公司内部。

But it's not just inside the company.

你正在塑造创新的全球大格局。

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

把那些想法传播出去，你真的在实现现实。

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

我们不制造电脑。

We don't build computers.

我们实际上也不构建云。

We actually don't build clouds.

事实证明，我们是一家计算平台公司。

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

所以没有人能从我们这里购买任何产品。

And so nobody can buy anything from us.

这很奇怪。

That's the weird thing.

我们从上到下进行设计和整合，以实现设计和优化，但随后我们在每一层都开放整个平台，使其能够集成到其他公司的产品、服务、云、超级计算机和OEM电脑中。

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

所以惊人的是，如果没有先说服他们，我就无法做到我目前所做的事情。

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

因此，GTC 的大部分内容都是在描绘一个未来：当我的产品最终准备好时，他们会说：‘你怎么花了这么长时间？’

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

是的。

Yeah.

所以你长期以来一直相信的是广义上的规模定律。

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

那么，你是否仍然相信规模定律？

So are you still a believer in the in the scaling laws?

是的。

Yeah.

我们现在有了更多的规模定律。

We have more scaling laws now.

所以我认为你已经概述了四种：预训练、后训练、推理时和智能体规模定律。

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

当你思考未来——无论是深远的未来还是近期的未来时，你最担心哪些阻碍因素？这些障碍让你夜不能寐，必须克服它们才能继续扩展。

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 that keep you up at night that you have to overcome in order to keep scaling?

我们可以回过头来看看，过去人们认为的阻碍是什么。

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

嗯。

Mhmm.

最初，我们提出了预训练的扩展定律。

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

你知道，人们当时认为——这很合理——我们拥有的高质量数据量会限制我们所能达到的智能水平。

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

而这条扩展定律非常重要。

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

模型越大，相应地需要越多的数据，才能产生更智能的AI。

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

这就是预训练阶段。

And so that was pretraining.

埃利阿斯·苏斯科弗说，我们没数据了，类似这样的话。

And Elias Susskover, Elias said, we're out of data or something like that.

预训练结束了，之类的话。

Pre training is over or something like that.

整个行业恐慌了，觉得这是AI的末日。

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

当然，这显然不是真的。

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

我们会继续增加用于训练的数据量。

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

其中很多数据可能是合成的，这也让很多人感到困惑。

A lot of that data is probably gonna be synthetic, and that also confused people.

你知道的？

You know?

人们没有意识到的是，他们已经忘了，我们用来训练、互相教导、互相交流的大部分数据本质上都是合成的。

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

你知道吧？

You know?

它是合成的，因为它不是自然产生的。

I it's synthetic because it didn't come out of nature.

是你创造的。

You created it.

我在使用它。

I'm consuming it.

我修改它，增强它。

I modify it, augment it.

我重新生成它。

I regenerate it.

其他人也在使用它。

Somebody else consumes it.

因此，我们现在达到了一个水平，AI能够获取真实数据，对其进行增强和合成生成海量数据，而这种后训练部分仍在持续扩展。

And so 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.

因此，我们可以使用的由人类生成的数据量将越来越小。

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

我们用于训练模型的数据量将继续增长，直到不再受数据限制，而是受计算能力限制。

The amount of data that we use to train model is gonna continue to scale to the point where we're no longer limited training is no longer limited by data is now limited by compute.

之所以如此，是因为大部分数据都是合成的。

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

接下来的阶段是推理阶段。

Then the next phase is test time.

我仍然记得有人告诉我，推理？哦，是的。

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

那很简单。

That's easy.

预训练才是难点。

Pre pretraining, that's hard.

这些是人们正在讨论的庞大系统。

These are giant systems that people are talking about.

推理必须很简单。

Inference must be easy.

所以推理芯片将会是小小的芯片，你知道的，它们不像英伟达的芯片。

So inference chips are gonna be little tiny chips, and, you know, they're not they're not like NVIDIA's chips.

哦，那些芯片会很复杂且昂贵。

Oh, those are gonna be complicated and expensive.

而且，你知道，我们可以制造——嗯，是这样。

And, you know, we could make and this is Mhmm.

在未来，推理将成为最大的市场，而且会很简单，我们会把它变成大宗商品。

And in the future, inference is gonna be the biggest market, and it's gonna be easy, and we're gonna commoditize it.

你知道，每个人都可以制造自己的芯片。

You know, everybody can build their own chips.

这对我来说一直很不合逻辑，因为推理就是思考。

And and and that was always illogical to me because inference is thinking.

我认为思考是困难的。

And I think thinking is hard.

思考远比阅读困难。

Thinking is way harder than reading.

是的。

Mhmm.

你知道，预训练只是记忆和泛化，寻找模式和关系。

You know, pretraining is just memorization and generalization, you know, and looking for patterns and relationships.

你是在阅读，而不是思考、推理、解决问题，面对那些未曾经历的新体验，将其分解为可解决的片段，然后通过第一性原理推理，或借助以往的经验、案例，或者通过探索和搜索，尝试不同的方法。

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

整个测试时扩展、推理的过程，本质上就是思考。

That whole process of post of of test time scaling, inference is really about thinking.

