# Jensen Huang: NVIDIA - The $4 Trillion Company & the AI Revolution | Lex Fridman Podcast #494

https://www.youtube.com/watch?v=vif8NQcjVf0
Translation: zh-CN

[00:00] 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.
  以下是与英伟达首席执行官黄仁勋的对话，他是人类文明史上最重要、最具影响力的公司之一。

[00:11] 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.
  英伟达是推动人工智能革命的引擎，其许多成功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卓越的投资和决策。

[00:26] This is the Lex Freedman podcast.
  这是 Lex Freedman 播客。

[00:28] And now, dear friends, here's Jensen Huang.
  现在，亲爱的朋友们，让我们欢迎黄仁勋。

[00:33] You've propelled Nvidia into a uh new era in AI, moving beyond his focus on chip scale design to now rack scale design.
  您已将英伟达推入人工智能的一个新时代，从专注于芯片规模设计转向现在的机架规模设计。

[00:41] And I think it's fair to say that uh 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，在很长一段时间里，英伟达的胜利在于打造出最好的 GPU。

[00:47] and you still do, but now you've expanded that to 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
  您仍然这样做，但现在您已将其扩展到 GPU、CPU、内存、网络、存储、电源、散热、软件、机架本身、您已宣布的 Pod，甚至数据中心的极端协同设计。

[01:01] Center. So, let's talk about extreme code design.
  中心。那么，我们来谈谈极端代码设计。

[01:03] What uh is the hardest part of uh co-designing a system with that many complex components and design variables?
  在设计一个拥有如此多复杂组件和设计变量的系统时，最难的部分是什么？

[01:10] Yeah, thanks for that question.
  是的，谢谢你的问题。

[01:13] So first of all, the reason why extreme code design is necessary is because the problem no longer fits inside one computer to be accelerated by one GPU.
  首先，极端代码设计之所以必要，是因为问题不再适合在一台计算机内，由一个GPU来加速。

[01:24] The problem that you're trying to solve is you would like to go faster than the number of computers that you add.
  你试图解决的问题是，你想比你添加的计算机数量运行得更快。

[01:32] So you added, you know, 10,000 computers, but you would like it to go a million times faster.
  所以你添加了，你知道，10000台计算机，但你想让它运行得快一百万倍。

[01:38] 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，你必须采用该算法，你必须分解该算法，你必须重构它，你必须分片处理流水线，你必须分片处理数据，你必须分片处理模型。

[01:52] 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
  现在突然之间，当你以这种方式分发问题时，不仅仅是扩展问题，而是你在分发问题，那么一切都会进入

[02:02] This is the AMD doll'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AMD的阿姆达尔定律问题，即你对某事物的加速量取决于它占总工作量的多少。

[02:13] 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地提高了，比如提高了数百万倍，你知道，我只将总工作量提高了一倍。

[02:25] Now all of a sudden, not only do you have to distribute the computation, you have to, you know, shard the pipeline somehow.
  现在突然之间，你不仅要分配计算，还要，你知道，以某种方式分片管道。

[02:35] Uh you also have to solve the networking problem because you've got all of these computers are all connected together.
  呃，你还必须解决网络问题，因为你所有的这些计算机都连接在一起。

[02:42] And so distributed computing at the scale that we do, the CPU is a problem, the GPU is a problem, the networking is a problem, the switching is a problem, and distributing the workload across all these computers are a problem.
  因此，在我们进行的规模上进行分布式计算，CPU是一个问题，GPU是一个问题，网络是一个问题，交换是一个问题，以及将工作负载分配到所有这些计算机上也是一个问题。

[02:57] It's just a massively complex computer science problem and so we just got to bring every technology to bear otherwise
  这只是一个极其复杂的计算机科学问题，所以我们必须运用所有技术来解决，否则

[03:05] we scale up linearly
  我们线性扩展

[03:08] or we scale up based on uh the capabilities of Moore's law which has largely slowed because Dernard's scaling has slowed.
  或者我们根据摩尔定律的能力进行扩展，而摩尔定律的扩展速度已经大大放缓，因为 Dernard 的扩展速度也放缓了。

[03:15] I'm sure there's trade-offs there.
  我相信那里会有权衡。

[03:17] Plus you have a completely disperate disciplines here.
  此外，这里还有完全不同的学科。

[03:19] I'm sure you have specialists in each one of these high bandwidth memory, the the networking, the NVL link, the nyx, the the optics and the copper that you're doing, the power delivery, the cooling, all that.
  我相信在这些高带宽内存、网络、NVL 链路、nyx、光学和铜线、供电、散热等方面，你都有专家。

[03:30] I mean, there's like world experts in each of those.
  我的意思是，在这些领域都有世界顶尖的专家。

[03:31] How do you get them in a room together to figure out that's why my staff is so large.
  你如何让他们聚在一起弄清楚，这就是为什么我的员工如此之多的原因。

[03:37] What's the pro can you take me through the process of the specialists and the generalists?
  专家和通才的过程，你能带我走一遍吗？

[03:39] Like, how do you put together the rack when you know this the set of things you have to shove into a rack together?
  比如，当你需要将一组东西塞进一个机架时，你是如何组装机架的？

[03:48] Yeah, like what does that process look like of designing it all together?
  是的，就像一起设计所有东西的过程是什么样的？

[03:51] There's the the first question which is what is extreme code design?
  第一个问题是，什么是极端代码设计？

[03:53] You're we're optimizing across the entire stack of software from architectures to chips to systems to system software to the algorithms to the applications.
  我们正在优化整个软件堆栈，从架构到芯片到系统到系统软件再到算法再到应用程序。

[04:03] That's one layer.
  这是一个层面。

[04:03] The second thing that you and
  你和

[04:06] I just talked about is goes beyond CPUs and GPUs and networking chips and scale up switches and scale out switches.
  我刚才谈到的是超越了CPU、GPU、网络芯片以及横向扩展和纵向扩展的交换机。

[04:15] And then of course you got to include power and cooling and all of that because you know all these computers are extremely extremely power power hungry.
  然后当然你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。

[04:26] 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，但总的来说它们仍然消耗大量电力。

[04:32] And so that's one the first question is what is it?
  所以这是第一个问题，它是什么？

[04:34] The second question is why is it 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.
  第二个问题是为什么，我们刚才谈到了原因，你知道你想分发工作负载，这样你就可以超越仅仅增加计算机数量的好处。

[04:47] And then the third question is how is it how do you do it?
  然后第三个问题是如何做到，你怎么做？

[04:51] and and uh that's the that's kind of the miracle of this company you know when you're designing a computer you have to have operating system of computers.
  然后，嗯，这就是这家公司的奇迹，你知道，当你设计一台计算机时，你必须有计算机的操作系统。

[05:00] 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，你应该首先考虑你想让公司生产什么。

[05:04] You know, I see a lot of companies organization
  你知道，我看到很多公司组织

[05:08] charts and they all look the same.
  图表，它们看起来都一样。

[05:09] Hamburger organization charts, software organization charts and car company organization charts.
  汉堡组织图、软件组织图和汽车公司组织图。

[05:15] And it doesn't make any sense to me.
  这对我来说毫无意义。

[05:17] You know, the goal of au of a company is to be the machinery, the mechanism, the system that produces the output and that output is the product that we like to create.
  你知道，公司的目标是成为生产产出的机器、机制、系统，而产出是我们喜欢创造的产品。

[05:30] It is also designed the architecture of the company should reflect the environment by which it exists.
  它还被设计成公司的架构应该反映其存在的环境。

[05:34] It almost directly says what you should do with the organization.
  它几乎直接说明了你应该如何处理这个组织。

[05:39] My direct staff is 60 people.
  我的直属员工有60人。

[05:42] You know, I don't have one-on-ones with them because it's impossible.
  你知道，我不能和他们进行一对一的谈话，因为这不可能。

[05:44] You can't have you can't have 60 people on your staff if you're, you know, going to get work done.
  如果你想完成工作，你不能有60名员工。

[05:51] And so you still have 60 reports.
  所以你仍然有60份报告。

[05:53] You still have more.
  你仍然有更多。

[05:56] Yeah.
  是的。

[05:56] And most stars at least have a foot in engineering.
  而且大多数明星至少在工程领域有所涉猎。

[05:59] almost all of them.
  几乎所有人都如此。

[05:59] There's experts in memory, there's experts in CPUs, there's experts in optical all Yeah.
  有内存专家，有CPU专家，有光学专家，是的。

[06:04] GPUs and
  GPU和

[06:08] Architecture, algorithms, design.
  架构、算法、设计。

[06:11] So, you constantly have an eye on the entire stack and you're having to do like intense discussions about the design of the entire stack.
  所以，你必须时刻关注整个技术栈，并就整个技术栈的设计进行深入的讨论。

[06:18] And no conversation is ever one person.
  而且没有任何一次对话是单方面的。

[06:21] That's why I don't do one-on-ones.
  这就是为什么我不进行一对一谈话。

[06:23] We present a problem and all of us attack it, you know, because we're doing extreme code design and literally the company is doing extreme code design all the time.
  我们提出一个问题，然后我们所有人一起解决它，因为我们正在进行极致的代码设计，而且公司实际上一直在进行极致的代码设计。

[06:33] So even if you're talking about a particular component like cooling networking, everybody's listening in.
  所以，即使你在谈论一个特定的组件，比如冷却网络，每个人都在听。

[06:40] Yeah.
  是的。

[06:41] And they can contribute well this doesn't work for the for the power distribution.
  他们可以贡献，嗯，这对于电源分配不起作用。

[06:44] This doesn't exactly this doesn't work for the for the memory.
  这不完全适用于内存。

[06:47] This doesn't work for this.
  这不起作用。

[06:49] Exactly. And whoever wants to tune out, tune out.
  没错。谁想走神就走神吧。

[06:53] You know what I'm saying?
  你知道我说什么吗？

[06:54] And the reason for that is because because the people who are on the staff, they they know when to pay attention.
  原因在于，在场的员工知道何时需要集中注意力。

[06:59] They're supposed, you know, something they could have contributed to, they didn't contribute to.
  他们本应该有所贡献，但却没有。

[07:02] I'm going to call them out, you know, and so, hey, come on, let's get in here.
  我会点名他们，所以，嘿，来吧，加入进来。

[07:07] So, as you mentioned, Nvidia is this
  所以，正如你提到的，英伟达是这个

[07:09] company that's adapting to the environment.
  一家适应环境的公司。

[07:11] So, at which point can you say, did the environment change?
  那么，在哪个时间点你可以说，环境改变了？

[07:16] began 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.
  在早期，从用于游戏的GPU，也许是早期深度学习革命，开始秘密地适应，到现在我们开始将其视为一个AI工厂。

[07:26] What does Nvidia do is produces AI.
  英伟达做什么？它生产人工智能。

[07:28] Let's build a factory that makes AI.
  让我们建造一个制造人工智能的工厂。

[07:31] I could I you could I could reason through it just systematically.
  我可以，你可以，我可以系统地推断出来。

[07:36] Um we started out as as an accelerator company but the problem with accelerators is that the application domain is too narrow.
  嗯，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狭窄。

[07:42] It has the benefit of being incredibly optimized for the job.
  它具有为工作进行极其优化的好处。

[07:45] You know, any specialist has that benefit.
  你知道，任何专家都有这种好处。

[07:49] The problem with intense specialization is that of course your market reach is narrower, but that's that's even fine.
  过度专业化的问题是，当然你的市场覆盖范围会更窄，但这甚至还可以。

[07:57] The problem is the market size also dictates your R&D capacity.
  问题是市场规模也决定了你的研发能力。

[08:04] And your R&D capacity ultimately dictates the influence and
  而你的研发能力最终决定了影响力和

[08:11] impact that you can possibly have in computing.
  你可能在计算方面产生的最大影响。

[08:15] And so when we first started out in acceler as an accelerator, very specific accelerator, we always we always knew that that had that was going to be our first step.
  所以当我们最初作为加速器，一个非常具体的加速器开始时，我们一直都知道那将是我们的第一步。

[08:21] We had to find a way to become accelerated computing.
  我们必须找到一种方法来成为加速计算。

[08:26] But the problem is when you become a computing company, it's too general purpose and it takes away from your specialization.
  但问题是，当你成为一家计算公司时，它的通用性太强，会削弱你的专业性。

[08:32] It turn I connected two words that are actually have fundamental tension.
  它转变为我连接了两个实际上存在根本性张力的词。

[08:38] The better computing company we become, the worse we become as a specialist.
  我们成为一家更好的计算公司，就越不像一个专家。

[08:43] The more of a specialist, the less capacity we have to do overall computing.
  越是专家，我们进行整体计算的能力就越弱。

[08:47] And so the that and I connected those two words together on purpose that the company has to find that really narrow path step by step by step to expand our aperture of computing but not give up on the most important specialization that we had.
  所以我和那个特意将这两个词联系在一起，公司必须一步一步地找到那条非常狭窄的道路，以扩大我们的计算视野，但又不放弃我们最重要的专业领域。

[09:07] Okay.
  好的。

[09:07] So the first step that we took beyond acceleration was we invented the programmable pixel shader.
  所以我们迈出的第一个超越加速的步骤是发明了可编程像素着色器。

[09:11] So that was
  所以那是

[09:13] The first step towards programmability.
  可编程性的第一步。

[09:16] Our, you know, that was our first journey towards moving into the world of computing.
  我们的，你知道，那是我们迈入计算世界的第一段旅程。

[09:20] The second thing that we did was we, we, uh, uh, created, uh, we put FP32 into our shaders.
  我们做的第二件事是，我们，我们，呃，呃，创建了，呃，我们将FP32放入了我们的着色器中。

[09:26] That FP32 step, I.E., compatible FP32 was a huge step in the direction of computing.
  那个FP32步骤，即兼容FP32，是迈向计算领域的一大步。

[09:30] It was the reason why, um, all of the people who were working on, on, um, stream processors and, you know, other types of data flow processors discovered us and they say, "Hey, all of a sudden, you know, we might be able to use these GPUs that's incredibly computationally intensive and it's now, you know, compliant with it.
  这就是为什么，嗯，所有从事流处理器和，你知道，其他类型数据流处理器工作的人都发现了我们，他们说：“嘿，突然之间，你知道，我们也许能够使用这些计算密集型的GPU了，而且它现在，你知道，与之兼容了。

[09:34] E. I can take my software that I was writing, you know, previously on CPUs and I could, you know, see about, you know, using the GPU for that.
  我可以用我以前在CPU上编写的软件，你知道，我就可以，你知道，考虑，你知道，用GPU来做这件事。

[10:04] And which led us to create, put C on top of FP32 was called, we call CG.
  这使我们能够创建，将C放在FP32之上，我们称之为CG。

[10:08] That CG path took us to eventually CUDA, CUDA.
  那条CG路径最终将我们带到了CUDA，CUDA。

[10:15] Step by step by step, um, uh, we, well, putting CUDA on GeForce, that, that was a strategic decision 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, but we did it anyways because we wanted to be a computing company. A computing company has a computing architecture. A computing architecture has to be compatible across all of the chips that we build.
  一步一步地，嗯，呃，我们，嗯，将CUDA放在GeForce上，那，那是一个战略决策，那非常非常难做，因为它花费了公司巨大的利润，我们当时负担不起，但我们还是做了，因为我们想成为一家计算公司。计算公司拥有计算架构。计算架构必须与我们构建的所有芯片兼容。

[10:42] Can you, can you take me through that decision?
  你能，你能带我回顾一下那个决定吗？

[10:43] So, putting CUDA on GeForce, could not afford to do.
  所以，将CUDA放在GeForce上，负担不起。

[10:45] Can you explain that decision?
  你能解释一下那个决定吗？

[10:47] Why, why boldly choose to do that anyway?
  为什么，为什么大胆地选择这样做？

[10:53] Excellent. That was, that was the first, I would, I would say that that was the first, um, the first strategic decision that, that is as close to an existential threat. 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.
  太好了。那是，那是第一个，我会，我会说那是第一个，嗯，第一个战略决策，那几乎是生死攸关的威胁。对于不知道的人来说，剧透一下，事实证明这是公司有史以来最 the most incredibly brilliant 决定之一。

[11:12] So CUDA turned out to be
  所以CUDA最终成为了

[11:18] an incredible foundation for computation in this AI infrastructure world.
  这是人工智能基础设施领域计算的令人难以置信的基础。

[11:23] So, so you're just setting the context.
  所以，你只是在设定背景。

[11:25] It turned out to be a good decision.
  事实证明这是一个明智的决定。

[11:27] Yeah, it turned out to have been a good decision.
  是的，事实证明这是一个明智的决定。

[11:31] I think the So, so here here's the way it went.
  我认为，所以，事情是这样的。

[11:32] So, we invented this thing called CUDA and um uh it expanded the the aperture of applications that that we can accelerate with our accelerator.
  所以，我们发明了名为CUDA的东西，它扩展了我们可以用我们的加速器加速的应用程序的范围。

[11:40] The question is how do we how do we attract developers to CUDA?
  问题是我们如何吸引开发者使用CUDA？

[11:48] Because a computing platform is all about developers.
  因为计算平台就是关于开发者的。

[11:50] And developers don't come to a computing platform just because you know it could perform something interesting.
  开发者不会仅仅因为你知道它可以执行一些有趣的事情就来到一个计算平台。

[11:54] They come to a computing platform because the install base is large.
  他们来到一个计算平台是因为安装基数很大。

[12:02] Because a developer like anybody else wants to develop software that reaches a lot of people.
  因为开发者和任何人都一样，想开发能够触达很多人的软件。

[12:09] 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。

[12:14] The architecture could attract enormous amounts of criticism.
  该架构可能会招致大量的批评。

[12:17] For example, no
  例如，不

[12:20] architecture has ever attracted more criticism than the x86.
  有史以来，没有哪种架构比 x86 受到更多的批评。

[12:25] You know, as as a less than less than elegant architecture, but yet it is the defining architecture of today.
  你知道，它是一种不够优雅的架构，但它却是当今的决定性架构。

[12:33] It 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 and so I've given you two examples where one is you know one is elegant the other one's barely aesthetic and so yet x86 survived.
  它给你一个例子，事实上，许多风险架构，尽管由世界上最聪明的计算机科学家精心设计，但却很大程度上失败了，所以我给你举了两个例子，一个你知道，一个是优雅的，另一个几乎没有美感，但 x86 却得以幸存。

[12:58] install base is everything.
  装机量就是一切。

[12:59] install base defines an architecture not everything else is secondary.
  装机量定义了架构，其他一切都是次要的。

[13:04] Okay. And so there were other architectures at the time.
  好的。所以当时还有其他架构。

[13:07] CUDA came out, Open CL was here.
  CUDA 出来了，Open CL 也在。

[13:09] There were you know there's several other competing architectures but the the thing that the decision that we made that was good was we said hey look ultimately it's about um installed base.
  你知道，还有其他几个竞争性架构，但我们做出的一个好的决定是，我们说，嘿，看，最终是关于装机量。

[13:20] And what is the best way we could get a new computing architecture into the world.
  那么，我们如何才能将一种新的计算架构推向世界呢？

[13:26] By that time frame GeForce had become successful.
  到那个时候，GeForce已经取得了成功。

[13:29] We were already selling millions and millions of GeForce GPUs a year.
  我们每年已经售出了数百万，数百万块GeForce GPU。

[13:33] And we said, you know, we we ought to put CUDA on GeForce and put it into every single PC whether customers use it or not and use it as a starting point of cultivating our installed base.
  我们说，你知道，我们应该将CUDA放到GeForce上，并将其放入每一台PC中，无论客户是否使用它，并将其作为培养我们安装基础的起点。

[13:46] Meanwhile, we'll go and attract developers and we went to universities and wrote books and taught classes and put CUDA everywhere.
  与此同时，我们将去吸引开发者，我们去了大学，写书，教课，并将CUDA推广到各处。

[13:59] And eventually people 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。

[14:02] There was no cloud and we could put a supercomputer in the hands of every researcher in school, every scientist, you know, every engineering school, every or every student in school and eventually something amazing will happen.
  那时还没有云计算，我们可以将超级计算机交给学校里的每一位研究员，每一位科学家，你知道，每一个工程学院，每一个或学校里的每一个学生，最终会发生一些令人惊叹的事情。

[14:15] Well, the problem was CUDA increased our cost of that GPU, which is
  嗯，问题是CUDA增加了我们GPU的成本，而这

[14:21] A consumer product, so tremendously it completely consumed all of the company's gross profit dollars.
  一款消费级产品，它极大地消耗了公司所有的毛利润。

[14:29] And so, at the time, the company was probably, you know, worth, I don't know, at the time, eight, was it like $8 billion or something like $67 billion or something like that.
  所以，当时，这家公司大概，你知道，价值，我不知道，当时，八，是像80亿美元还是像670亿美元还是什么的。

[14:40] 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。

[14:48] You know, our market cap went down to like $1.5 billion.
  你知道，我们的市值下降到大约15亿美元。

[14:51] And so, we were down we were down there for a while and and uh we clawed our way way back slowly, but we carried CUDA on GeForce.
  所以，我们低迷了一段时间，然后我们慢慢地努力恢复，但我们继续在GeForce上运行CUDA。

[14:59] I always say that Nvidia is the house that GeForce built because it was GeForce that took CUDA out to everybody.
  我一直说英伟达是GeForce建造的帝国，因为是GeForce将CUDA带给了所有人。

[15:10] Researchers, scientists, um they discovered CUDA on GeForce because they were all, you know, many of them were gamers.
  研究人员，科学家，嗯，他们在GeForce上发现了CUDA，因为他们都是，你知道，很多人都是游戏玩家。

[15:17] Um many of them built their own
  嗯，很多人自己构建了

[15:21] PCs anyways in a university lab.
  总之，在大学实验室里的个人电脑。

[15:25] many of them built clusters themselves you know.
  他们中的许多人自己构建了集群，你知道的。

[15:27] using using PC components and and so.
  使用，使用PC组件，等等。

[15:29] that you know that's kind of how we got going.
  你知道，这有点像是我们开始的方式。

[15:31] and then that became the platform the foundation for the deep learning revolution.
  然后它就成了深度学习革命的平台和基础。

[15:35] that was also another great great observation.
  这也是另一个非常非常棒的观察。

[15:37] yeah.
  是的。

[15:38] that existential moment do you remember.
  那个决定生死的时刻，你还记得吗？

[15:40] like what were those meetings like what were those discussions like deciding as a company risking everything.
  那些会议是怎样的？那些讨论是怎样的？作为一家公司，冒着一切风险做决定。

[15:48] well um I had I had to make it clear to the board what we were trying to do and and um uh the management team knew our gross margins were going to get crushed.
  嗯，我必须向董事会清楚地说明我们想做什么，管理团队知道我们的毛利率将会被压垮。

[16:00] So you could imagine a world where GeForce would carry the burden of CUDA and none of the gamers would appreciate it and none of the gamers would pay for it.
  所以你可以想象一个世界，GeForce将承担CUDA的负担，而没有玩家会欣赏它，也没有玩家会为此付费。

[16:10] 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。

[16:14] And so that you know, we we increased our cost by 50% and that consumed and we were a 35% gross margin.
  所以你知道，我们成本增加了50%，这消耗了，而我们是35%的毛利率。

[16:22] company.
  公司。

[16:25] And so it it was a it was quite a difficult decision to make, but you could imagine that someday this could go into workstations and it would go into supercomputers and and in those segments maybe we can capture more margin.
  所以这是一个非常艰难的决定，但你可以想象有一天它会进入工作站，进入超级计算机，在那些领域我们或许能获得更高的利润。

[16:38] Um so you you could you could reason your way into being able to afford this.
  嗯，所以你可以通过推理来证明你有能力负担得起。

[16:43] Uh but it still took it took a decade.
  呃，但它仍然花了十年时间。

[16:45] But that but that's more like conversation with the board convincing them.
  但这更像是与董事会进行说服他们的谈话。

[16:48] But you psychologically
  但你在心理上

[16:50] because Nvidia has continued to make bold bets that predict the future and in part especially now define the future.
  因为英伟达一直做出预测未来并尤其现在定义未来的大胆的赌注。

[16:59] So I'm almost looking for wisdom about how you were able to make those decisions to make leaps like that as a company.
  所以我几乎是在寻求关于你们公司是如何做出那些决策，进行那样飞跃的智慧。

[17:14] Well, f first of all, um I'm informed by by by a lot of curiosity.
  嗯，首先，我受到很多好奇心的驱动。

[17:18] Uh at some point there's a reasoning system
  呃，在某个时候，有一个推理系统

[17:25] that that convinces me uh so clearly.
  这让我非常清楚地确信。

[17:29] this outcome will happen.
  这个结果将会发生。

[17:32] that this will happen.
  这将会发生。

[17:35] And so I believe I believe it in my mind.
  所以我相信，我在心里相信它。

[17:37] And when I believe it in my mind, you know, you know how it is.
  当我心里相信它时，你知道，你知道它是怎么回事。

[17:39] You manifest a future.
  你显化一个未来。

[17:42] And that future is so convincing, there's no way it won't happen.
  而那个未来是如此令人信服，它一定会发生。

[17:45] There's a lot of suffering in in between, but you've got to believe what you believe.
  中间有很多痛苦，但你必须相信你所相信的。

[17:51] So you you you envision the future.
  所以你，你，你设想未来。

[17:55] Yeah.
  是的。

[17:56] And you essentially from a sort of engineering perspective manifest it.
  你基本上是从一种工程的角度来显化它。

[17:59] Yeah.
  是的。

[17:59] And and you you reason about how to get there.
  然后，然后你思考如何到达那里。

[18:01] You reason about why it it must exist.
  你思考它为什么必须存在。

[18:04] Um and and um and you know, I reason we all reason here.
  嗯，然后，嗯，你知道，我思考，我们都在这里思考。

[18:08] the management team will reason about it.
  管理团队会对此进行思考。

[18:11] All the people that I we spend a lot of time reasoning about it.
  所有我认识的人，我们花了很多时间来思考它。

[18:14] The thing the thing that the next part of it is probably a skill thing which is you know oftentimes in leadership uh the leadership stays quiet or they learn
  接下来的部分可能是一个技能问题，你知道，在领导力方面，领导者常常保持沉默，或者他们学习

[18:25] about something and then they do some manifesto and it's a brand new year and somehow at the end of the year next year we're going to have a brand new plan,
  关于某事，然后他们会制定一份宣言，新的一年开始了，不知何故，到明年年底，我们将有一个全新的计划，

[18:34] big huge layoff this way, big huge organization change this way, new mission statement, brand new logos, um you know that kind of stuff.
  大规模裁员，大规模组织变革，新的使命宣言，全新的标志，嗯，你知道的那种东西。

[18:42] Um, we've just never I never do things that way.
  嗯，我们从来没有，我从来不那样做。

[18:45] When I learn about something and it's starting to influence how I think, I'll make it very clear to everybody near me that, you know, this this is interesting.
  当我了解到某事并开始影响我的思考方式时，我会让周围的每个人都非常清楚，你知道的，这很有趣。

[18:55] Um, this is going to make a difference.
  嗯，这将带来改变。

[18:58] Uh, this is going to impact that.
  呃，这将对那产生影响。

[19:01] And I reason about things step by step by step.
  我一步一步地推理事情。

[19:03] often times I've already made up my mind but I'll take every possible opportunity external information new insights new discoveries uh new engineering you know revelations uh new milestones developed I'll take those opportunities and I'll use it to shape everybody else's belief system and
  很多时候我已经下定决心，但我会抓住每一个可能的机会，外部信息、新见解、新发现、呃，新的工程，你知道的，启示，呃，新开发的里程碑，我会抓住这些机会，并用它们来塑造其他人的信念体系，

[19:25] I'm doing that literally every single day.
  我几乎每天都在这样做。

[19:28] I'm doing that with my board.
  我和我的董事会一起这样做。

[19:30] I'm doing that with my management team.
  我和我的管理团队一起这样做。

[19:32] I'm doing that with my employees.
  我和我的员工一起这样做。

[19:34] I'm trying to shape their belief system such that when I come the day I say, "Hey, let's buy Melanox."
  我试图塑造他们的信念体系，以便当我来的时候，我说：“嘿，我们买下Melanox吧。”

[19:43] It's completely obvious to everybody that we absolutely should.
  对每个人来说，我们绝对应该这样做，这是显而易见的。

[19:48] On the day that on the day that I that I said, "Hey guys, let's go all in on deep learning."
  在我说：“嘿，伙计们，让我们全力投入深度学习”的那一天。

[19:54] And let me tell you why.
  让我告诉你为什么。

[19:56] I've already been laying down the bricks to different organizations inside the company.
  我已经为公司内部的不同组织打下了基础。

[19:59] every organization and every everybody many of the people might have heard everything.
  每个组织，每个人，很多人可能都听说了所有的事情。

[20:06] most of the company heard hears of course pieces of it and on the day that I announce it um
  公司的大部分人都听说了其中的一些内容，当然，在我宣布的那天，嗯

[20:16] everybody's kind of bought into many pieces of it and in a lot of ways I like to announce these things and I imagine um that that the employees are kind of
  每个人都接受了其中的许多部分，在很多方面我都喜欢宣布这些事情，我想象着嗯，员工们都有点

[20:27] saying you know Jensen what took you so

[20:29] long and and in fact I've been shaping

[20:32] their belief system for some time and

[20:34] therefore leadership

[20:36] sometimes it looks like you're leading

[20:38] from behind

[20:39] >> but you've been shaping their you know

[20:41] to the point where on the day that I

[20:42] declared it 100% buy in but that's what

[20:45] you want you want to bring everybody

[20:47] along you know otherwise we announce

[20:49] something about deep learning and

[20:50] everybody goes what are you talking

[20:51] about you know you announce something

[20:54] about let's go allin on this thing and

[20:56] and your your management team your board

[20:59] your employees your customers, they're

[21:01] kind of like, where's this coming from?

[21:02] You know, this is insane. And so, so,

[21:05] uh, GTC, in fact, if you go back in

[21:07] time, you look at look at the keynotes,

[21:11] I'm also shaping the belief system of my

[21:14] partners and the industry and and I'm

[21:17] using that to shape, you know, the

[21:18] belief system of my own employees and

[21:21] and and so by the time that I announce

[21:23] something, like, for example, we just

[21:25] now we just announced Grock, we've been

[21:28] late. I've been talking about the

[21:30] stepping stones for two and a half

[21:32] years. You guys just go back and oh my

[21:35] gosh, they've been talking about it for

[21:37] two and a half years. And so I've been

[21:39] laying the foundation step by step by

[21:40] step. So when the time comes you

[21:42] announce it, everybody's, you know, what

[21:43] took you so long?

[21:44] >> But it's not just inside the company.

[21:45] You're shaping the landscape, the

[21:47] broader global landscape of innovation.

[21:49] Like putting those ideas out there, you

[21:51] really are manifesting reality.

[21:53] >> We don't build computers. We actually

[21:54] don't build clouds. We don't, as it

[21:57] turns out, we're a computing platform

[21:58] company and so nobody can buy anything

[22:01] from us. That's the weird thing. You

[22:03] know, we ver we vertically

[22:06] design vertically integrate to design

[22:08] and optimize, but then we open up the

[22:11] entire platform at every single layer to

[22:14] be integrated into other companies

[22:17] products and services and clouds and

[22:19] supercomputers and OEM computers and and

[22:22] so the amazing thing is I can't do what

[22:25] I do without having convinced them

[22:27] first. And so most of GTC is about

[22:31] manifesting a future that by the time

[22:33] that we my product is ready, they're

[22:36] going what took you so long? Yeah. Uh so

[22:40] one of the things you've been a believer

[22:42] for a long time is uh scaling laws

[22:45] broadly defined. So are you still a

[22:48] believer in the in the scaling laws?

[22:49] >> Yeah, we have more scaling laws now.

[22:51] >> So I think uh you've outlined four of

[22:53] them with pre-training, post- training,

[22:55] test time, and agentic scaling. What do

[22:58] you think when you think about the

[23:00] future, deep future and the near-term

[23:03] future, what are the blockers that

[23:06] you're most concerned about that keep

[23:08] you up at night that you have to

[23:09] overcome in order to keep scaling?

[23:12] >> Well, we can go back and reflect on what

[23:14] people thought were blockers.

[23:16] >> Mhm. So in the beginning we were the

[23:18] first the pre pre-training scaling law

[23:21] you know people thought uh well

[23:23] rightfully so that the amount of data

[23:25] that we have high quality data that we

[23:28] have um will limit the intelligence that

[23:30] we achieve and that scaling law was an

[23:32] important very important scale law the

[23:34] larger the model the correspondently

[23:36] more data uh results in a better with a

[23:39] results in a smarter AI and so that was

[23:42] pre-training and Ilas Susker Ilas

[23:45] we're out of data or something like

[23:47] that. Pre-training is over or something

[23:48] like that. The the industry panicked,

[23:51] you know, that this is the end of AI.