而且是关于推理的。

And and it's about reasoning.

是关于规划的。

It's about planning.

是关于搜索的。

It's about search.

这还涉及到，那怎么可能算是计算轻量的呢？

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

我们当时对这一点绝对是正确的。

And we were absolutely right about that.

你知道吗？

You know?

所以，推理时的扩展是非常耗费计算资源的。

So so test time scaling is intensely compute intensive.

那么问题来了，好吧。

Then the question is, okay.

我们现在处于推理和推理时扩展阶段。

Now we're at inference and we're at test time scaling.

那之后又是什么呢？

What's beyond that?

显然，我们现在已经创造出了一个具有代理能力的人。

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

而这个单一的智能体拥有一个我们现在已经开发出来的大语言模型。

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

但在推理阶段，这个智能体系统会自行去开展研究、访问数据库，并使用各种工具。

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

它最重要的一项工作是生成并启动大量子智能体，这意味着我们现在正在组建大型团队。

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.

通过雇佣更多员工来扩展NVIDIA，比我自己扩展要容易得多。

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

是的。

Mhmm.

因此，下一个扩展定律，就是智能体扩展定律。

And so the next scaling laws, the agentic scaling law.

这有点像是对AI进行乘法级放大。

It's kind of like multi multiplying AI.

是的。

Mhmm.

通过乘法提升AI，我们可以像你希望的那样快速生成代理。

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

所以，你知道，你现在有四种扩展法则。

And so, you know, I you're have four scaling laws.

当我们使用智能代理系统时，它们会产生大量数据。

And and as we use the age agentic systems, they're gonna create a lot more data.

它们会创造大量的经验。

They're gonna create a lot of experiences.

其中一些，我们会说，哇。

Some of it, we're gonna say, wow.

这真的很好。

This is really good.

我们应当记住这一点。

We ought to memorize this.

是的。

Mhmm.

这个数据集然后会回到预训练阶段。

That dataset then comes all the way back to pretraining.

我们对其进行记忆和泛化。

We memorize and generalize it.

接着我们会对其进行优化，并微调回后训练阶段。

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

然后我们通过推理时的处理，以及智能体系统，进一步增强它并将其应用到产业中。

Then we enhance it even more with test time, you know, and the agents agents agentic systems, you know, put it on to the industry.

因此，这个循环将会不断持续下去。

And so this loop, this cycle is gonna go on and on and on.

本质上，智能的扩展将依赖于一个因素，那就是计算力。

It kinda comes down to, basically, intelligence is gonna scale by one thing, and it's compute.

但这里有一个棘手的问题，你需要提前预判和预测，因为这些组件中有些需要不同的硬件才能最优地运行。

But there's a tricky thing there that you have to anticipate and predict, which is some of these components requires different kind of hardware to really do it optimally.

因此，你必须预判人工智能创新将走向何方。

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

例如，确保通过稀疏性进行分布。

For example, make sure that's dispersed with sparsity.

完美。

Perfect.

在硬件方面，你不能在一周内就突然改变方向。

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

你必须提前预判它会是什么样子。

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

这太可怕了，也太难做到了。

That's that's so scary and difficult to do.

对吧？

Right?

例如，这些AI模型架构大约每六个月就会被发明一次。

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

是的。

Yeah.

对吧？

Right?

而系统架构和硬件架构大约每三年更新一次。

And system architectures and hardware architectures kinda every three years.

所以你需要预测两三年后可能发生的状况。

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

嗯。

Mhmm.

有几种方法可以做到这一点。

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

首先，我们可以自己开展内部研究，这也是我们进行基础研究的原因之一。

First of all, we could do research internally ourselves, and that's one the reasons why we have basic research.

我们也有应用研究。

We have applied research.

嗯。

Mhmm.

我们自己创建模型。

We create our own models.

因此，我们在这里拥有第一手的实践经验。

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

这正是我所说的协同设计的一部分。

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

我们也是全球唯一一家与几乎所有AI公司合作的AI公司。

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

在可能的范围内，我们努力了解人们所面临的问题。

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

所以你在倾听整个行业、AI实验室的风声？

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

没错。

That's right.

你必须倾听并从每个人身上学习，然后最后一部分是拥有一种灵活的架构，能够随势而变。

You gotta listen and and learn from everybody, and have a have a and then the the last part is to have an architecture that's that's flexible, that can adapt and move with the wind.

CUDA 的一个优势在于，它一方面是一个令人难以置信的加速器。

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

另一方面，它也非常灵活。

On the other hand, it's really flexible.

因此，这种在专业化与通用化之间的绝佳平衡至关重要——专业化让我们能够加速 CPU，而通用化则让我们能够适应不断变化的算法。

And so that balance incredible balance between specialization, otherwise, we can't accelerate the the CPU, versus generalization so that we can adapt with changing algorithms, that's really, really important.

这正是 CUDA 为何如此坚韧，同时我们还能持续改进它的原因。

That's the reason why why CUDA has been so resilient on the one hand, and yet we continue to enhance it.