[23:54] And of course, of course, that's that's

[23:56] obviously not true. Um, we're going to

[23:58] keep on scaling the amount of data that

[23:59] we h have to to train with. A lot of

[24:02] that data is probably going to be

[24:03] synthetic. And that also confused

[24:06] people, you know, and and what people

[24:08] don't realize is they've kind of

[24:10] forgotten that most of the data that

[24:12] that we are training uh that we teach

[24:15] each other with, inform each other with

[24:16] this is synthetic. You know, I it's

[24:19] synthetic because it didn't come out of

[24:22] nature. You created it. I'm consuming

[24:25] it. I modify it, augment it, I

[24:30] regenerate it, somebody else consumes

[24:32] it. And so so we've now reached a level

[24:35] where AI is able to

[24:39] take ground truth, augment it,

[24:43] enhance it, synthetically generate an

[24:46] enormous amount of data and that part of

[24:49] post training um continues to scale. And

[24:51] so the amount of data that we could use

[24:53] that is human generated will be smaller

[24:56] and smaller and smaller. the amount of

[24:58] data that we use to uh train model uh uh

[25:02] is going to continue to scale to the

[25:04] point where we're no longer limited

[25:06] training is no longer limited by data is

[25:09] now limited by compute and the reason

[25:11] for that is most of the data is

[25:12] synthetic then the next phase is uh test

[25:16] time and um I I still remember people

[25:20] people telling me that inference oh yeah

[25:22] that's easy pre pre-training that's hard

[25:25] these are giant systems that people are

[25:26] talking about inference must be easy and

[25:29] so inference chips are going to be

[25:30] little tiny chips and you know they're

[25:32] not they're not like Nvidia's chips oh

[25:34] those are going to be complicated and

[25:36] expensive and you know we could make and

[25:38] this is and in the future inference is

[25:41] going to be the biggest market and it's

[25:42] going to be easy and we're going to

[25:43] commoditize and you know everybody can

[25:45] build their own chips and and and that

[25:48] was always illogical to me because

[25:51] inference is thinking and I think

[25:54] thinking is hard thinking is way harder

[25:57] than reading.

[25:59] >> You know, pre-training is just

[26:01] memorization and generalization, you

[26:03] know, and looking for patterns and

[26:05] relationships. You're reading and

[26:07] reading versus thinking, reasoning,

[26:10] solving problems, taking un unexplored

[26:16] experiences, new experiences, and

[26:18] breaking it down into de decomposing it

[26:21] into, you know, solvable pieces that we

[26:24] then go off either through first

[26:26] principal reasoning or, you know,

[26:28] through through uh previous examples,

[26:30] prior experiences, you know, or or or

[26:33] just uh uh exploration. and and search

[26:36] and you know trying different things and

[26:39] that whole process of post of of test

[26:42] time scaling. Uh inference is really

[26:45] about thinking and and it's about

[26:47] reasoning. It's about planning. It's

[26:49] about search. It's about and so how

[26:51] could that possibly be computed? And we

[26:54] were absolutely right about that you

[26:56] know so so test time scaling is

[26:58] intensely comput intensive.

[27:01] Then the question is okay now we're at

[27:02] inference and we're at test time

[27:04] scaling. What's beyond that? Well,

[27:06] obviously

[27:08] uh we have now created you know one

[27:10] agentic person and that one agentic

[27:14] person has a large language model that

[27:15] we've now we've now you know developed.

[27:18] But during test time, that agentic

[27:20] system goes off and does research and

[27:23] bangs on databases and it goes on and

[27:26] you know uses tools and one of the most

[27:28] important things it does is spins off

[27:30] and spawns off a whole bunch of sub

[27:32] aents which means we're now creating

[27:34] large teams. It's so much easier to

[27:38] scale Nvidia by hiring more employees

[27:42] than it is to scale myself.

[27:44] >> And so the next scaling law is the

[27:45] agentic scaling law. It's kind of like

[27:48] multip multiplying

[27:50] AI. Multiplying AI, we could spin off

[27:53] agents as fast as you want to spin off

[27:55] agents. And so, you know, I you have

[27:58] four scaling laws. And and as we use the

[28:01] a agentic systems, they're going to

[28:03] create a lot more data. They're going to

[28:04] create a lot of experiences. Some of it

[28:06] we're going to say, "Wow, this is really

[28:08] good. We ought to memorize this."

[28:12] >> That data set then comes all the way

[28:13] back to pre-training. We memorize and

[28:16] generalize it. We then refine it and

[28:18] fine-tune it back into post training.

[28:22] Then we enhance it even more with test

[28:24] time, you know, in the agent agents

[28:27] agentic systems, you know, put it onto

[28:29] the indust industry. And so this loop,

[28:31] the cycle is going to go on and on and

[28:34] on. It kind of comes down to basically

[28:37] intelligence is going to scale by one

[28:39] thing and it's compute. But there's a

[28:42] tricky thing there that you have to

[28:43] anticipate and predict which is some of

[28:46] these components. It requires different

[28:50] kind of hardware to really do it

[28:52] optimally. So you have to anticipate

[28:55] where the AI innovation is going to

[28:56] lead. For example, mixture of experts

[28:58] with sparity.

[28:59] >> Perfect.

[29:00] >> With hardware, you can't just pivot on a

[29:03] week's notice. You have to anticipate

[29:04] what that's going to look like. That's

[29:07] >> that's so scary and difficult to do,

[29:09] right? For example, uh these AI model

[29:12] architectures are being invented about

[29:14] once every six months.

[29:16] >> Yeah. Right. And uh system architectures

[29:20] and hardware architectures

[29:23] kind of every 3 years. And so you need

[29:27] to anticipate what likely is going to

[29:29] happen, you know, 2 3 years from now.

[29:33] And there's a couple ways that you could

[29:34] do that. First of all, we could do

[29:35] research internally ourselves. And

[29:36] that's one of the reasons why we have

[29:38] basic research. We have applied

[29:39] research. We create our own models. And

[29:41] so we have we have hands-on life

[29:44] experience right here. This is part of

[29:46] the code design that I'm talking about.

[29:48] >> We're also the only AI company in the

[29:49] world that works with literally every AI

[29:51] company in the world. And to the extent

[29:52] that we can um uh we try to get a sense

[29:55] of of what are the challenges that

[29:57] people are experiencing.

[29:58] >> So you're listening to the whispers

[30:00] across the industry, the adabs.

[30:02] >> That's right. You got to listen and and

[30:04] learn from everybody and have a have a

[30:06] and then the the last part is to have an

[30:08] architecture that's that's flexible that

[30:11] can adapt and move with the wind and one

[30:13] of the benefits of of CUDA is that it's

[30:16] you know on the one hand an incredible

[30:19] accelerator on the other hand it's

[30:21] really flexible and so that balance

[30:24] incredible balance between

[30:26] specialization

[30:28] otherwise we can't accelerate the the

[30:30] CPU versus generalization so that we can

[30:33] adapt with changing algorithms. That's

[30:35] really really important. That's the

[30:37] reason why why um CUDA has been so

[30:39] resilient um on the one hand and yet we

[30:43] continue to enhance it. We're at CUDA

[30:44] 13.2 and so we're invol evolving the

[30:48] architecture so fast that we can stay

[30:50] with you know with with the modern al

[30:54] algorithms. Um for example

[30:57] uh when mixture of experts came out uh

[30:59] that's the reason why we had MVLink 72

[31:02] instead of MVLink 8. We could now take

[31:04] an entire 4 trillion 10 trillion

[31:07] parameter model and put it in one

[31:09] computing domain as if it's running on

[31:11] one GPU. Um I people probably didn't

[31:17] notice I said it but if you look at the

[31:20] architecture of the Grace Blackwell

[31:23] racks it was completely focused on doing

[31:26] one thing processing the LLM.

[31:30] All of a sudden one year later you're

[31:32] looking at a Vera Rubin rack. It has

[31:35] storage accelerators. It has this

[31:38] incredible new CPU called Vera. It has

[31:41] Vera Rubin and MVLink72 to run the LLMs.

[31:46] It also has this new additional rack

[31:48] called Gro. And so this entire rack

[31:51] system is completely different than the

[31:55] previous one and it's got all these new

[31:57] components in it. And the reason for

[31:58] that is because the last one was

[32:00] designed to run

[32:02] large language models inference and this

[32:06] one is to run agents and agents bang on

[32:08] tools and Obviously the design of the

[32:12] system

[32:13] had to have been done before claude

[32:17] code, codeex, open claw. So you were

[32:20] anticipating the future essentially and

[32:22] that that comes from what? From the

[32:23] whispers, from the understanding what

[32:25] all the state of the artist is.

[32:26] >> No, it's it's easier than that. Uh you

[32:28] you just reason about it. Uh first of

[32:31] all just reason

[32:34] no matter no matter what happens at some

[32:37] point in order for that large language

[32:40] model to be a digital worker. Let's just

[32:42] let's just use that metaphor. Let's say

[32:45] that we want the LM to be a digital

[32:47] worker. What does it have to do? It has

[32:49] to access ground truth. That's our file

[32:52] system. It has to be able to do

[32:53] research. It doesn't know everything. We

[32:56] don't have and I don't want to wait

[32:57] until this AI becomes, you know,

[32:59] universally smart about everything past,

[33:03] present, and future before I make it

[33:05] useful. And so therefore, I might as

[33:07] well let it go do research. It's

[33:09] obviously if it wants to help me, it's

[33:11] got to use my tools. You know, a lot of

[33:13] people would say, you know, um AI is

[33:16] going to completely destroy software. We

[33:18] don't need software anymore. We don't

[33:19] even need tools anymore. That's

[33:20] ridiculous. Let's let's use the let's

[33:23] use a thought experiment. Uh, and you

[33:25] could just sit there, enjoy a glass of

[33:27] whiskey and and think about all these

[33:30] things and it would become completely

[33:32] obvious like if I were to create

[33:35] the most amazing ro the most amazing

[33:38] agent that we can imagine in the next 10

[33:40] years, let's say be a humanoid robot. If

[33:43] that human or robot were to be created,

[33:46] is it more likely that the human or

[33:47] robot comes into my house and uses the

[33:50] tools that I have to do the work that it

[33:53] needs to do? Or does his hand turns into

[33:56] a 10- pound hammer in one instance,

[33:59] turns into a scalpel in another

[34:02] instance, and in order to boil water, it

[34:04] beams, you know, microwaves out of its

[34:07] fingers, you know, or is it more likely

[34:09] just to use the microwave, you know, and

[34:11] the first time it goes up to the

[34:12] microwave. It probably doesn't know how

[34:14] to use it. But that's okay. It's

[34:16] connected to the internet. It reads the

[34:19] manual of this microwave, reads it

[34:23] instantly, becomes an expert, and so

[34:25] uses it.

[34:26] >> And so I I think the I just described in

[34:29] fact almost all of the

[34:32] properties of Open Claw.

[34:34] >> Mhm.

[34:35] >> You know, that it's going to use tools,

[34:36] that it's going to access files, it's

[34:38] going to be able to do research, it has

[34:40] IO subsystem. And when you're done

[34:42] reasoning through it, reasoning about it

[34:44] through through it in that way, um then

[34:47] you say, "Oh my gosh, the impact to the

[34:51] future computing is deeply profound."

[34:53] And the reason for that is I think we've

[34:55] just reinvented the computer. And then

[34:58] now you say, "Okay, when did we reason

[35:00] about that? When did we reason about

[35:02] Open Claw?" If you take the Open Claw

[35:05] schematic that I used at GTC,

[35:08] you will find it two years ago.

[35:11] Literally two years ago at GTC, I was

[35:14] talking about Asgentic systems that

[35:18] exactly reflect open claw today and and

[35:22] of course the confluence of of many

[35:25] things had to happen. First of all, we

[35:27] needed claude and and GPT and you know

[35:30] all of these models to reach a level of

[35:33] capability. So so their innovation and

[35:35] their breakthroughs and their continual

[35:36] advances was really important. And then

[35:39] of course somebody had to create a an

[35:41] open- source you know um project that

[35:45] that uh was sufficiently robust you know

[35:48] and sufficiently complete and that we

[35:51] can all we can all put to put to work

[35:53] and and I think openclaw did for did for

[35:56] agentic systems what chat GPT did for

[35:59] generative systems and and I just think

[36:00] it's a very big deal.

[36:02] >> Yeah, it's a really special moment. I'm

[36:04] not exactly sure why it captured

[36:07] so much of the world's attention, but it

[36:08] did more than cloud code and codeex and

[36:11] so on because consumers could reach it.

[36:13] >> Sure. Yeah. But there there's also so

[36:16] much of this is vibes and and Peter uh I

[36:19] had a podcast with him. He's a wonderful

[36:21] human being. So part of it is also the

[36:23] humans that represent the thing. Part of

[36:25] it is memes and the

[36:27] >> cuz we're all trying to figure it out.

[36:28] There's really serious and complicated

[36:30] security concerns about when you have

[36:33] such powerful technology, how do you

[36:34] hand over your data so they can do

[36:36] useful stuff, but then there's scary

[36:38] things associated with that. And we as a

[36:40] civilization, as individual people and

[36:41] as a civilization figuring out how to

[36:43] find that right balance.

[36:44] >> Yeah, we we uh we jumped on it right

[36:46] away and we sent a bunch of security

[36:48] experts this way

[36:49] >> and we did this thing called Open Shell.

[36:51] It's it's already been integrated into

[36:54] into open claw

[36:55] >> and Nvidia put forward Nemo claw.

[36:58] >> Yep. Exactly.

[36:59] >> The install is super easy. It makes sure

[37:02] that uh it's secure.

[37:03] >> We give you two out of three rights.

[37:05] Agentic systems can can access sensitive

[37:07] information. It can execute code and it

[37:10] can communicate externally.

[37:11] >> Mhm.

[37:14] >> We could keep things safe if we gave you

[37:16] two out of those three capabilities at

[37:18] any time, but not all three. And out of

[37:21] those two out of three capabilities, we

[37:23] also give you access control based on

[37:25] based on um whatever rights that you're

[37:27] given by enterprise. And then we

[37:29] connected to a policy engine that all

[37:31] these enterprises already have. And so

[37:34] um we're going to try to do our best to

[37:36] to uh help Open Claw become a a better

[37:39] claw. So you eloquently explained how we

[37:43] have a long history of blockers that we

[37:45] thought were going to be blockers and we

[37:46] overcame them. But now looking into the

[37:48] future, what do you think might be the

[37:49] blockers now that it's clear that agents

[37:52] will be everywhere? So it's obviously

[37:54] we're going to need compute. So what is

[37:56] going to be the blocker for that

[37:58] scaling? Power is a concern, but it's

[38:01] not the only concern. But that's the

[38:03] reason why we're pushing so hard on

[38:05] extreme code design so that we can

[38:09] improve the tokens per second per watt

[38:13] orders of magnitude every single year.

[38:17] And so in the last 10 years, Moors law

[38:19] would have progressed computing about a

[38:22] 100 times in the last 10 years. We

[38:24] progressed and scaled up computing by a

[38:27] million times in the last 10 years. And

[38:29] so we're going to keep on we're going to

[38:30] keep on doing that through extreme code

[38:31] design. Um so energy efficiency per per

[38:35] watt completely affects the revenues of

[38:38] a company. It affects the revenues of a

[38:41] factory and we're just we're just going

[38:44] to push that to the limit so that we can

[38:46] keep on driving token cost down as fast

[38:49] as we can. you know, the our computer

[38:52] price is going up, but our token

[38:55] generation effectiveness is going up so

[38:57] much faster that token cost is coming

[39:00] down. It's just it it's coming down an

[39:02] order of magnitude every year.

[39:04] >> So power that's an interesting one. So

[39:06] the the way to try to get around the

[39:09] power blocker is to try to with the

[39:11] tokens per second per watt try to make

[39:12] it more and more efficient. Of course,

[39:14] there's the question, how do we get more

[39:15] power?

[39:16] >> We should also get more power.

[39:17] >> That's a really complicated one. And

[39:18] you've talked about small module nuclear

[39:20] power plants. There's all kinds of ideas

[39:22] for energy. Uh how much does it keep you

[39:25] up at night? Uh the the bottlenecks in

[39:27] the supply chain of AI like ASML with

[39:30] EUV lithography machines, TSMC with

[39:33] advanced packaging like cos and uh SK HX

[39:36] with high bandwidth memory all all the

[39:39] time and we're working on all the time.

[39:41] No company in history has ever grown at

[39:45] a scale that we're growing while

[39:47] accelerating that growth. It's

[39:49] incredible.

[39:50] >> Yeah.

[39:50] >> And it's hard for people to even

[39:51] understand this in the overall world of

[39:55] AI computing. We're increasing share.

[39:58] And so supply chain upstream and

[40:00] downstream are really important to us. I

[40:04] spent a lot of time um informing all the

[40:08] CEOs that I work with what are the

[40:10] dynamics that's going to cause uh the

[40:13] growth to continue or even accelerate.

[40:15] It's part of the reasons why to the

[40:17] entire right hand side of me were CEOs

[40:21] of practically the entire IT industry

[40:24] upstream and practically the entire

[40:29] infrastructure industry downstream. Mhm.

[40:32] And they were all there were several

[40:34] hundred CEOs and I don't think there's

[40:36] ever been keynotes where several hundred

[40:38] CEOs show up. And and part of it is I'm

[40:42] telling them about our business

[40:44] condition now. I'm telling them about

[40:47] the growth drivers in the very near

[40:49] future and what's happening. And I'm

[40:51] also describing where are we going to go

[40:52] next so that they could use all of this

[40:55] information and all of the dynamics that

[40:57] are here to inform how they want to

[41:00] invest.

[41:01] And so so I I inform them that way like

[41:04] I inform my own employees. And then of

[41:06] course then I make trips out to them and

[41:09] make sure that hey listen I want you to

[41:11] know this quarter, this coming year,

[41:13] this next year these things are going to

[41:16] happen and and if you look at the CEOs

[41:19] of the DRAM industry um the number one

[41:22] DRAM in the in the world was DDR memory

[41:26] for CPUs in data centers. About three

[41:31] years ago, I was able to convince

[41:33] several of the CEOs that even though at

[41:36] the time HBM memory was used quite

[41:39] scarcely, you know, and and barely by

[41:41] supercomputers,

[41:42] um that this was going to be a

[41:44] mainstream memory for data centers in

[41:46] the future. And at first it sounded

[41:48] ridiculous, but several of the CEOs

[41:50] believed me and decided to invest in

[41:53] building HBM memories. Another memory

[41:56] was rather odd to put into a data center

[41:59] is the low power memories that we use

[42:01] for cell phones. And we wanted them to

[42:04] adapt them for supercomputers in the

[42:07] data center. And they go, cell phone

[42:09] memory for supercomputers. And I

[42:11] explained to them why. Well, look at

[42:13] these two memories, LPDDR5,

[42:16] HBM4.

[42:17] The volumes are so incredible. All three

[42:20] of them had record years in history. And

[42:22] these are these are 45 year old

[42:24] companies. And so, you know, I that's

[42:28] part of my job is to

[42:30] inform and shape,

[42:33] inspire,

[42:35] you know. So, you're not just

[42:36] manifesting the the future and maybe

[42:39] inspiring Nvidia, the the the different

[42:42] engineers of the company. You're you're

[42:44] manifesting the supply chain of the

[42:46] future. So you're having conversations

[42:48] with TSMC, with ASML,

[42:50] >> upstream, downstream,

[42:51] >> upstream, downstream. So that's the

[42:53] thing.

[42:53] >> GEV, Caterpillar.

[42:56] >> Yeah, that's downstream from us. Yeah.

[42:58] Yeah. There you go.

[42:59] >> Yeah. The whole thing. I mean, but

[43:00] that's so

[43:02] >> there's so much incredibly difficult

[43:04] engineering that happens in the the

[43:07] entire semiconductor industry. And it's

[43:09] just feels scary how intricate the

[43:14] supply chain is, how many components

[43:16] there are, but it works somehow.

[43:18] Exactly. The deep science, the deep

[43:21] engineering, the incredible

[43:22] manufacturing, and so much of the

[43:24] manufacturing is already robotics, but

[43:26] we have a couple of hundred suppliers

[43:28] that contribute the technology that goes

[43:31] into our 1.3 million component rack.

[43:35] Mhm.

[43:36] >> Each rack is 1.3 one and a half million

[43:40] components. There are 200 suppliers

[43:43] across the Vera Rubin rack.

[43:45] >> So, it's interesting that you don't list

[43:46] that as the thing that keeps you up at

[43:47] night in the list of blockers.

[43:49] >> But I'm doing I'm doing all the things

[43:51] necessary to

[43:52] >> Okay.

[43:53] >> See, I can go to sleep because I checked

[43:55] it off. I said, "Okay, you know, I I go

[43:57] I I can go to sleep and I go, well,

[44:00] let's see what um re let's reason about

[44:02] this. What's important for us?" Um

[44:04] because okay let's reason about this uh

[44:07] because we changed the system

[44:09] architecture from the original DGX1 that

[44:12] you remembered to uh MVLink 72 rack

[44:15] scale computing.

[44:15] >> Mhm.

[44:16] >> What's going to what does that what does

[44:17] that mean? What does that mean to uh

[44:20] software? What does that mean to

[44:22] engineering? What does that mean uh to

[44:24] how we design and test and what does

[44:26] that mean to the supply chain? Well, one

[44:28] of the things that it meant was we moved

[44:32] um supercomput superco computer

[44:34] integration at the data center into

[44:37] supercomputer manufacturing in the

[44:40] supply chain.

[44:41] >> Mhm.

[44:42] If you're doing that, you also have to

[44:45] recognize you're going to move one and

[44:46] and if if if you're if you're, you know,

[44:50] total footprint of whatever data center

[44:53] you're going to build, let's say you

[44:55] would like to have, you know, 50 gawatts

[44:58] of supercomputers that are running

[45:01] simultaneously

[45:02] and it takes one week to manufacture

[45:05] that 50 gawatts of supercomputers. Then

[45:09] each week in the supply chain, the

[45:11] supercomputers are going to need a

[45:12] gigawatt of power. And so so we're going

[45:15] to need the supply chain to increase the

[45:17] amount of power it has to build test to

[45:20] build and test the supercomputers in the

[45:23] supply chain before I ship it.

[45:24] >> Well, MVLink72 literally builds

[45:26] supercomputers in the supply chain and

[45:28] ships them two, three tons at a time per

[45:31] rack. It used to be come they used to

[45:34] come in parts and we used to assemble

[45:36] them inside the data center. But that's

[45:38] impossible now because MVLink 72 is so

[45:40] dense. And so that's an example. And I

[45:42] would have to go into, you know, I fly

[45:45] into the supply chain, go meet my

[45:47] partners, and hey, I said, guess what?

[45:49] So here's what we're going to do with

[45:51] this is the way we used to build our

[45:53] DGXs. We're going to build them this

[45:55] way. This is going to be so much better

[45:56] because we're going to need them for

[45:58] inference. The market for inference is,

[46:00] you know,

[46:02] coming. The inflection point for

[46:03] inference is coming. It's going to be a

[46:04] big market. And so I first explain to

[46:06] them what's going on, why it's going to

[46:08] happen, and then I then I ask them to

[46:12] make several billion dollars of capital

[46:15] investments each

[46:17] and because they, you know, they trust

[46:19] me and and I I I'm very respectful of

[46:22] them and I I give them every opportunity

[46:24] to question me and I spend time to

[46:26] explain things to people and I reason

[46:28] about it. I draw them pictures and I

[46:30] reason about it in first principles and

[46:32] by by the time I'm done with them

[46:34] there's no what to do.

[46:35] >> So it's a lot of is about relationships

[46:37] and building a shared view of the

[46:39] future.

[46:40] >> Yeah.

[46:41] >> Uh but do you worry about certain

[46:44] bottlenecks? I mean what are the biggest

[46:45] bottlenecks in the supply chain? Are are

[46:47] you worried about it ASML V tooling? Are

[46:49] you are you worried about the the

[46:51] packaging co-as packaging of TSMC about

[46:54] how fast it could scale? like you said,

[46:56] you're not only growing incredibly fast,

[46:59] you're accelerating a growth. So it it

[47:01] it feels like every everybody in the

[47:03] supply chain and those are certainly

[47:05] bottlenecks would have to scale up.

[47:07] >> Are you having conversations with them

[47:09] like how can you scale up faster?

[47:12] >> Do you worry about it?

[47:13] >> No.

[47:13] >> Okay.

[47:14] >> Because because I told them what I

[47:16] needed, they understood what I need.

[47:18] They told me what they're going to go do

[47:20] and I believe in what they're going to

[47:21] do.

[47:22] >> Interesting. That's great to hear. So

[47:24] maybe if we can just linger on the power

[47:26] for a little bit. Uh what are your hopes

[47:28] for how to solve the energy problem? One

[47:30] of the areas le that I'm um that I would

[47:34] love I would love love us to talk about

[47:36] and just get the message out. You know

[47:39] um our our our power grid is designed

[47:44] for the worst case condition with some

[47:47] margin.

[47:49] Well, 99% of the time we're nowhere near

[47:52] the worst case condition because the

[47:53] worst case condition is a few days in

[47:55] the winter, a few days in the summer and

[47:58] extreme weather. Most of the time we're

[48:01] nowhere near the worst case condition

[48:03] and we're probably running around call

[48:05] it 60% of peak. And so 99% of the time

[48:12] our power grid has excess power and

[48:15] they're just sitting idle. But they have

[48:17] to be there sitting idle because just in

[48:19] case when the time comes hospitals have

[48:21] to be powered and you know

[48:22] infrastructure has to be powered and

[48:24] airports have to run and so on so forth.

[48:26] And so the question that I have is

[48:28] whether we could go and um help them

[48:32] understand and create contractual

[48:34] agreements and design computer

[48:36] architecture systems, data centers such

[48:38] that when they need

[48:42] um the maximum power for infrastructure

[48:45] in society that the data centers would

[48:48] get less.

[48:49] >> But that's in a very rare instance

[48:50] anyways. And during that time, we either

[48:52] have our backup generator for that

[48:54] little part of it or we just have our

[48:56] computers shift the workload somewhere

[48:57] else or we have the computers just run

[49:00] slower. You know, we could degrade our

[49:02] performance, reduce our power

[49:05] consumption and provide for, you know,

[49:08] slightly longer latency response, you

[49:10] know, when somebody asks for, you know,

[49:12] asked for an answer. And so I think that

[49:14] that that way of using computers of

[49:17] building data centers instead of

[49:20] expecting 100% uptime

[49:22] and these contracts that are really

[49:24] really quite rigorous it's putting a lot

[49:27] of pressure on the grid to be able to

[49:29] now they're going to have to increase

[49:31] from their maximum. I just want to use

[49:33] their excess. It's just sitting there.

[49:36] Yeah. That's not talked about enough. So

[49:38] what's what's this what's stopping

[49:39] there? Is it regulation? Is it

[49:42] bureaucracy?

[49:43] >> I think it's it's a throughway problem.

[49:45] Uh it starts with the end customer. The

[49:47] end customer puts puts requirements on

[49:51] the data centers that they can never

[49:55] not be available. Okay. So that the end

[49:58] customer expects perfection. Now in

[50:00] order to deliver that perfection, you

[50:02] need a combination of backup generators

[50:05] and your grid power supplier to deliver

[50:08] on perfection. And so everybody's got to

[50:11] have 69s.

[50:13] >> Well, I think first of all, right now,

[50:16] we ought to have everybody understand

[50:17] that when the customer asks for these

[50:19] things, you got somebody, you have

[50:21] somebody in your data center operations

[50:23] team disconnected from the CEO. I bet

[50:25] the CEO doesn't know this. I'm going to

[50:27] talk to all the CEOs. The CEOs are

[50:29] probably not paying any attention to the

[50:32] contracts that are being signed. And so

[50:34] everybody wants to sign the best

[50:36] contract of course and they go down to

[50:38] the cloud service providers and the

[50:40] contract the the two contract

[50:42] negotiators that are you I could just

[50:45] see them now

[50:46] >> you know negotiating these multi-year

[50:47] contracts both sides want you know the

[50:51] best contract as a result

[50:54] the CSPs then have to go down to the

[50:57] utilities and they expect the nine the

[50:59] 69s and so I think I think the first

[51:01] thing is just make sure that that all of

[51:04] the customers, the CEOs of the customers

[51:06] realize what they're asking for. Now,

[51:09] the second thing is we have to build

[51:10] data centers that gracefully degrade.

[51:13] And so, if the power, if the utility of

[51:15] the grid tells us, listen, we're going

[51:17] to have to back you down to about 80%.

[51:19] We're going to say that's no problem at

[51:21] all.

[51:21] >> Mhm.

[51:21] >> We're just going to move our workload

[51:23] around. We're going to make sure that

[51:25] data is never lost, but we can reduce

[51:27] the computing rate and use less energy.

[51:31] the quality of service degrades a little

[51:33] bit for the critical workloads I shift

[51:35] that somewhere else right away so I

[51:38] don't have that problem and so you know

[51:40] whoever whichever data center still has

[51:42] 100% uptime and so how difficult of an

[51:45] engineering problem is that the smart

[51:46] dynamic allocation of power in the data

[51:48] center

[51:49] >> as soon as you could specify you could

[51:50] engineer it beautifully put

[51:54] so long as it obeys the laws of physics

[51:56] on first principles I think we're good

[51:58] >> what was the third thing you were

[51:59] mentioning um so the Second thing is the

[52:02] the data centers

[52:03] >> and the third thing is we need the

[52:05] utilities

[52:06] to also recognize that this is an

[52:09] opportunity

[52:10] >> and and instead of instead of saying

[52:12] look um it's going to take me 5 years to

[52:15] increase my grid capability uh if you if

[52:19] you have if you're willing to take power

[52:21] of this level of guarantee

[52:24] I can make them available for you next

[52:26] month and at this price and so if

[52:30] utilities He's also offered more

[52:33] segments of power delivery promises,

[52:36] then I think everybody will figure out

[52:38] what to do with it. Yeah. But there's

[52:39] just way too much waste in the in the

[52:41] grid right now. We we should go after

[52:43] it.

[52:43] >> Uh you've uh highly lauded Elon and uh

[52:47] Xi's accomplishment in Memphis in

[52:49] building um Colossus Supercomputer

[52:53] probably in record time in just 4

[52:54] months. It's now at 200,000 GPUs and

[52:57] growing very quickly. Is there something

[53:00] that you could speak to the understand

[53:02] about his approach that's instructive to

[53:04] the broadly to all the data center

[53:06] creators that's um that enabled that

[53:09] kind of accomplishment his approach to

[53:11] engineering his approach to the whole

[53:13] management of construction everything

[53:15] first of all Elon is deep in so many

[53:18] different topics um uh yet he's also a

[53:22] really good systems thinker

[53:24] >> and so he's able to think through

[53:26] multiple disciplines and and um uh he

[53:30] obviously

[53:32] uh pushes things questions everything

[53:35] whether number one is it necessary

[53:39] number two does it have to be done this

[53:40] way and number you know does it have

[53:43] does it have to take this long and and

[53:46] so so he he has he has the he has the

[53:49] ability uh to question everything uh to

[53:53] the point where everything is down to

[53:55] its minimal amount that's necess

[53:57] necessary. You can't take anything else

[53:59] out and and yet yet the the uh the the

[54:03] the the necessary um capabilities of the

[54:07] product retains, you know, and so he's

[54:09] he is as minimalist as you could

[54:11] possibly imagine and he does it at a

[54:13] system system scale. Um I I also love

[54:16] the fact that he he is um he is

[54:19] represented he he is he is present at

[54:23] the point of action.

[54:24] >> Mhm. you know, he'll just go there and

[54:27] if there's a problem, he'll just go

[54:29] there and show me the problem. You know,

[54:31] when you do all of this in combination,

[54:34] you overcome a lot of previous this is

[54:37] just the way we do it.

[54:38] >> Um, you know, I'm I'm waiting for them.

[54:42] I, you know, I mean, just everybody has

[54:44] a lot of excuses. And and so and then

[54:47] and then the last thing is when when you

[54:49] act personally with so much urgency, uh

[54:52] it causes everybody else to act with

[54:53] urgency, you know, and and every

[54:56] supplier has a lot of customers going

[54:57] on. Every supplier has a lot of projects

[54:59] going on. And he he make it he made it

[55:02] he makes it his business that he's the

[55:05] top priority of everybody else's, you

[55:06] know, projects. And so he does that by

[55:08] demonstrating it.

[55:09] >> Yeah. I've been in a bunch of those

[55:10] meetings. is it's fun to watch cuz

[55:12] really not enough people ask the

[55:14] question like okay so uh can this be

[55:18] done a lot faster and how why does it

[55:20] have to take this long yeah

[55:22] >> and then that becomes an engineering

[55:24] question often and yes I think when you

[55:26] get the ground truth of actually I

[55:29] remember um one of the times I was

[55:31] hanging out with him he literally is

[55:32] going through the entire process how to

[55:34] plug in cables into a rack and he's was

[55:37] working with engineer on the ground

[55:39] that's doing that task and he's just

[55:41] trying to understand what does that

[55:42] process look like so it can be less

[55:44] errorprone

[55:46] and just building up that intuition from

[55:48] every single task involved in uh putting

[55:50] together the data center. You start to

[55:53] immediately get a sense at the detailed

[55:56] scale and at the broad system scale of

[56:00] where the inefficiencies are and so you

[56:02] can make it more and more and more

[56:03] efficient. Plus, you have the big hammer

[56:05] of being able to say, "Let's do it

[56:07] totally different."