我们现在已进入 CUDA 13.2 版本，我们正在以极快的速度演进架构，以便跟上现代先进算法的步伐。

We're at CUDA 13.2, and so we're involve evolving the architecture so fast that we can stay with, you know, with with the with the modern art algorithms.

例如，当混合专家模型问世时，这就是为什么我们采用了 NVLink 72 而不是 NVLink 8。

For example, when mixed use experts came out, that's the reason why we had NVLink 72 instead of NVLink eight.

我们现在可以将一个包含 4 万亿甚至 10 万亿参数的模型，完整地部署在一个计算域中，就像它运行在单张 GPU 上一样。

We could now take an entire 4,000,000,000,000, 10,000,000,000,000 parameter model and put it in one computing domain as if it's running on one GPU.

人们可能没注意到我这么说，但如果你仔细观察 Grace Blackwell 机架的架构，就会发现它完全专注于一件事：处理大语言模型。

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.

突然间，一年后，你看到的是维拉·鲁宾机架。

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

它配备了存储加速器。

It has storage accelerators.

它搭载了一款名为维拉的全新CPU。

It has this incredible new CPU called Vera.

它拥有维拉·鲁宾和NVLink 72，用于运行大语言模型。

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

它还配备了一个名为Grok的新机架。

It also has this new additional rack called Grok.

因此，整个机架系统与之前的版本完全不同，集成了所有这些新组件。

And so this entire rack system is completely different than the previous one, and it's got all these new components in it.

之所以如此，是因为上一代是为运行MOE大语言模型推理设计的，而这一代则是为运行智能代理设计的。

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

而智能代理会频繁调用工具。

And agents bang on tools.

而且

And

显然，这个系统的设计必须在clawed code、codex、Open Claw之前就已完成，你实际上是在预见未来。

Obviously, the design of the system had to have been done before clawed code, codex, Open Claw, that you were anticipating the future, essentially.

是的。

Yeah.

而这来自于什么？

And that and that comes from what?

来自于对前沿技术的风声和理解，不是吗？

From the whispers, from the understanding what all the state of the art is No.

其实没那么复杂。

It's it's easier than that.

你只需要进行推理。

You you just reason about it.

首先，你只需要进行推理。

First of all, you you just reason.

不管发生什么，总有一个时刻，为了让这个大型语言模型成为数字员工，我们就用这个比喻吧。

No matter no matter what happens at some point, in order for that large language model to be a digital worker let's just let's just use that metaphor.

假设我们希望这个语言模型成为一个数字员工。

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

它需要做些什么？

What does it have to do?

它必须能够访问真实数据。

It has to access ground truth.

那就是我们的文件系统。

That's our file system.

它必须能够进行研究。

It has to be able to do research.

它不可能什么都知道。

It doesn't know everything.

我们不可能等，也不愿意等到这个AI变得对过去、现在和未来的一切都普遍聪明了，才让它变得有用。

We don't have and I don't wanna wait until this AI becomes, you know, universally smart about everything past, present, and future before I make it useful.

因此，我还不如让它去进行研究。

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

显然，如果它想帮我，就必须使用我的工具。

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

很多人会说，人工智能会彻底摧毁软件。

You know, a lot of people would say, you know, AI is gonna completely destroy software.

我们再也不需要软件了。

We don't need software anymore.

我们甚至不需要任何工具了。

We don't even need tools anymore.

这太荒谬了。

That's ridiculous.

让我们来做个思想实验吧。

Let's let's use the let's use a thought experiment.

你可以坐在那里，喝着一杯威士忌，思考所有这些问题，然后一切就会变得显而易见。

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

比如说，如果我要在未来的十年内创造出我们能想象的最棒的智能体，假设它是一个人形机器人。

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

如果这个人形机器人真的被创造出来了，它更有可能走进我家，使用我现有的工具来完成它需要做的工作吗？

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

还是说，它的手会瞬间变成一把十磅重的锤子，又在另一个时刻变成一把手术刀？

Or does this hand turns into a 10 pound hammer in one instance, turns into a scalpel in another instance?

为了烧水，它直接从手指里发射出微波？

And in order to boil water, it beams, you know, microwaves out of its fingers.

你知道吧？

You know?

还是说，它更可能只是去用微波炉？

Or is it more likely just to use a microwave?

你知道吧？

You know?

而当它第一次走到微波炉前时，很可能根本不知道怎么用。

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

但这没关系。

But that's okay.

它连接着互联网。

It's connected to the Internet.

它会阅读这台微波炉的说明书，读完后立刻成为专家，嗯。

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

所以它会使用它。

And so it uses it.

因此，我认为我刚刚描述的，实际上几乎涵盖了OpenCLaw的所有特性。

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

嗯。

Mhmm.

你知道吗？

You know?

它会使用工具，会访问文件，能够进行研究。

That it's gonna use tools, that it's gonna access files, it's gonna be able to do research.

它拥有I/O子系统。

It has IO subsystem.

当你以这种方式反复思考和推敲之后，你会说：天哪。

And when you're done reasoning through it reasoning about it through through it in that way, then you say, oh my gosh.

这对未来计算的影响极为深远。

The impact to the future computing is deeply profound.

之所以如此，是因为我认为我们刚刚重新发明了计算机。

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

然后你现在会说：好吧。

And then now you say, okay.

我们什么时候开始思考过这个问题？

When did we reason about that?

我们什么时候思考过OpenCLaw？

When did we reason about OpenCLaw?

如果你拿我在GTC上使用的OpenCLaw架构图来看，你会发现它早在两年前就已存在。

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

就在两年前的GTC大会上，我就在谈论与如今OpenClaw完全一致的智能体系统。

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

当然，许多因素必须同时汇聚在一起。

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

首先，我们需要Claude、GPT以及所有这些模型达到一定的能力水平。

First of all, we needed Claude and and GPT and, you know, all of these models to reach a level of capabilities.

因此，它们的创新、突破和持续进步至关重要。

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

然后，当然，必须有人创建一个开源项目，这个项目足够稳健、足够完整，让我们都能实际应用。

And then, of course, somebody had to create a an open source, you know, project that that was sufficiently robust, you know, and sufficiently complete, and that we can all we can all put to put to work.

我认为，OpenClaw之于智能体系统，就像ChatGPT之于生成式系统一样，这真的是一件大事。

And and I think Open Claw did for did for agentic systems what chat GPT did for generative systems, and and I just think it's a very big deal.

是的。

Yeah.

这是一个非常特别的时刻。

It's a really special moment.

我不太确定为什么它吸引了全世界如此多的关注，但它带来的影响远超Coda和Codex等工具。

I'm not exactly sure why it captured so much of the world's attention, but it did more than clogged code and codex and so on.

因为普通消费者能够接触到它。

Because consumers could reach it.

当然。

Sure.

是的。

Yeah.

但这其中还有很多是氛围和感觉，彼得，我和他做过一期播客。

But there there's also so much of this is vibes, and and Peter, I had a podcast with him.

一个非常棒的人。

A wonderful human being.

所以，这部分也在于代表这些事物的人。

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

毫无疑问。

And No doubt.

其中一部分是迷因，因为我们都在试图理解它。

Part of it is memes and the because we're all trying to figure it out.

当你拥有如此强大的技术时，如何在让它们为你做有用的事情的同时保护你的数据，这涉及严肃而复杂的安全问题，而这些也伴随着令人担忧的方面。

There's really serious and complicated security concerns about when you have such powerful technology, how do you hand over your data so they can do useful stuff, but then there's scary things associated with that.

作为人类文明、作为个体，我们都在努力寻找这个正确的平衡点。

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

是的。

Yeah.

我们立刻就行动了，并派了一群安全专家去处理这件事。

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

嗯。

Mhmm.

我们做了一件叫OpenShell的事情。

And we did this thing called OpenShell.

它已经被集成到OpenClaw中了。

It's it's already been integrated into into OpenClaw.

NVIDIA 推出了 NemoClaw。

And NVIDIA put forward NemoClaw.

是的。

Yep.

没错。

Exactly.

它们安装起来非常简单。

They install super easy.

它确保了安全性。

It makes sure that it's secure.

我们给你三个权限中的两个。

We give you two out of three rights.

智能系统可以访问敏感信息。

Agenic systems can can access sensitive information.

它可以执行代码，并且可以与外部通信。

It can execute code, and it can communicate externally.

是的

Mhmm.

如果我们随时只赋予你这三种能力中的两种，而不是全部三种，我们就能确保安全。

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

在这三种能力中的两种中，我们还会根据企业赋予你的权限提供访问控制。

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

然后我们将它连接到这些企业 already 拥有的策略引擎。

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

因此，我们将尽力帮助 Open Claw 成为更好的 Claw。

And so we're gonna try to do our best to to help Open Claw become a a better claw.

你生动地解释了我们过去如何克服了许多原本以为会成为障碍的问题。

So you eloquently explained how we have a long history of blockers that we thought were gonna be blockers, we overcame them.

但展望未来，你认为可能会出现哪些新的障碍？

But now looking into the future, what do you think might be the blockers?

既然现在已经明确代理将无处不在。

Now that it's clear that agents will be everywhere.

所以显然我们需要计算资源。

So it's obviously we're gonna need compute.

那么，这种扩展的瓶颈会是什么？

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

电力是一个问题，但并不是唯一的问题。

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

但这就是我们大力推动极致协同设计的原因，以便每年都能将每瓦特的令牌处理量提升数个数量级。

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

在过去十年中，摩尔定律使计算能力提升了约100倍。

And so in the last ten years, Moore's law would have progressed computing about a 100 times in the last ten years.

而我们在过去十年中将计算能力提升了百万倍。

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

因此，我们将继续通过极致协同设计来实现这一目标。

And so we're gonna keep on we're gonna keep on doing that through extreme codesign.

因此，能效和每瓦性能完全影响公司的收入。

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

这会影响工厂的收入，我们会将这一点推向极限，以便尽可能快地降低每个令牌的成本。

It affects the revenues of a factory, and we're just we're just gonna push that to a limit so that we could keep on driving token cost down as fast as we can.

你知道，我们的计算机价格在上涨，但我们的令牌生成效率提升得更快，因此每个令牌的成本正在下降。

You know, the our computer price is going up, but our token generation effectiveness is going up so much faster that token cost is coming down.