[56:08] >> Yeah.

[56:08] >> And remove all possible blockers.

[56:10] >> That's right.

[56:11] >> Is there parallels in the Nvidia extreme

[56:14] systems code design approach that you

[56:15] see in the way Elon approaches systems

[56:17] engineering?

[56:18] >> Well, first of all, the code design is a

[56:20] ultimate systems engineering problem.

[56:22] And so, we approach we approach the work

[56:24] that we do from that first from that

[56:26] principle. Um the other thing that we do

[56:29] uh and this is this is a a philosophy

[56:32] that a thought

[56:35] a a state of mind I guess a method that

[56:39] I started uh 30 years ago and it's

[56:42] called the speed of light. The speed of

[56:44] light is not just about the speed. Speed

[56:46] of light is my my shorthand for u what's

[56:50] what's the limit of what physics can do.

[56:53] And so every single everything

[56:55] everything that we do is compared

[56:56] against the speed of light. Um memory

[56:58] speed uh math speed uh power cost time

[57:04] effort number of people manufacturing

[57:07] cycle time. And uh when you think about

[57:10] latency versus throughput, uh when you

[57:13] think about cost versus throughput, cost

[57:16] versus capacity, all of these things, uh

[57:21] you test against the speed of light to

[57:23] achieve all of these different

[57:26] constraints separately.

[57:28] And then when you consider it together,

[57:31] you know, you have to make compromises

[57:33] because a system that achieves extremely

[57:35] low latency versus achie a system that

[57:37] achieves very high throughput are

[57:39] architected fundamentally differently.

[57:42] But you want to know what's the speed of

[57:44] light of a system that achieves high

[57:47] throughput? What's the speed of light of

[57:49] a system that achieves low latency? And

[57:52] then when you think about the total

[57:53] system, you can make trade-offs. And so

[57:56] I I force everybody to think about

[57:57] what's this what the f the first

[57:59] principles the limits

[58:01] >> the physical limits

[58:03] um for everything before we you know

[58:06] before we uh do anything and and we test

[58:10] everything against that and so that's a

[58:12] good frame of mind I don't love the

[58:15] other methods which is continuous

[58:17] improvement

[58:19] >> the the problem with continuous

[58:20] improvement it it first of all you

[58:23] should engineer something from first

[58:26] principles at the speed you know with

[58:27] speed of light thinking limited only by

[58:30] physical limits and and physics limits

[58:34] and um after that of course you would

[58:37] improve it over time um but I don't like

[58:41] going into a problem and somebody says

[58:42] hey you know it takes 74 days to do this

[58:45] today

[58:46] >> right now and um we can do it for you in

[58:48] 72 days

[58:49] >> you know I rather strip it all back to

[58:51] zero

[58:52] >> and so first of all explain to me why

[58:54] it's 74 is in the first place and let's

[58:57] know let's think about what's possible

[58:58] today and if I were to to build it

[59:01] completely from scratch you know how

[59:03] long would it take often times you'd be

[59:05] surprised and might come to 6 days now

[59:08] the rest of the 6 days to 74 could be

[59:12] very wellreasoned and compromises and

[59:15] you know cost reductions and all kinds

[59:17] of different things but at least you

[59:19] know what they are and then now that you

[59:21] know that six days possible

[59:24] Then the conversation from 74 to 6

[59:28] surprisingly much more effective

[59:30] >> in such incredibly complex systems that

[59:32] you're working with is simplicity

[59:33] sometimes a good huristic to to reach

[59:36] for I mean if I can just

[59:40] I mean the pod the Vera Rubin pod that

[59:42] you announced is just incredible uh

[59:44] we're talking about seven chips seven

[59:47] chip types five purpose-built rack types

[59:49] 40 racks 1.2 two quadrillion

[59:51] transistors, nearly 20,000 Nvidia dies,

[59:56] over 1100 Ruben GPUs, 60 exoflops, 10

[59:59] pabytes per second of scale bandwidth.

[01:00:01] Uh, that's all just one

[01:00:03] >> that's just one pod.

[01:00:04] >> That's just

[01:00:06] >> Yeah, that's just one pod.

[01:00:07] >> I mean, so you have the and then even

[01:00:09] the the NVL72 rack alone is 1.3 million

[01:00:14] components, 1300 chips, 4,000 lb crammed

[01:00:17] into a single 19inch wide rack. And Lex,

[01:00:20] we'll probably kind of crank out about

[01:00:21] 200 of these pods a week just to put in

[01:00:24] perspective

[01:00:25] >> the the amount of different components.

[01:00:27] I suppose simplicity is impossible, but

[01:00:30] is that a metric that you kind of reach

[01:00:33] for in trying to design things?

[01:00:35] >> You know, the phrase the phrase that I

[01:00:37] use most often is we we need things to

[01:00:40] be as complex as necessary but as simple

[01:00:42] as possible. And and so the question is

[01:00:45] is all that complexity there necessary?

[01:00:48] And we ought to test for that and we

[01:00:50] ought to challenge that. And then after

[01:00:52] that everything else above it, you know,

[01:00:55] it's gratuitous.

[01:00:56] >> But it's some of the most incredible

[01:00:58] semiconductor industry broadly, but what

[01:01:00] Nvidia is doing uh

[01:01:04] some of the greatest engineering in

[01:01:05] history. So these systems are just truly

[01:01:08] truly marvels of engineering.

[01:01:10] >> It is the most complex computer the

[01:01:12] world has ever made. Yeah, the

[01:01:13] engineering teams. I mean, I don't it's

[01:01:15] not a competition, but I don't know if

[01:01:16] if it was like an Olympics of uh

[01:01:18] engineering teams. I mean, TSMC does

[01:01:20] incredible engineering. Like I said,

[01:01:22] ASML at every scale, but Nvidia is going

[01:01:25] to give them a run for their money.

[01:01:27] >> Just incredible, incredible teams,

[01:01:29] >> gold medal medalist in every single in

[01:01:31] every single sport, all assembled right

[01:01:33] here

[01:01:33] >> and have to work together and report

[01:01:35] directly to you. This is wonderful. Uh

[01:01:37] you've recently traveled to China.

[01:01:40] Uh so it's interesting to ask you uh

[01:01:44] China's been incredibly successful in

[01:01:46] building up its technology sector. What

[01:01:48] do you understand about um how China is

[01:01:52] able to over the past 10 years build so

[01:01:54] many incredible world-class companies,

[01:01:57] world-class engineering teams and just

[01:01:59] this technology ecosystem

[01:02:01] >> that produces so many um incredible

[01:02:04] products. whole bunch of reasons for

[01:02:06] well first of all let's let's start

[01:02:07] let's start with some facts 50% of the

[01:02:10] world's AI researchers are Chinese

[01:02:13] plus or minus and they're mostly in

[01:02:17] China still we have many of them here

[01:02:20] but there's amazing researchers still in

[01:02:22] China um they their tech industry showed

[01:02:27] up at precisely the right time at the

[01:02:30] time of the mobile cloud era uh their

[01:02:33] way of contributing was software and So

[01:02:35] this is a country's in incredible

[01:02:37] science and math. Uh really well

[01:02:40] educated kids. Um uh their tech industry

[01:02:44] was created during the era of software.

[01:02:48] They're very comfortable with modern

[01:02:49] software.

[01:02:52] China is not one giant economic country.

[01:02:56] It's got many provinces and cities with

[01:02:59] mayors all competing with each other.

[01:03:01] That's the reason why there's so many EV

[01:03:03] companies. That's the reason why there's

[01:03:05] so many AI companies. That's the reason

[01:03:06] why there's so many every company you

[01:03:08] could imagine. Um they all create some

[01:03:12] of them and and um as a result they have

[01:03:16] insane competition internally and you

[01:03:19] know what remains is an incredible

[01:03:22] company. Um they also have a um social

[01:03:27] culture where where it's family first,

[01:03:31] friends second and company third.

[01:03:34] And so

[01:03:36] um

[01:03:38] the amount of conversation that goes

[01:03:42] back and forth between they're

[01:03:45] essentially open source all the time. So

[01:03:48] the fact that they contribute more to

[01:03:49] open source is so sensible because

[01:03:52] they're probably what are we protecting?

[01:03:53] You know my engineers their brothers are

[01:03:56] in that company their friends are in

[01:03:58] that company and they're all

[01:03:59] schoolmates. you know the schoolmate

[01:04:01] concept it's a you know one schoolmate

[01:04:04] your brother for life and um and so they

[01:04:08] they they share knowledge very very

[01:04:10] quickly and so there's no sense keeping

[01:04:14] technology hidden you might as well put

[01:04:16] it on open source and so the open source

[01:04:18] community then amplifies accelerates the

[01:04:21] the innovation process so you get this

[01:04:24] rapid incredibly great talent rapid

[01:04:27] innovation because of open source and

[01:04:29] just you the the nature of friends and

[01:04:32] and um insane competition among compet

[01:04:36] among the company what emerges is

[01:04:38] incredible stuff and so this is the

[01:04:42] fastest innovating

[01:04:44] country in the world today and this is

[01:04:46] something that has everything that

[01:04:47] everything that I've just said is

[01:04:49] fundamental to just how the kids were

[01:04:51] grown the fact that they have excellent

[01:04:53] education the fact that they parents

[01:04:56] want them to do well in school the fact

[01:04:58] that they their culture that way. These

[01:05:00] are, you know, these are just the thing

[01:05:02] about their country and they showed up

[01:05:04] at a precisely the time when technology

[01:05:06] is going through that exponential.

[01:05:09] >> Plus, culturally, it's pretty cool to be

[01:05:11] an engineer. It connects to all the

[01:05:14] components that you're mentioning.

[01:05:16] >> It's a it's a builder nation.

[01:05:18] >> It's a builder nation.

[01:05:19] >> Yeah, it's a builder nation. Um, our

[01:05:21] country's leaders, incredible, but

[01:05:23] they're mostly lawyers. They're

[01:05:25] country's leaders and because we're

[01:05:26] they're trying to keep us safe. uh rule

[01:05:29] of law, uh governing. Their country was

[01:05:33] built out of poverty and so most of

[01:05:37] their leaders are incredible engineers,

[01:05:40] some of the brightest minds.

[01:05:43] To take a small tangent because you

[01:05:44] mentioned open source, I have to uh go

[01:05:47] to Perplexity here, who you have been a

[01:05:49] a fan of a long time.

[01:05:51] >> I love it. Yeah.

[01:05:52] >> And thank you for releasing open source

[01:05:54] Neatron 3 Super, which you can also use

[01:05:57] inside Perplexity. look stuff up.

[01:05:59] >> Yeah.

[01:05:59] >> Uh which is uh 120 billion parameter

[01:06:02] open weight uh model.

[01:06:05] >> Uh what's your vision

[01:06:08] with open source? So you mentioned China

[01:06:11] with with Deep Seek with Minia with all

[01:06:14] these companies really pushing forward

[01:06:17] the open- source uh AI movement and

[01:06:20] Nvidia is really leading the way in um

[01:06:24] close to state-of-the-art open source

[01:06:26] LMS. What's your vision there?

[01:06:28] >> First,

[01:06:30] if we're going to be a great AI

[01:06:32] computing company, we have to understand

[01:06:33] how AI models are evolving.

[01:06:36] >> One of the things that I love about

[01:06:38] Neotron 3 is it's it's not a just a pure

[01:06:41] transformer model. It's transformer and

[01:06:44] SSM. And uh we were early in uh

[01:06:48] developing the the uh conditional GANs

[01:06:51] which that progressive GANs which led

[01:06:54] step by step to diffusion. And so um the

[01:06:57] fact that we're doing basic research in

[01:07:00] model architecture and in different

[01:07:02] domains gives us visibility into you

[01:07:06] know what kind of computing systems

[01:07:07] would do a good job for future models

[01:07:09] and so it is part of our extreme

[01:07:11] codeesign strategy. Second,

[01:07:14] um I think we we right rightfully

[01:07:18] recognize that on the one hand we want

[01:07:22] worldclass models as products and they

[01:07:25] should be proprietary.

[01:07:27] On the other hand, we also want AI to

[01:07:30] diffuse into every industry and every

[01:07:32] country, every researcher, every

[01:07:35] student.

[01:07:37] And if everything is proprietary, it's

[01:07:40] hard to do research and it's hard to

[01:07:42] innovate on top of around with. And so

[01:07:47] open source is fundamentally necessary

[01:07:50] for many industries to join the AI

[01:07:53] revolution.

[01:07:54] Nvidia has the scale and we have the

[01:07:57] motives to not only skills, scale and

[01:08:02] motivation to build and continue to

[01:08:06] build these AI models for as long as we

[01:08:08] shall live. And so therefore, we ought

[01:08:10] to do that. We can open up, we can

[01:08:12] activate every industry, every

[01:08:15] researcher, you know, every country to

[01:08:18] be able to join the AI revolution.

[01:08:21] There's a third reason which is for that

[01:08:24] to recognizing that AI is not just

[01:08:27] language. These AIs will likely use uh

[01:08:31] tools and models and sub aents that were

[01:08:36] trained on other modalities of

[01:08:38] information. Maybe it's biology or

[01:08:40] chemistry or um you know laws of physics

[01:08:44] or you know fluids and thermodynamics

[01:08:47] and not all of it is in language

[01:08:49] structure. And so somebody has to go

[01:08:51] make sure that weather prediction,

[01:08:55] biology, AI, AI for biology, physical

[01:09:00] AI, all of that stuff stays can be

[01:09:03] pushed to the limits and pushed to the

[01:09:04] frontier. We don't build cars, but we

[01:09:07] want to make sure every car company has

[01:09:08] access to great models. We don't we

[01:09:11] don't discover drugs, but I want to make

[01:09:12] sure that Lily has the world's best

[01:09:15] biology AI systems so that they can go

[01:09:17] use it for discovering drugs. And so

[01:09:20] these three fundamental reasons both in

[01:09:22] in recognizing that AI is not just the

[01:09:25] language that AI is really broad that we

[01:09:27] want to engage everybody into the world

[01:09:29] of AI and then also codees of AI.

[01:09:32] >> Well, I have to say once again, thank

[01:09:34] you uh for open sourcing really truly

[01:09:37] open sourcing uh Neatron 3. And

[01:09:39] >> yeah, I appreciate you were saying that

[01:09:40] we open source the models, we open

[01:09:42] source the weights, we open source the

[01:09:43] data, we open source how we created it.

[01:09:46] >> Yeah, it's pretty amazing.

[01:09:48] It's really It's really incredible.

[01:09:51] You're originally from Taiwan and have a

[01:09:53] close relationship with TSMC. So I have

[01:09:56] to ask uh TSMC I think uh also is a

[01:10:00] legendary company in terms of the

[01:10:02] engineering teams in terms of the

[01:10:03] incredible engineering work that they

[01:10:05] do. uh what uh what do you understand

[01:10:08] about TSMC culture and their approach

[01:10:10] that explains how they're able to

[01:10:13] achieve this singular unmatched success

[01:10:16] in uh everything they're doing with

[01:10:18] semiconductors? You know, first of all,

[01:10:21] the deepest misunderstanding about TSMC

[01:10:25] is that that um

[01:10:29] their technology

[01:10:31] is all they have. that somehow they they

[01:10:35] have a really great transistor and if

[01:10:37] somebody shows up another transistor

[01:10:39] game over

[01:10:41] >> it's the technology and of course you

[01:10:44] know I I don't mean just the trans

[01:10:46] transistor the metalization systems the

[01:10:48] packaging the 3D packaging the silicon

[01:10:50] photonics the you know all of the

[01:10:52] technology that they have that

[01:10:53] technology is really what makes the

[01:10:55] company special their technology makes

[01:10:57] the company special

[01:10:59] but their ability to orchestrate

[01:11:04] the the demands the the dynamic demands

[01:11:08] of hundreds of companies in the world as

[01:11:11] they're moving up, shifting out, you

[01:11:14] know, increasing, decreasing, push

[01:11:16] pushing out, pulling in, um changing

[01:11:20] from customer to customer, uh wafer

[01:11:23] starting, wafer stopping,

[01:11:26] uh emergency wafer starts, you know, all

[01:11:29] of this dynamics of the world's

[01:11:32] complexity as the world is shapeshifting

[01:11:35] all the time and somehow they're running

[01:11:38] a factory with high throughput, high

[01:11:41] yields, really great costs, excellent

[01:11:44] customer service. They they take their

[01:11:47] work ser they take their promises

[01:11:49] seriously. when your wafer because they

[01:11:51] know that you're help they're helping

[01:11:52] you run your company when the wafers

[01:11:54] when the wafers were promised to show up

[01:11:56] the wafers show up you know so that you

[01:11:58] could run your company appropriately and

[01:12:00] so their system their manufacturing

[01:12:02] system is completely miraculous I would

[01:12:05] say then the second thing is their

[01:12:06] culture this culture is uh

[01:12:08] simultaneously

[01:12:10] uh technology focused on one hand

[01:12:12] advancing technology

[01:12:14] simultaneously customer serviceoriented

[01:12:16] on the other hand a lot of

[01:12:19] C companies are very customer

[01:12:20] serviceoriented, but they're not very

[01:12:22] technology

[01:12:24] excellent. They're they're not at the

[01:12:25] bleeding edge of technology or a lot of

[01:12:27] companies who are tech at the bleeding

[01:12:28] edge of technology, but they're not the

[01:12:30] best customer service oriented company.