它每年都在下降一个数量级。

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

所以电力，这是一个有趣的问题。

So power, that's an interesting one.

为了解决电力瓶颈，我们需要通过提高每瓦特的每秒令牌数，让效率变得越来越高。

So the the way to try to get around the power blocker is to try to with the tokens per second per watt, try to make it more and more efficient.

当然，还有一个问题是我们如何获得更多的电力。

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

我们也应该获得更多电力。

We should also get more power.

那是一个

That's a

非常复杂的一个问题。

really complicated one.

你提到过小型模块化核电厂。

You've talked about small module nuclear power plants.

关于能源，有各种各样的想法。

There's all kinds of ideas for energy.

供应链中的瓶颈让你多晚睡不着觉？比如ASML的EUV光刻机、台积电的先进封装技术如Colossus，以及SK海力士的高带宽内存。

How much does it keep you up at night, The the bottlenecks in the supply chain of AI, like ASML with EUV lithography machines, TSMC with advanced packaging, like Colossus, and SK Hynix with high bandwidth memory.

一直如此。

All the time.

我们一直在努力解决这个问题。

And we're working on it all the time.

历史上从未有任何一家公司能在如此规模上增长，同时还能加速这一增长。

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

这太不可思议了。

It's incredible.

是的

Yeah.

人们甚至很难理解这一点。

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

在整个人工智能计算领域，我们的市场份额正在增长。

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

因此，上下游供应链对我们来说至关重要。

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

我花了很多时间向我合作的各位首席执行官传达这些信息。

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

哪些动态将推动增长持续甚至加速？

What are the dynamics that that's gonna cause the growth to continue or even accelerate?

这正是为什么我右侧的几乎所有人，都是上游整个IT行业和下游整个基础设施行业的首席执行官的原因之一。

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

嗯

Mhmm.

而且当时在场的有好几百位CEO。

And they were all there were several 100 CEOs.

我认为以前从未有过这么多CEO出席的主题演讲。

And I don't think there's ever been keynotes where several 100 CEOs show up.

其中一部分原因是我正在向他们介绍我们当前的业务状况。

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

我正在向他们阐述近期的增长驱动力以及正在发生的变化。

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

同时，我也在描述我们下一步的走向，以便他们能利用这些信息和所有相关动态来指导他们的投资决策。

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

因此，我以这种方式向他们传达信息，就像我向自己的员工传达一样。

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

然后，当然，我还会亲自去拜访他们，确保他们明白。

And then, of course, then I make trips out to them and make sure that, hey.

听好了。

Listen.

我想让你们知道，本季度、即将到来的一年、下一年，这些事情都会发生。

I want you to know this quarter, this coming year, this next year, these things are gonna happen.

如果你看一下DRAM行业的首席执行官们，世界上最大的DRAM产品是用于CPU和数据中心的DDR内存。

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

大约三年前，我成功说服了几位首席执行官，尽管当时HBM内存使用非常少，仅限于超级计算机，但它未来将成为数据中心的主流内存。

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 and barely by supercomputers, that this was going to be a mainstream memory for data centers in the future.

起初，这听起来很荒谬，但几位首席执行官相信了我，并决定投资开发HBM内存。

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

另一种被奇怪地引入数据中心的内存是手机使用的低功耗内存，我们希望将它们改造用于数据中心的超级计算机。

Another memory was rather odd to put into a data center is the low power memories that we use for cell phones, and we wanted them to adapt them for supercomputers in the data center.

他们说：‘手机内存用于超级计算机？’

And they go, cell phone memory for supercomputers?

嗯。

Mhmm.

我向他们解释了原因。

And and I explained to them why.

看看这两种内存，LPDDR5和HBM4。

Well, look at these two memories, LPDDR five, HBM four.

它们的出货量简直惊人。

The volumes are so incredible.

这三家公司都创下了历史最高纪录，而它们都是已有四十五年历史的企业。

All three of them had record years in history, and these are these are forty five year old companies.

所以，你知道，我的工作一部分就是传递信息、塑造方向、激发灵感。

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

你知道吧？

You know?

所以你不仅仅是预见并可能激励了英伟达以及公司的不同工程师。

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

你是在塑造未来的供应链。

You're you're manifesting the supply chain of the future.

所以你正在与台积电、阿斯麦进行对话。

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

上游，下游。

Upstream, downstream.

上游，下游。

Upstream, downstream.

所以，这就是EV的问题，

So that's the thing EV,

卡特彼勒。

Caterpillar.

是的。

Yeah.

这属于我们下游的环节。

That's downstream from us.

是的。

Yeah.

是的。

Yeah.

就在那儿，是的。

There you Yeah.

整个事情。

The whole thing.

我的意思是，但半导体行业确实涉及大量极其复杂的工程，供应链如此错综复杂、组件如此之多，简直让人感到害怕，但它 somehow 就能正常运转。

I mean, but that's so there's so much incredibly difficult engineering that happens in the the entire semiconductor industry, and it just feels scary how intricate the supply chain is, how many components there are, but it works somehow.

没错。

Exactly.