[01:12:32] And so it just depends on somehow

[01:12:34] they've they've balanced these two and

[01:12:36] they're world class at both. Um and then

[01:12:40] probably the third thing is the

[01:12:41] technology that I most value in them uh

[01:12:44] that they created this you know this

[01:12:46] this uh intangible called trust. I trust

[01:12:50] them to put my company on top of them.

[01:12:54] That's a very big deal. But they trust I

[01:12:56] mean there's a really close relationship

[01:12:57] there that you've established and that

[01:12:59] trust is established based on many years

[01:13:01] of performance. But there's human

[01:13:03] relationships involved there as well.

[01:13:05] three decades. I don't know how many

[01:13:07] tens, hundreds of billions of dollars of

[01:13:09] business we've done through them and we

[01:13:11] don't have a contract.

[01:13:14] That's pretty great. Amazing. Okay.

[01:13:16] There's a story uh that in 2013 the

[01:13:20] founders of TSMC, Morris Jang, offered

[01:13:22] you the chance to become TSMC's chief

[01:13:25] executive.

[01:13:26] Uh and you said you already had a job.

[01:13:28] Is this story true?

[01:13:30] >> Story is true. I didn't I didn't dismiss

[01:13:32] it. Yeah. Um uh but I was I was deeply

[01:13:35] honored and and of course of course um

[01:13:39] uh I knew then as I know now TSMC is one

[01:13:42] of the most consequential companies in

[01:13:43] history.

[01:13:44] >> Yeah. And and Morris is one of the the

[01:13:48] highest regarded executive and and um

[01:13:51] business and personal friend that I've

[01:13:54] that I've had in my life. And um

[01:13:57] uh for him to ask is uh uh um I I was

[01:14:02] humbled and and really honored.

[01:14:05] Um but but the work that I'm doing here

[01:14:07] is really important and I've seen you

[01:14:09] know in my mind anyways in my mind's eye

[01:14:12] what Nvidia was going to be and what the

[01:14:15] impact that we could have and um uh it

[01:14:18] was really important work

[01:14:21] and it's my responsibility you know my

[01:14:23] sole responsibility to make this happen

[01:14:25] and so I I um uh I declined it you know

[01:14:31] not not because it wasn't an incredible

[01:14:33] offer Uh it it's an unbelievable offer.

[01:14:36] Um but but I simply couldn't take it.

[01:14:38] >> I think Nvidia, both Nvidia and TSMC are

[01:14:41] two of the greatest companies in the

[01:14:44] history of human civilization. Running

[01:14:46] either one, I'm sure, is incredibly

[01:14:48] complicated effort and it takes you have

[01:14:50] to truly be allin.

[01:14:52] >> Yeah.

[01:14:52] >> Uh everybody at every scale, not just at

[01:14:54] the CEO level, everybody is really truly

[01:14:56] allin.

[01:14:57] >> Yeah.

[01:14:58] >> To accomplish this kind of complexity.

[01:15:00] >> See, now I can help both companies.

[01:15:02] >> Exactly. Um, so Nvidia is now the most

[01:15:06] valuable company in the world. I have to

[01:15:08] ask, what is the Nvidia's biggest moat

[01:15:12] as the folks in the tech sector say?

[01:15:15] >> Mhm.

[01:15:15] >> The edge you have that protects you from

[01:15:18] the competition.

[01:15:20] Our single

[01:15:22] most important uh property as a company

[01:15:27] is the install base of our computing

[01:15:30] platform. Our single most important

[01:15:33] thing is the invol today is our is the

[01:15:35] installed base of CUDA. Now the reason

[01:15:38] why uh

[01:15:41] 20 20 years ago of course there was no

[01:15:44] installed base but what makes and if

[01:15:47] somebody if somebody came up with with a

[01:15:50] guda or a tuda uh it wouldn't make any

[01:15:52] difference at all. And the reason for

[01:15:55] that is because because it's never been

[01:15:57] just about the technology. The

[01:15:59] technology of course was incredible

[01:16:01] visionary. Um but it's the fact that the

[01:16:04] company was dedicated to it, stuck with

[01:16:07] it, expanded its reach. Um it wasn't

[01:16:11] three people that that made CUDA

[01:16:13] successful. It was 43,000 people that

[01:16:15] made CUDA successful. and the several

[01:16:18] million developers that believed in us

[01:16:21] um that trusted that we were going to

[01:16:23] continue to make CUDA 1 2 3 13 that they

[01:16:27] decided to port and dedicate their

[01:16:29] software on top of it, their mountain of

[01:16:31] software on top of it. And so the

[01:16:33] install base is the number one most

[01:16:36] important advantage. that installed base

[01:16:39] when you amplified with the velocity of

[01:16:42] our execution at the scale that we're

[01:16:44] talking about. No company in history had

[01:16:47] ever built systems of this complexity

[01:16:50] period. And then to build it once a year

[01:16:53] is impossible.

[01:16:56] And and

[01:16:58] that velocity combined with the

[01:17:00] installed base in the developer's mind

[01:17:04] is just going to now take the

[01:17:05] developer's mind. From the developers

[01:17:07] perspective, if I support CUDA

[01:17:11] tomorrow, it will be 10 times better. I

[01:17:13] just have to wait 6 months on average.

[01:17:16] Not only that, if I develop it on CUDA,

[01:17:19] I reach a few hundred million people

[01:17:22] computers. I'm in every cloud. I'm in

[01:17:25] every computer company. I'm in every

[01:17:27] single industry. I'm in every single

[01:17:29] country.

[01:17:31] So if I created an open source package

[01:17:33] and I put it on CUDA first,

[01:17:35] I get these both attributes

[01:17:38] simultaneously.

[01:17:40] And not only that,

[01:17:43] I trust 100%

[01:17:46] that Nvidia is going to keep CUDA around

[01:17:48] and maintain it and improve it and keep

[01:17:51] optimizing the libraries for as long as

[01:17:54] they shall live.

[01:17:56] You could take that to the bank. And

[01:17:58] that last part, trust,

[01:18:00] you put all that stuff together, if I

[01:18:03] were a developer today, I would target

[01:18:05] CUDA first. I would target CUDA most.

[01:18:09] And that's the reason that that I think

[01:18:11] in the final analysis is our first

[01:18:14] that's even our first

[01:18:17] >> core advantage. Our second one is our

[01:18:19] ecosystem.

[01:18:21] >> The fact that we vertically integrated

[01:18:23] this incredibly complex system, but we

[01:18:26] integrated horizontally into every

[01:18:28] single every single company's computers.

[01:18:30] We're in the Google cloud, we're in

[01:18:31] Amazon, we're in Azure.

[01:18:33] >> You know, we're ramping up AWS like

[01:18:35] crazy right now. We're in new companies

[01:18:38] like Corewave and Nscale. We're in

[01:18:41] supercomputers at Lily. We're in

[01:18:43] enterprise computers. We're at the edge

[01:18:45] in radio base stations. You know, I it's

[01:18:48] just crazy. One architecture is in all

[01:18:50] these different systems. We're in cars,

[01:18:52] we're in robots, we're in satellites,

[01:18:54] we're out in space. And so, so the fact

[01:18:56] that you have this one architecture and

[01:18:58] the ecosystem is so broad, it basically

[01:19:00] covers every single industry in the

[01:19:02] world. Well, how does the how does the

[01:19:05] CUDA install base evolve into the future

[01:19:08] with AI factories as a moat? What do you

[01:19:11] what do do you think it's possible that

[01:19:13] Nvidia of the future is all about the AI

[01:19:15] factory? Well, the the unit of computing

[01:19:18] used to be GPU to us, then it became a

[01:19:21] computer. Then it became a cluster. Now

[01:19:24] it's an entire AI factory. when I see a

[01:19:26] computer, when I see what Nvidia builds

[01:19:28] in the old days, I would, you know, I

[01:19:30] visualize the chip

[01:19:32] >> and then and then when I announced a new

[01:19:34] product, you know, new generation, like

[01:19:36] ladies and gentlemen, we're announcing

[01:19:37] ampear today. I pick up the chip.

[01:19:39] >> Yeah.

[01:19:40] >> That was my mental model what I was

[01:19:42] building.

[01:19:43] >> Today, I don't I wouldn't picking up the

[01:19:45] chip is kind of still adorable,

[01:19:47] >> but it's adorable. It It's not It's not

[01:19:51] my mental model of what I'm doing. My

[01:19:53] mental model is this giant gigawatt

[01:19:56] thing that has power generation. It's

[01:19:59] connected to the grid. It's got cooling

[01:20:01] systems and networking of incredible

[01:20:03] monstrosity. You know, 10,000 people are

[01:20:07] in there trying to install it. Hundreds

[01:20:09] of networking engineers in there.

[01:20:11] Thousands of engineers behind it trying

[01:20:13] to power it up.

[01:20:14] >> You know, powering up one of those

[01:20:15] factories, as you know, it's not

[01:20:17] somebody going, "It's on now."

[01:20:20] takes thousands of people to bring it

[01:20:22] up.

[01:20:22] >> So mentally you're actually when you're

[01:20:24] thinking about a single unit of compute,

[01:20:26] you're like literally when you go to bed

[01:20:28] at night, you're thinking now about

[01:20:30] collection of racks. So pods, not

[01:20:32] individual chips,

[01:20:33] >> entire infrastructure. And I'm hoping my

[01:20:35] next click is when I'm thinking about

[01:20:37] building computers, it's, you know,

[01:20:38] planetary scale. That would be the next

[01:20:41] click. What do you think about the space

[01:20:44] angle that Elon has talked about doing

[01:20:46] compute in space uh for solving some of

[01:20:50] the it makes some of the energy issues

[01:20:54] in terms of scaling energy easier

[01:20:56] cooling issues is not easy. Yeah,

[01:20:58] >> cooling well there's a large number of

[01:21:01] engineering complexities involved with

[01:21:02] that.

[01:21:03] >> So what you know Nvidia has also

[01:21:05] announced that

[01:21:07] >> you're already thinking about that.

[01:21:08] >> Yeah, we're already there. Uh, Nvidia

[01:21:11] GPUs are the first GPUs in space and um

[01:21:15] I I didn't realize it was it was so

[01:21:17] interesting to I would have declared it

[01:21:19] maybe we're in space, you know, little

[01:21:23] little astronaut suit on one of our

[01:21:24] GPUs.

[01:21:27] Um but but we've been in space. Uh it's

[01:21:29] the right place to do a lot of imaging.

[01:21:31] >> Mhm.

[01:21:32] >> You know, because those satellites have

[01:21:33] really high resolution imaging systems

[01:21:36] and they're sweeping the Earth, you

[01:21:37] know, continuously now. And um uh you

[01:21:41] want you know centimeter scale you know

[01:21:43] imaging that is done continuously uh for

[01:21:47] the world so that you know you'll

[01:21:49] basically have real time telemetry of

[01:21:50] everything. Uh you don't want to beam

[01:21:54] that back down to earth. It's just you

[01:21:56] know pabytes and pabytes of data. You

[01:21:59] got to just do AI right there at the

[01:22:00] edge. Throw away everything you don't

[01:22:02] need. You've seen before didn't change

[01:22:04] and then just keep the stuff that that

[01:22:06] you need. And so AI ought to be done at

[01:22:08] the edge. Um obviously we have we have

[01:22:11] uh 24/7 solar if we put it at the polars

[01:22:15] and um uh

[01:22:19] but you know there's no conduction, no

[01:22:21] convection and so you know you're pretty

[01:22:24] much just radiation

[01:22:26] and um uh but you know space is big I

[01:22:29] guess. You know we're just going to put

[01:22:30] big giant radiators out there.

[01:22:32] >> How crazy of an idea do you think it is?

[01:22:33] Like is this is this 5 years out, 10

[01:22:35] years out, 20 years out? So, uh, we're

[01:22:39] talking about blockers for AI scaling.

[01:22:41] You know, I'm just so much more

[01:22:42] practical. I I look for where where um I

[01:22:46] next next bucket of opportunities are

[01:22:49] first.

[01:22:51] Meanwhile, I'm cultivating space. And

[01:22:54] so, I send I send engineers uh to go

[01:22:56] work on the problem. We're we're

[01:22:58] starting to we're learning a lot about

[01:22:59] it. Um, how do we deal with radiation?

[01:23:01] How do we deal with degrading

[01:23:03] performance? How do we deal with um uh

[01:23:05] continuous uh testing and addestation of

[01:23:09] of um def defects and and um you know

[01:23:12] how do we deal with redundancy and how

[01:23:14] do we degrade uh gracefully and things

[01:23:16] like that and so we could we could do uh

[01:23:18] what what about software? How do you

[01:23:20] think about software and and redundancy

[01:23:22] and performance out in space? Uh make it

[01:23:25] so that so that the computer never

[01:23:28] breaks. It just gets slower, you know.

[01:23:31] And um I so we could start doing a lot

[01:23:34] of engineer exploration up front, but in

[01:23:36] the meantime, my my favorite answer is

[01:23:39] get eliminate waste.

[01:23:41] >> You know, we've we've got all that idle

[01:23:43] power. I want to evacuate it as fast as

[01:23:45] possible.

[01:23:47] >> Yeah. There Yeah, there's a lot of low

[01:23:49] hanging fruit here on Earth uh that we

[01:23:51] can utilize uh for the AI scaling. Uh

[01:23:54] quick pause, quick 30 second thank you

[01:23:57] to our sponsors. Check them out in the

[01:23:59] description. It really is the best way

[01:24:01] to support this podcast. Go to

[01:24:03] lexfreedman.com/sponsors.

[01:24:06] We got Perplexity for curiositydriven

[01:24:09] knowledge exploration, Shopify for

[01:24:11] selling stuff online, Element for

[01:24:14] Electrolytes, Finn for customer service

[01:24:16] AI agents, and Quo for a phone system

[01:24:20] like calls, texts, contacts for your

[01:24:22] business. Choose wisely, my friends. And

[01:24:25] now back to my conversation with Johnson

[01:24:28] Kuang. Do you think Nvidia may be worth

[01:24:33] 10 trillion at some point? Let's let's

[01:24:36] ask it this way. What does the future of

[01:24:40] the world look like where that where

[01:24:42] that's true?

[01:24:45] I think that Nvidia's growth is is um

[01:24:51] uh extremely likely and in my mind

[01:24:54] inevitable. And let me explain why.

[01:24:57] We're the largest computer company in

[01:24:58] history.

[01:25:00] That alone should beg the question why.

[01:25:04] And the reason for of course uh two

[01:25:06] reasons. First two foundational

[01:25:08] technical reasons. The first reason is

[01:25:11] that computing went from being a

[01:25:13] retrievalbased file retrieval system.

[01:25:16] Almost everything is a file. We we pre

[01:25:19] pre-write something, we pre-record

[01:25:21] something, you know, we we draw

[01:25:23] something, we put it on the web, we put

[01:25:24] in a file, and we we use a recommener

[01:25:27] system, some smart filter to figure out

[01:25:30] what to retrieve for you. And so we were

[01:25:32] a pre-recording, human pre-recording and

[01:25:34] file retrieving system. That's what a

[01:25:37] computer is largely

[01:25:39] to now AI computers are contextually

[01:25:42] aware which means that it has to process

[01:25:45] and generate tokens in real time. So we

[01:25:47] went from a retrievalbased computing

[01:25:49] system to a generativebased computing

[01:25:52] system.