深层的科学，还有

The deep science, the

深层的工程、惊人的制造技术，而大部分制造过程已经实现了机器人化。

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

但我们有几百家供应商，为我们的130万个组件机架提供技术支持。

But we have a couple of 100 suppliers that contribute the technology that goes into our 1,300,000 component rack.

嗯。

Mhmm.

每个机架包含一百三十万到一百五十万个组件。

Each rack is 1.3, one and a half million components.

在Verirubin机架中，有200家供应商提供支持。

There are 200 suppliers across the Verirubin Rack.

有趣的是，你并没有把这一点列为你晚上睡不着觉的障碍因素。

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

但我正在做所有必要的事情来

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

好的。

Okay.

去看一下？

To see?

我可以安心睡觉了，因为我已经确认过了。

I can go to sleep because I checked it off.

我说，你知道吗，我可以安心睡觉了，我想，好吧，我们来看看。

I said, You know, I I go I I can go to sleep, I go, well, let's see.

我们来理性分析一下这个问题。

What read let's reason about this.

对我们来说什么是重要的？

What's important for us?

嗯。

Mhmm.

因为，好吧。

Because okay.

我们来理性分析一下这个问题。

Let's reason about this.

因为我们已经将系统架构从你记得的原始DGX架构改为NVLink 72机架级计算。

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

嗯。

Mhmm.

那这意味着什么？

What's gonna what does that what does that mean?

什么

What

意思

does

这对软件意味着什么？

that mean to software?

这对工程意味着什么？

What does that mean to engineering?

这对我们的设计和测试意味着什么？

What does that mean to how we design and test?

这对供应链意味着什么？

How and what does that mean to the supply chain?

这意味着我们把超级计算机在数据中心的集成工作，转移到了供应链中的超级计算机制造环节。

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

嗯

Mhmm.

如果你这么做，你还必须意识到，你将要转移的是什么——比如，假设你希望建成的数据中心总共需要50吉瓦的超级计算机同时运行，而制造这50吉瓦的超级计算机需要一周时间，那么在供应链的每个星期里，这些超级计算机都需要一吉瓦的电力。

If you're doing that, you also have to recognize you're gonna move what and and if if if your if your, you know, total footprint of whatever data center you're gonna build, let's say, you would like to have, you know, 50 gigawatts of supercomputers that are running simultaneously, and it takes one week to manufacture that 50 gigawatts of supercomputers, then each week in the supply chain, the supercomputers are gonna need a gigawatt of power.

所以我们需要供应链提升其电力供应能力，以便在发货前完成超级计算机的制造和测试。

And so so we're gonna need the supply chain to increase the amount of power it has to build, test to build and test the supercomputers in the supply chain before I ship it.

嗯。

Mhmm.

NVLink 72 真正做到了在供应链中制造超级计算机，并以每机架两到三吨的规模发货。

Well, NVLink seventy two literally builds supercomputers in the supply chain and ships them two, three tons at a time per rack.

过去，这些设备都是以零部件形式运来，我们再在数据中心内进行组装。

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

但现在这已经不可能了，因为NVLink 72的密度实在太高了。

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

所以这是一个例子，我得深入说明——我会飞去供应链现场，见我的合作伙伴，然后说：

And so that's an example, and I would have to go into you know, I fly into the supply chain, go meet my partners, and, hey.

我说，你猜怎么着？

I said, guess what?

所以，我们过去建造DGX的方式是这样的。

So here's what I'm gonna do with this is the way we used to build our DGXs.

我们现在要这样建造它们。

We're gonna build them this way.

这样会好得多，因为我们将来需要它们来做推理。

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

推理的市场正在到来。

The market for inference is, you know, coming.

推理的转折点即将到来。

The inflection point for inference is coming.

这将是一个巨大的市场。

It's gonna be a big market.

所以我首先向他们解释了正在发生什么，以及为什么会这样。

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

然后我要求他们各自进行数十亿美元的资本投资。

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

因为他们信任我，而且我非常尊重他们，我会给他们充分的机会提问，花时间向他们解释事情，并进行理性分析。

And because they, you know, they trust me, and and I I I'm very respectful of them, and I I give them every opportunity to question me, and I spend time to explain things to people, and I reason about it.

我会画图给他们看，并用第一性原理进行推理。

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

等我讲完的时候，他们就知道该做什么了。

And by the by by the time I'm done with them, there's know what to do.

所以这很大程度上关乎人际关系和建立对未来的共同认知。

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

是的。

Yeah.

但你是否担心某些瓶颈问题？

But do you worry about certain bottlenecks?

我的意思是，供应链中最大的瓶颈是什么？

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

你是否担心ASML的EUV设备？

Are you are you worried about ASMLs, EUV tooling?

你担心台积电的封装成本以及其扩展速度吗？

Are you are you worried about the the packaging, cost packaging of TSMC, about how fast it could scale?

正如你所说，你不仅增长得极其迅速，而且还在加速增长。

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

所以，供应链中的每个人，这些无疑都是瓶颈，都必须扩大规模。

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

你和他们有过沟通吗？

Are you having conversations with them?

比如，你们如何能更快地扩大规模？

Like, how can you scale up faster?

你对此感到担忧吗？

Do you worry about it?

不担心。

No.

好的。

Okay.

因为我告诉了他们我的需求。

Because because I told them what I needed.

他们理解了我需要什么。

They understood what I need.

他们告诉我他们会做什么，我相信他们会做到。

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

有意思。

Interesting.

是的。

Yeah.

听到这个真好。

That's great to hear.

所以，也许我们可以再稍微聊聊电力问题。

So maybe if we can just linger on the power for a little bit.

你对解决能源问题有什么期望？

What are your hopes for how to solve the energy problem?

莱克斯，我想和我们讨论的一个领域是，我非常希望我们能传达这个信息。

One of the areas, Lex, that I'm I'm that I would love I would love love us to talk about and just get the message out.

你知道，我们的电网是按照最恶劣的情况设计的，并留有一定的余量。

You know, our our our power grid is designed for the worst case condition with some margin.

但在99%的时间里，我们远未达到最恶劣的情况，因为最恶劣的情况只发生在冬季的几天、夏季的几天以及极端天气时。

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.

大多数时候，我们离最恶劣的情况相去甚远，实际运行大概在峰值的60%左右。

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

因此，99%的时间里，我们的电网都有多余的电力，这些电力却处于闲置状态。

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

但它们必须保持待命，以防万一——当关键时刻到来时，医院必须供电，基础设施必须运行，机场也必须正常运作，等等。

But they have to be there sitting idle because just in case, when the time comes, hospitals have to be powered and, you know, infrastructure has to be powered and airports have to run and so on and so forth.

所以我的问题是，我们能否帮助他们理解并建立合同协议，设计计算机架构和数据中心，使得当社会基础设施需要最大电力时，数据中心的用电量可以相应减少。

And so the question that I have is whether we could go and help them understand and create contractual agreements and design computer architecture systems, data centers, such that when they need the maximum power for infrastructure in society, that the data centers would get less.

嗯。

Mhmm.

但这种情况本身就已经非常罕见了。

But that's in a very rare instance anyways.

在那段时间里，我们要么为这部分使用备用发电机，要么让计算机将工作负载转移到其他地方，要么让计算机运行得慢一些。

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

我们可以降低性能，减少功耗，在有人请求答案时，稍微增加一点延迟响应时间。

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

因此，我认为这种使用计算机、建设数据中心的方式——不再追求100%的正常运行时间以及那些极其严苛的合同——给电网带来了巨大压力，迫使电网必须从最大负荷进一步提升。

And so I think that that that way of using computers, of building data centers, instead of expecting a 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 gonna have to increase from their maximum.

我只是想利用它们的多余产能。

I just wanna use their excess.

嗯。

Mhmm.

它就白白闲置在那里。

It's just sitting there.

是的。

Yeah.

这一点讨论得还不够。

That's not talked about enough.

那么，是什么阻碍了这一点呢？

So what's what's this what's stopping there?

是监管问题吗？

Is it regulation?

是

Is it

官僚主义吗？

bureaucracy?

我认为这是一个三方面的问题。

I think it's it's a three way problem.

问题始于最终用户。

It starts with the end customer.

最终用户对数据中心提出了永远不能停机的要求。

The end customer puts puts requirements on the data centers that they can never not be available.

好吗？

Okay?

因此，最终客户期望的是完美。

So that the end customer expects perfection.

为了实现这种完美，你需要备用发电机和电网供应商共同协作来达成完美。

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

所以每个人都必须达到六个九的可用性。

And so everybody's gotta have six nines.

嗯哼。

Mhmm.

我认为，首先，现在我们必须让每个人都明白，当客户提出这些要求时，你的数据中心运营团队中有人与CEO脱节了。

Well, I think, first of all, right now, we ought to have everybody understand that when the customer asks for these things, you got somebody you have somebody in your data center operations team disconnected from the CEO.

我打赌CEO并不知道这一点。

I bet the CEO doesn't know this.

我要跟所有CEO谈谈。

I'm gonna talk to all the CEOs.

首席执行官们可能根本没留意那些正在签署的合同。

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

所以每个人都想签下最好的合同，这当然了。

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

他们去找云服务提供商，合同由两位谈判代表负责——我都能想象到他们现在的情形，嗯。

And they go down to the cloud service providers and the contract the the two contract negotiators that are you I could just see them now Mhmm.

你知道的，在谈判这些多年期合同。

You know, negotiating these multiyear contracts.

双方都想要最好的合同，你知道的。

Both sides want, you know, the best contract.

因此，云服务提供商不得不去找公用事业公司，并期望达到六个九的可靠性。

As a result, the CSPs then have to go down to the utilities, and they expect the nine the six nines.

所以我认为，首先要确保所有客户——客户的首席执行官们——都明白他们所要求的是什么。

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

第二件事是我们必须建造能够优雅降级的数据中心。

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

因此，如果电力公司或电网告诉我们：‘我们需要将你的用电量下调至80%左右’，我们会说：‘没问题，完全没问题。’

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

是的。

Mhmm.

我们只是把工作负载重新分配一下。

We're just gonna move our workload around.