[01:25:53] We're going to need a lot more

[01:25:55] processing in this new world than in the

[01:25:57] old world. We need a lot of storage in

[01:25:59] the old old world. We need a lot of

[01:26:02] computation in this new world. And so so

[01:26:06] that's that's the first part of it. We

[01:26:08] fundamentally changed computing and the

[01:26:10] way how computing is done. The only

[01:26:12] thing that would cause it to go back is

[01:26:15] if this way of computation, this way of

[01:26:18] computing generating information that's

[01:26:20] contextually relevant, situationally

[01:26:22] aware that is grounded on new insight

[01:26:27] before it generates information. this

[01:26:30] computationintensive way of doing

[01:26:32] computing would only go back if it's not

[01:26:35] effective. So if for the last 1015 years

[01:26:38] while working on deep learning if at any

[01:26:41] single moment

[01:26:43] I would have come to the conclusion that

[01:26:45] that you know what this is not going to

[01:26:48] work out I think this is a dead end or

[01:26:50] it's not going to scale it's not going

[01:26:51] to solve this modality it's not going to

[01:26:53] be used in this application then of

[01:26:55] course I would feel very differently

[01:26:56] about it but I think the last five years

[01:26:59] has given me more confidence than the

[01:27:02] last 10 years the previous 10 years the

[01:27:05] second idea

[01:27:06] is computers because it was a storage

[01:27:09] system. It was largely a warehouse.

[01:27:13] We're now building factories.

[01:27:16] Warehouses don't make much money.

[01:27:20] Factories directly correlates with a

[01:27:23] company's revenues.

[01:27:25] And so

[01:27:28] the computer did two things. Not only

[01:27:30] did it change the way it did it, its

[01:27:34] purpose in the world changed. It's no

[01:27:36] longer a computer, it's a factory. It's

[01:27:40] a factory is used for generation of

[01:27:42] revenues.

[01:27:45] We're now seeing not only is this

[01:27:47] factory generating products, commodities

[01:27:50] that people want to consume,

[01:27:53] we're seeing that the commodities are so

[01:27:55] interesting, so valuable so to so many

[01:27:58] different audiences that the tokens are

[01:28:00] starting to segment like iPhones.

[01:28:03] >> Mhm.

[01:28:04] >> You have a free tokens, you have premium

[01:28:07] tokens, and you have several tokens in

[01:28:09] the middle. And so intelligence, as it

[01:28:12] turns out, you know, is a scalable

[01:28:14] product. There's extremely high

[01:28:16] intelligence products, tokens that you

[01:28:18] could that are used for specialized

[01:28:20] things. People be willing to pay, you

[01:28:22] know, the idea that somebody's willing

[01:28:24] to pay $1,000 per million tokens is just

[01:28:29] around the corner. It's not if, it's

[01:28:31] only when. And so so now we're seeing

[01:28:34] that the commodity that this factory

[01:28:37] makes is actually valuable and is

[01:28:39] revenue generating and profit

[01:28:40] generating. How now the question is how

[01:28:43] many of these factories can does the

[01:28:46] world need?

[01:28:48] How much how many tokens does the world

[01:28:49] need?

[01:28:52] And um

[01:28:54] how much is society willing to pay for

[01:28:56] these tokens?

[01:28:58] And

[01:29:00] what would happen to the world's economy

[01:29:02] if the productivity were to improve so

[01:29:05] substantially?

[01:29:07] What would happen? Are we are we going

[01:29:09] to discover new drugs, new products, new

[01:29:11] services? And so when you take these

[01:29:14] things in combination, I am absolutely

[01:29:16] certain

[01:29:18] that the world's GDP is going to

[01:29:20] accelerate in growth. I'm absolutely

[01:29:23] certain the percentage of that GDP that

[01:29:27] will be used for computation

[01:29:30] will be a 100 times more than the past

[01:29:33] because it's no longer a storage unit.

[01:29:35] It's a product generation unit. And so

[01:29:39] when you look at it in that context and

[01:29:42] then you back into what is Nvidia's what

[01:29:45] does Nvidia what does Nvidia do and how

[01:29:48] much of that

[01:29:50] new economics new industry would we have

[01:29:54] to benefit to address I think we're

[01:29:56] going to be a lot lot bigger and then

[01:29:58] the rest of it to me is um you go is it

[01:30:02] possible for Nvidia to be a you know $3

[01:30:05] trillion revenues company in the near

[01:30:07] future

[01:30:08] The answer is of course yes. And the

[01:30:10] reason for that is because it's not

[01:30:12] limited by any physical limits. There's

[01:30:15] nothing that I see that says, you know,

[01:30:18] gosh, um, $3 trillion is not possible.

[01:30:22] And as it turns out, Nvidia supply chain

[01:30:26] is the burden is shared by 200

[01:30:29] companies.

[01:30:31] and the fact that we scale out on the

[01:30:34] backs of with the partnership of this

[01:30:37] ecosystem.

[01:30:38] The question is do we have the energy to

[01:30:40] do so? And surely we will have the

[01:30:44] energy to do so. And so all of these

[01:30:46] things combined

[01:30:49] that number is just a number you know

[01:30:51] and I still remember Nvidia was a Nvidia

[01:30:54] was a the first time we crossed a

[01:30:56] billion dollars.

[01:30:58] I was reminded of of a CEO who told me,

[01:31:00] you know, Jensen, it's theoretically

[01:31:02] impossible for a fabulous semiconductor

[01:31:04] company to exceed a billion dollars. And

[01:31:08] and um I won't bore you with why, but

[01:31:10] but the of course is illogical and

[01:31:13] there's a lot of evidence we're not. And

[01:31:15] then there somebody told me, you know,

[01:31:16] Jensen, you'll never be more than $25

[01:31:19] billion because of some other company.

[01:31:22] Somebody told me that you'll never be,

[01:31:24] you know, because and then so so the the

[01:31:27] those aren't principled first principle

[01:31:31] reason thinking and the simple the

[01:31:34] simple way to think about that is what

[01:31:36] is it that we make and how large is the

[01:31:39] opportunity that we can create. Now

[01:31:42] Nvidia is not in the market share

[01:31:44] business. Almost everything that I just

[01:31:46] talked about don't exist.

[01:31:47] >> Mhm.

[01:31:48] >> That's the part that's hard.

[01:31:51] You know, if Nvidia was a was a was a

[01:31:54] $10 billion company trying to take

[01:31:56] Nvidia's share, then it's easy to to see

[01:31:59] for shareholders that oh yeah, if they

[01:32:02] could just take 10% share, they could be

[01:32:05] this much larger. But it's hard for

[01:32:08] people to imagine how large we could be

[01:32:10] because there's nobody I could take

[01:32:12] share from,

[01:32:13] >> you know, and so so I think that that's

[01:32:16] one of the challenges for the world is

[01:32:18] is um the imagination of the future. But

[01:32:21] I got plenty of time and I'll keep

[01:32:22] reasoning about it and I'll keep talking

[01:32:24] about it and every single GTC will

[01:32:26] become more and more real,

[01:32:27] >> you know, and and and then more and more

[01:32:29] people will talk about one of these

[01:32:30] days, you know, we'll we'll get there.

[01:32:32] But I'm 100% we'll get there.

[01:32:34] >> Yeah. this view of uh you know token

[01:32:37] factories essentially this token per

[01:32:39] second per watt and every token having

[01:32:42] value like it's an actual thing that

[01:32:44] brings value and it brings different

[01:32:46] kinds of value different amounts of

[01:32:48] value to different people but it's value

[01:32:49] that's the actual product is really

[01:32:51] could be loosely thought of as the token

[01:32:53] and so you have a bunch of token factors

[01:32:54] and it's very easy first principles to

[01:32:57] imagine a future given all the potential

[01:32:59] things that AI can solve that you're

[01:33:01] going to need an exponential number more

[01:33:04] of token factories.

[01:33:05] >> Yeah.

[01:33:06] >> And and what's really interesting, the

[01:33:08] reason why I was so excited about it,

[01:33:09] the iPhone of tokens arrived.

[01:33:11] >> What do you call Wait, are you saying

[01:33:12] open clause iPhone?

[01:33:14] >> Yeah,

[01:33:14] >> that's interesting. Uh

[01:33:15] >> agents.

[01:33:16] >> Yeah, agents. True.

[01:33:18] >> Agents in general. The iPhone of tokens

[01:33:20] arrived. Uh it is the fastest growing

[01:33:22] application in history. It went straight

[01:33:24] up.

[01:33:25] >> Yeah,

[01:33:25] >> went straight up.

[01:33:26] >> That says something.

[01:33:27] >> Yep. There's no question OpenClaw is the

[01:33:30] iPhone of tokens. Yeah, there's

[01:33:32] something truly as you know

[01:33:34] something truly special happening from

[01:33:37] about December where people really woke

[01:33:40] up to the power of claw code of codeex

[01:33:42] of open claw. Um, I mean, I've

[01:33:46] embarrassed to admit that on the way

[01:33:48] here in the airport,

[01:33:51] I've

[01:33:52] this first time I've done this in

[01:33:54] public, I was programming quote unquote

[01:33:57] by talking to my laptop and I was

[01:33:59] embarrassed because I was pretending

[01:34:01] like I'm talking to a human colleague.

[01:34:03] Mhm.

[01:34:03] >> Uh I'm not sure how I feel about the

[01:34:05] future where everybody

[01:34:07] >> is walking around talking to their AI,

[01:34:10] but it's such an efficient way to get

[01:34:12] stuff done

[01:34:13] >> and and it's it's more likely that your

[01:34:15] AI is bothering you all the time. And

[01:34:18] the reason for that is because it's

[01:34:20] getting stuff done so fast.

[01:34:21] >> Yeah.

[01:34:22] >> Is reporting back to you. I got that

[01:34:23] done. You know, what do you want me to

[01:34:25] do next? You know, it that's the part

[01:34:27] that I think most people don't realize

[01:34:29] is mo the person who's going to be

[01:34:31] chatting with them, texting them most is

[01:34:34] their is their claws or lobster.

[01:34:37] >> What an incredible future. Uh I read

[01:34:40] that you attribute a lot of your success

[01:34:42] to your ability to work harder than

[01:34:43] anyone and withstand more suffering than

[01:34:45] anyone.

[01:34:46] So, we can list many of the things that

[01:34:49] entails. I mean, dealing with failure,

[01:34:52] the constant engineering problems we've

[01:34:54] talked about, the the human problems,

[01:34:58] uncertainty, responsibility, exhaustion,

[01:35:00] embarrassment, the near-death company

[01:35:02] moments that you've mentioned,

[01:35:05] um, but also the pressure now as the CEO

[01:35:08] of this

[01:35:10] company that economies and nations

[01:35:14] strategize around, uh, plan their, um,

[01:35:19] financial allocations around plan their

[01:35:21] in AI infrastructure around how do you

[01:35:23] deal with this much pressure?

[01:35:26] What gives you strength given

[01:35:29] how many nations and peoples depend on

[01:35:33] you?

[01:35:37] I'm conscious about the fact that

[01:35:41] um Nvidia success is very important to

[01:35:45] United States.

[01:35:47] We generate enormous amounts of tax tax

[01:35:49] revenues. Uh we establish technology

[01:35:52] leadership for our nation. Technology

[01:35:54] leadership is important for national

[01:35:56] security. National security not just in

[01:35:59] one aspect of national security. All

[01:36:01] aspects of national security. When our

[01:36:03] country is more prosperous,

[01:36:05] we could do a better job with domestic

[01:36:07] policies and helping social social

[01:36:09] benefits because we're generating so

[01:36:12] much re-industrialization in the United

[01:36:15] States. We're creating mountains of

[01:36:16] jobs. We're helping shift um how we how

[01:36:23] we how we build things uh back to United

[01:36:26] States in so many different plants,

[01:36:28] chips, computers, and of course these AI

[01:36:31] factories. I'm completely aware that

[01:36:35] that um and I have I have the benefit

[01:36:37] and this is a real real um a real gift

[01:36:41] uh with with uh mainstream investors,

[01:36:45] teachers, policemen who have somehow for

[01:36:49] whatever reason invested in Nvidia or

[01:36:52] because they watch Jim Kramer um bought

[01:36:55] some stock and now are millionaires.

[01:36:57] >> Mhm. And um I I am completely aware of

[01:37:02] that circumstance. I'm aware of the

[01:37:04] circumstance that that Nvidia

[01:37:07] uh

[01:37:09] is central to a very large network of

[01:37:13] ecosystem partners behind us and

[01:37:14] downstream from us. And so the way the

[01:37:17] way I deal with that is exactly what I

[01:37:20] just did. I reason about

[01:37:23] what is it what is it that we're doing?

[01:37:26] um what is it causing? What's the impact

[01:37:28] that has other people benefit you know

[01:37:31] positively or even even um uh through

[01:37:35] great burden for example the supply

[01:37:36] chain

[01:37:38] and and the question is

[01:37:42] uh therefore what are you going to do

[01:37:43] about it and almost everything that I

[01:37:46] feel I break it down I reason about okay

[01:37:50] what's the circumstance what is what has

[01:37:52] changed what's hard um and what am I

[01:37:55] going to do about it and I I break it

[01:37:57] down, decompose the problem. And the de

[01:38:01] the decomposition

[01:38:03] of these

[01:38:05] circumstances

[01:38:07] turns it into manageable things that I

[01:38:09] can do. And the only thing that I after

[01:38:11] that I could do is did you do it? Did

[01:38:14] you either do it or did you get somebody

[01:38:16] else to do it? And if you didn't do it,

[01:38:18] you you reason that you need to do it

[01:38:20] and you didn't do it and you get didn't

[01:38:21] get anybody else to do it, then stop

[01:38:23] crying about it, you know. And so, and

[01:38:27] so, so I I'm I'm fairly I'm fairly uh

[01:38:31] >> uh tough on myself. And but I also break

[01:38:34] things down so that so that um uh I

[01:38:37] don't panic. Uh I can go to sleep

[01:38:39] because I've made the list of things

[01:38:41] that needed to be done. And I've made

[01:38:43] sure that everything that could put our

[01:38:46] company in harm's way, could put my

[01:38:48] partners in harm's way, put our industry

[01:38:50] in harm's way, I've told somebody.

[01:38:54] Everything that I feel could put anybody

[01:38:57] in harm's way, I've told someone. And

[01:39:00] I've told that someone who could do

[01:39:02] something about it. And so I've gotten

[01:39:04] it off my chest or I'm doing something

[01:39:06] about it. And so after that, Lex, what

[01:39:10] else can you do?

[01:39:10] >> So given all the in insane intense

[01:39:13] amount of suffering on the journey of

[01:39:17] building up Nvidia,

[01:39:19] you have you hit low points

[01:39:21] psychologically?

[01:39:22] >> Oh yeah. Oh yeah, sure. All the time.

[01:39:26] All the time.

[01:39:27] >> And there you just break down the

[01:39:29] problem

[01:39:30] >> into pieces.

[01:39:31] >> Yeah.

[01:39:31] >> See what you can do about it.

[01:39:33] >> And and part of And you know, Lex, part

[01:39:36] of it part of it is forgetting. One of

[01:39:39] the most important attributes of AI

[01:39:41] learning as you know is right systematic

[01:39:43] forgetting. You you need to know when to

[01:39:46] forget some things. You can't memorize

[01:39:48] everything. You can't keep everything.

[01:39:50] and and you know you want to you don't

[01:39:52] want to carry everything. One of the

[01:39:53] things that I do very quickly is I

[01:39:55] decompose the problem. I reason about

[01:39:57] the problem and I I share the load with

[01:39:59] it. When I say I tell everybody, I'm

[01:40:02] essentially sharing that burden.

[01:40:03] >> Yeah.

[01:40:04] >> As quickly as possible.

[01:40:06] Whatever worries me, tell somebody else.

[01:40:08] Don't just keep it, you know, decompo.

[01:40:10] Don't don't freak them out. decompose

[01:40:13] the problem into smaller parts and get

[01:40:16] people to so and and inspire them to be

[01:40:19] able to go do something about it. But

[01:40:21] part of it is just just forgetting, you

[01:40:23] know, I a lot of it is you got to be

[01:40:25] tough on yourself, you know, just come

[01:40:27] on, stop crying about it, let's get

[01:40:29] going, you know, and and then you get

[01:40:31] out of bed. And then the other part is

[01:40:33] is um you you you're attracted to the

[01:40:36] next shiny light, the next future, you

[01:40:38] know, the next opportunity, the next

[01:40:40] Okay, that's behind us. Let what's next?

[01:40:43] It's a lot. I think you know you watch

[01:40:45] this with great athletes. They they um

[01:40:48] just worry about the next point.

[01:40:49] >> Mhm.