我们会确保数据不会丢失，但可以降低计算速率，减少能耗。

We're gonna make sure that data's never lost, but we can reduce the computing rate and use less energy.

对于关键工作负载，服务质量会略有下降。

The quality of service degrades a little bit for the critical workloads.

我会立刻把它们迁移到别处，这样就不会有这个问题了。

I shift that somewhere else right away, so I don't have that problem.

所以，无论哪个数据中心仍能保持100%的正常运行时间。

And so, you know, whoever whichever data center still has a 100% uptime.

所以

And so

在数据中心中实现这种智能动态电力分配，工程难度有多大？

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

只要你能明确需求，就能设计出来。

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

说得真好。

Beautifully put.

只要它符合物理学的基本原理，我认为就没问题。

So long as it obeys the laws of physics on first principles, I think we're good.

你刚才提到的第三点是什么？

What was the

你刚才提到的第三点是什么？

third thing you were mentioning?

第二点是数据中心。

So the second thing is the the data centers.

第三点是我们需要公用事业公司也认识到这是一次机遇。

And the third thing is we need the utilities to also recognize that this is an opportunity.

嗯嗯。

Mhmm.

而不是说，看，我要花五年时间才能提升我的电网能力。

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

如果你愿意接受这种级别的供电保障，我可以在下个月为你提供，并且价格是这样的。

If you if you have, if you're willing to take power of this level of guarantee, I can make them available for you next month and at this price.

如果公用事业公司也能提供更多样化的电力供应承诺，那么我想每个人都会知道该如何利用它。

And so if utilities also offered more segments of power delivery promises, then I think everybody will figure out what to do with it.

是的。

Yeah.

但现在的电网浪费实在太严重了。

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

我们应该去解决这个问题。

We we should go after it.

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，可能仅用四个月就创造了纪录。

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

现在已经有二十万块GPU，并且增长得非常迅速。

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

你能谈谈他的方法吗？这对所有数据中心建设者都有启发意义，是什么让他能实现这样的成就？

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

他的工程方法，他对整个建设管理的方式，方方面面。

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

首先，马斯克深入钻研了这么多不同领域，同时他也是一个非常出色的系统思考者。

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

嗯。

Mhmm.

因此，他能够跨越多个学科进行思考。

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

而且他显然会推动一切，质疑每一件事：第一，这有必要吗？

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

第二，非得这么做不可吗？

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

还有这些数字，你知道，真的需要这么长时间吗？

And the numbers, you know, does it have does it have to take this long?

所以他有能力质疑一切，直到每件事都缩减到最必要的最小程度。

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

再削减任何东西都不可能了。

You can't take anything else out.

但产品的必要功能却依然保留着，你知道的。

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

因此，他可以说是你能想象到的最极简主义者，并且他是在系统层面做到这一点的。

And so he's he is as minimalist as you could possibly imagine, and he does it at a system system scale.

我也非常喜欢他确实有在场这一点。

I I I also love the fact that he he is he is represented.

他亲临行动现场。

He he is he is present at the point of action.

是的。

Mhmm.

你知道，他就会直接过去。

You know, he'll just go there.

如果有问题，他会直接过去，然后给我指出问题。

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

你知道，当你把所有这些结合起来时，就能克服很多以前‘我们一直就这样做’的惯性。

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

嗯。

Mhmm.

你知道，我正在等他们。

You know, I'm I'm waiting for them.

你知道吗？

I you know?

我的意思是，每个人都有很多借口。

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

所以，最后一点是，当你亲自以如此紧迫的态度行动时，就会促使其他人也紧迫起来。

And so and and then the last thing is when when you act personally with so much urgency, it causes everybody else to act with urgency.

你知道的。

You know.

而且每个供应商都有很多客户在忙。

And and every supplier has a lot of customers going on.

每个供应商都有很多项目在进行。

Every supplier has a lot of projects going on.

他总是让别人觉得，他的事是所有人项目中最重要的，你知道的。

And he he make it he made it he makes it his business that he's the top priority of everybody else's, you know, projects.

所以他通过实际行动来体现这一点。

And so he he does that by demonstrating it.

是的。

Yeah.

我参加过很多这样的会议。

I've been in a bunch of those meetings.

这挺有意思的，因为真的没多少人会问这样的问题，比如：好吧。

This is it's fun to watch because, really, not enough people ask the question, like, okay.

所以这能做得更快吗？怎么做？

So can this be done a lot faster, and how?

为什么非得花这么长时间？

Why does it have to take this long?

是的。

Yeah.

没错。

Right.

然后这通常就变成了一个工程问题。

And then that becomes an engineering question often.

是的，我想起有一次我和他在一起时，他正在亲历整个将电缆插入机架的过程。

And, yes, I think when you get the ground truth of actually I remember one of the times I was hanging out with him, he literally is going through the entire process of how to plug in cables into a rack.

他当时正和现场的一名工程师一起做这项任务，只是想了解这个流程究竟是怎样的，以便减少错误。

And he's was was working with an engineer on the ground that's doing that task, and he's just trying to understand what does that process look like so it can be less error prone.

并通过参与数据中心建设中的每一项任务来建立这种直觉。

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