[01:40:50] >> The last point is behind them. The

[01:40:52] embarrassment,

[01:40:54] the you know, the setback, you know, and

[01:40:57] and then and because I do so much of my

[01:40:59] job publicly,

[01:41:01] >> you know, Lex, you do a fair amount of

[01:41:03] your job publicly, too. And so, so I do

[01:41:05] a lot of my job publicly. And so, um,

[01:41:08] you know, I I say a lot of things that

[01:41:10] that seem sensible at the time or funny

[01:41:13] at the time. Mostly it's just because

[01:41:14] it's funny to me at the time and then,

[01:41:16] you know, you reflect on it's less

[01:41:18] money, but but

[01:41:20] >> Yeah. No, trust me, I know. But you

[01:41:22] basically allow yourself to be pulled by

[01:41:25] the light of the future. Forget the past

[01:41:27] and just keep

[01:41:28] >> That's right.

[01:41:28] >> Keep keep working towards that. I mean,

[01:41:30] you did say there's this kind of famous

[01:41:33] thing you said that um if you knew how

[01:41:36] hard it would be to build Nvidia, uh it

[01:41:40] turned out to be what is it a million

[01:41:42] times more hard than you anticipated

[01:41:45] that you wouldn't do it?

[01:41:46] >> Yeah.

[01:41:47] >> But is isn't you know when I hear that

[01:41:51] that's probably true about everything

[01:41:52] worth doing, right?

[01:41:53] >> Exactly. That is by the way what I was

[01:41:56] trying to explain is that there's a

[01:41:59] there's a incredible superpower of being

[01:42:03] um being being uh have the mind of a

[01:42:06] child.

[01:42:07] >> Yeah.

[01:42:07] >> You know, and I say to myself often

[01:42:09] times when I look at something and and

[01:42:12] almost almost everything

[01:42:15] um my first thought is how hard can it

[01:42:17] be?

[01:42:19] >> You know, and so and so you get yourself

[01:42:22] into that mode. How hard could it be?

[01:42:24] and and nobody's ever done it. It looks

[01:42:27] gigantic. It's going to cost hundreds of

[01:42:30] billions of dollars. It's going to take,

[01:42:32] you know, all this. And you just go,

[01:42:33] "Yeah, but how hard could it be?" You

[01:42:35] know, how hard could it be?

[01:42:37] >> And and so, so you got to get yourself

[01:42:39] into that state of mind. You don't want

[01:42:41] to you don't want to actually

[01:42:43] overstimulate

[01:42:45] everything and all the setbacks and all

[01:42:47] the trials and tribulations and all the

[01:42:49] disappointments. You don't want to

[01:42:50] simulate all that in advance. You don't

[01:42:51] want to know that. You don't you don't

[01:42:53] you want to go into a new experience

[01:42:55] thinking it's going to be perfect. It's

[01:42:57] going to be great. It's going to be

[01:42:58] incredibly fun. And then while you're

[01:43:00] there, you know, you need to have you

[01:43:03] need to have endurance. You need to have

[01:43:05] grit so that when the setbacks actually

[01:43:07] happened and those setbacks are going to

[01:43:09] surprise you, the disappoints

[01:43:11] disappointments aren't going to surprise

[01:43:13] you. You know, the embarrassments are

[01:43:14] going to surprise you, the humiliations

[01:43:16] are going to surprise you. Um you just

[01:43:18] can't let now you just got to turn on

[01:43:20] the other bit which is just forget about

[01:43:22] it. move on, keep keep moving. And and

[01:43:24] to the extent that

[01:43:27] to the extent that my assumptions

[01:43:31] about the future and why the future is

[01:43:34] going to manifest,

[01:43:36] so long as those assumptions and that

[01:43:38] input

[01:43:40] doesn't change or didn't change

[01:43:42] materially, then I should expect that

[01:43:44] the output won't change. And so my

[01:43:46] simulated output of the future is still

[01:43:50] going to happen. And if it's still going

[01:43:51] to happen, I'm still going to go after

[01:43:54] it. I believe it's going to, you know,

[01:43:55] and so there's a combination of two or

[01:43:58] three human characteristics.

[01:44:01] The ability to go into a into an

[01:44:03] experience fresh-minded,

[01:44:05] the ability to forget the setbacks,

[01:44:08] the ability to believe in yourself,

[01:44:11] you know, to believe what you believe

[01:44:13] and stay stay true to that belief. Um,

[01:44:16] but you're constantly re-evaluating.

[01:44:19] >> Mhm. This combination of three, four,

[01:44:22] five things I think is is really

[01:44:25] important for resilience. And and um

[01:44:30] and you know, I I'm I'm fortunate that

[01:44:32] that whatever whatever life experience

[01:44:33] has led to this, I've got kind of those

[01:44:36] four or five things. You know, I'm

[01:44:38] always curious, always learning. I'm

[01:44:40] always learning from everybody. You

[01:44:42] know, I'm always asking my and because

[01:44:44] I'm humble about about about everything,

[01:44:47] I'm always thinking, gosh, they did that

[01:44:49] so nicely. They did that so wonderfully.

[01:44:52] You know, I wonder what they're thinking

[01:44:54] through. How do they, you know, so I'm

[01:44:56] simulating everybody in a lot of ways,

[01:44:59] you know, I'm emulating almost everybody

[01:45:00] I watch, right? you're you're empathetic

[01:45:02] towards towards everything that they do

[01:45:04] that that you're observing and respect

[01:45:06] and and so you you're constantly

[01:45:09] learning and you know

[01:45:11] >> you're now one of the wealthiest people

[01:45:13] on earth, one of the most successful

[01:45:16] humans on earth. Is it harder to be

[01:45:19] humble and to be able to do you feel the

[01:45:22] effect of money and power and fame in

[01:45:26] making it harder for you to

[01:45:28] sort of be wrong in your own head enough

[01:45:32] to

[01:45:34] hear out an opinion of somebody else

[01:45:35] when it disagrees with you and learn

[01:45:37] from them? Those kinds of things.

[01:45:41] >> Um, surprisingly, no. And and I would I

[01:45:44] would actually go the other way because

[01:45:46] I do so much of my work publicly.

[01:45:50] When I'm wrong, pretty much everybody

[01:45:52] sees it.

[01:45:53] >> You get humbled.

[01:45:54] >> Yeah. And and uh and when I'm wrong,

[01:45:57] when I'm wrong or it didn't turn out

[01:45:59] that way or um you know, I mean, most of

[01:46:03] the things that that I say outside um

[01:46:06] I'm fairly certain about. And the reason

[01:46:08] for that is because because it's going

[01:46:10] to impact somebody else and I want to be

[01:46:12] quite concerned about that and quite

[01:46:14] quite circumspect about that. Um for

[01:46:17] stuff that that I'm reasoning about

[01:46:18] inside a meeting, you know, a lot of

[01:46:21] things could turn out differently. And

[01:46:23] so, but it doesn't ever stop me from

[01:46:25] reasoning. The way that that the way

[01:46:27] that I manage and lead, I you know, I'm

[01:46:30] constantly reasoning in front of people.

[01:46:32] Even when I'm talking to you, you can

[01:46:34] kind of see me kind of reasoning through

[01:46:35] things.

[01:46:35] >> And I want to make sure that you

[01:46:37] understand what I'm saying, not because

[01:46:38] I told you,

[01:46:40] >> because I'm so humble about what I'm

[01:46:42] about to tell you.

[01:46:43] >> I kind of show you the steps that I got

[01:46:45] there.

[01:46:46] >> And then you could decide whether you

[01:46:47] believe what I said in the end. And so

[01:46:49] I'm doing that all day long in meetings

[01:46:52] with all of my employees. I'm constantly

[01:46:54] reasoning through. Let me tell you, let

[01:46:55] me tell you what how I see it. And I

[01:46:57] reason through it. It gives everybody

[01:46:59] the opportunity to intercept and say, "I

[01:47:02] disagree with that part."

[01:47:04] >> The nice thing about reasoning through

[01:47:05] things and letting and letting people

[01:47:07] interact with it is that they don't have

[01:47:09] to disagree with your outcome.

[01:47:12] They can disagree with your reasoning

[01:47:13] steps and they could pull me in

[01:47:16] different directions and then we can

[01:47:18] reason forward. And so we're we're kind

[01:47:20] of, you know, collective

[01:47:25] patharching method and it's really

[01:47:27] fantastic.

[01:47:29] >> Yeah. You have this way about you of

[01:47:32] when you're explaining stuff, I can feel

[01:47:34] you actually reasoning on the spot about

[01:47:37] it with a constant open-mindedness where

[01:47:40] you could I I could feel like I could

[01:47:42] steer your thinking. Yeah. And that's a

[01:47:44] that's really beautiful that you've been

[01:47:46] able to maintain that after so many

[01:47:48] years of success and pain. I think

[01:47:50] sometimes pain makes you close you down

[01:47:54] a bit.

[01:47:55] >> Yeah.

[01:47:56] >> And I I think to maintain

[01:47:57] >> tolerance for embarrassment I think is

[01:48:00] >> that's that's the tolerance. I mean

[01:48:01] that's a real thing.

[01:48:03] >> Yeah. There's many years of embarrassing

[01:48:05] yourself. Even those meetings knowing

[01:48:07] that there's people around you where you

[01:48:09] declared one idea and it was shown that

[01:48:12] that idea was wrong and be able to admit

[01:48:14] that and to grow from that. That's not

[01:48:15] that's very difficult on a human level.

[01:48:17] >> Yeah. Well, you know, they knew I was

[01:48:20] they knew that recently my first job was

[01:48:22] was, you know, cleaning toilets. So,

[01:48:25] >> I'm glad you maintain that same spirit

[01:48:27] of Denny's um the the work. I mean, that

[01:48:30] that was beautiful. your whole journey

[01:48:32] from starting from Denny's is a

[01:48:33] beautiful one. Uh let me ask you about

[01:48:37] video games. So I'm a big gaming fan.

[01:48:40] >> Yeah.

[01:48:41] >> So I have to say thank you to Nvidia for

[01:48:44] many years of incredible graphics.

[01:48:46] Um

[01:48:47] >> by the way it it is GeForce is our still

[01:48:50] to this day.

[01:48:51] >> Yeah.

[01:48:51] >> Our number one marketing strategy.

[01:48:55] Right. People learn about Nvidia while

[01:48:57] they're in their teenage years.

[01:48:59] >> Mhm. And then they go to college and

[01:49:01] they know who Nvidia is and they and

[01:49:02] then in the beginning it's just you know

[01:49:05] playing Call of Duty you know you know

[01:49:06] Fortnite and then later they're using

[01:49:08] CUDA and then later they're using Nvidia

[01:49:10] and you know Blender and Do and Auto.

[01:49:16] >> I mean I should say I I mentioned to a

[01:49:18] friend that I'm uh talking with you. He

[01:49:21] said oh they make

[01:49:23] great gaming GPUs.

[01:49:25] >> Yeah. Exactly. Exactly. you know,

[01:49:27] there's there's more to it, but but

[01:49:30] yeah. Yeah, people really love the it

[01:49:32] really brought a lot of joy to a lot of

[01:49:34] people. The the the hardware really

[01:49:36] brings these worlds to life.

[01:49:38] >> Uh there was some controversy around

[01:49:41] this uh with DLSS 5. Yeah.

[01:49:44] >> Can you explain to me the drama around

[01:49:45] this? Uh I guess people gamers online

[01:49:49] were concerned that it makes games look

[01:49:52] like AI slop.

[01:49:53] >> Yeah.

[01:49:54] >> Uh what do you think of this drama?

[01:49:56] Yeah, I think their their perspective

[01:49:59] makes sense and I could see where

[01:50:02] they're coming from because I don't love

[01:50:04] AI slob myself. You know, all of the the

[01:50:07] AI generated content increasingly

[01:50:10] um looks similar and they're all

[01:50:13] beautiful and and I can so I can I'm

[01:50:16] empathetic towards what they're what

[01:50:17] they're thinking. Um that's just not

[01:50:19] what DLSS 5 is trying to do. I showed

[01:50:22] several examples of it, but DLSS5

[01:50:26] is 3D conditioned, 3D guided. It's

[01:50:30] ground truth structure data guided. And

[01:50:33] so, so the artist determine the

[01:50:35] geometry. We are completely truthful

[01:50:39] to the geometry maintain so in every

[01:50:42] single frame. Um it's uh conditioned by

[01:50:45] the textures, the artistry of the

[01:50:48] artist. And so every single frame it

[01:50:51] enhances but it doesn't change anything.

[01:50:55] Now the question is the question about

[01:50:57] enhancing.

[01:50:59] DLSS5 also lets because it's the system

[01:51:02] is open you could train your own models

[01:51:05] to determine and you could even in the

[01:51:08] future prompt it you know I want it to

[01:51:10] be a toune shader. I want it to look

[01:51:12] like this kind of, you know, so you can

[01:51:13] give it even an example and it would

[01:51:16] generate in the style of that all

[01:51:19] consistent with the artistry, you know,

[01:51:22] the style, the intent of the artist. And

[01:51:26] so all of that is done for the artist so

[01:51:30] that they can create something that is

[01:51:32] more beautiful um but still in the style

[01:51:35] that they want.

[01:51:37] I think that they got the impression

[01:51:40] that the the games are going to come out

[01:51:43] the way the games are shipped the way

[01:51:45] they do and then we're going to

[01:51:47] post-process it. That's not what DLSS is

[01:51:50] intended to do. DLSS is integrated with

[01:51:53] the artist. And so it's it's about

[01:51:55] giving the artist the tool of AI, the

[01:51:58] tool of generative AI. They could decide

[01:52:00] not to use it. You know,

[01:52:01] >> I think people are very sensitive to

[01:52:02] human faces.

[01:52:03] >> Yeah. And we're now living in this

[01:52:05] moment, which I think is a is a

[01:52:07] beautiful one, which is people are

[01:52:09] sensitive to AI slop.

[01:52:10] >> Yeah.

[01:52:11] >> It it puts a mirror to ourselves to help

[01:52:14] us realize that what we seek as

[01:52:15] imperfections, what we seek is sometimes

[01:52:18] not perfect graphics, it helps us

[01:52:20] understand what we find compelling in

[01:52:23] the worlds we create.

[01:52:25] >> And that's beautiful. And as long as

[01:52:26] it's tools that help us create those

[01:52:27] worlds.

[01:52:28] >> Yeah, that's right.

[01:52:29] >> It's it's wonderful.

[01:52:30] >> That's right. It's yet another tool. and

[01:52:31] they want the generative uh models to

[01:52:35] generate the opposite of photoreal.

[01:52:38] >> Mhm.

[01:52:38] >> Yeah. It'll do that too. And so it's

[01:52:40] just yet another tool. I think the um

[01:52:43] the gamers might might also appreciate

[01:52:46] that that um in the last couple years we

[01:52:50] we introduced

[01:52:53] uh skin shaders

[01:52:55] to the game developers and many of those

[01:52:58] games have skin shaders that include

[01:53:00] subs subsurface scattering that make

[01:53:03] skin look more skin-like. And so the

[01:53:06] industry's game developers are looking

[01:53:09] for more and more and more tools to

[01:53:12] express their art. And so this is just

[01:53:14] yet more one more tool they could decide

[01:53:16] what to use.

[01:53:16] >> Ridiculous question. Uh what do you

[01:53:18] think is the greatest or most

[01:53:20] influential game ever made? Maybe from

[01:53:22] Nvidia's perspective.

[01:53:24] >> Doom.

[01:53:25] >> Doom. Unquestionably. That was the start

[01:53:27] of the 3D. I would say Doom from a from

[01:53:30] a the intersection of the cultural

[01:53:32] implication as well as the industry

[01:53:35] turning a PC into a gaming device. That

[01:53:39] was a very important moment. Now, of

[01:53:40] course, flight simulation companies were

[01:53:42] before it

[01:53:44] >> and um but they just didn't have the

[01:53:46] popularity that Doom did to have made

[01:53:48] the industry turned the PC from a office

[01:53:51] automation tool into a personal computer

[01:53:55] for families and gamers and things like

[01:53:57] that. And so Doom was really impactful

[01:53:58] there. From a from an actual game

[01:54:00] technology perspective, I would say

[01:54:02] Virtual Fighter. And so we we're great

[01:54:05] friends with both of them, you know. And

[01:54:07] then there's games more recently. I

[01:54:09] mean, Cyberpunk 2077,

[01:54:12] really nice GPU,

[01:54:15] accelerated graphics, like fully ray

[01:54:17] traced,

[01:54:17] >> fully ray traced. Um, also I like I

[01:54:20] personally I'm a huge fan of Skyrim, uh,

[01:54:22] Elder Scrolls and the, you know, it's

[01:54:25] been released a long long time ago, but

[01:54:27] people release mods and they

[01:54:31] I mean it it's like a different game and

[01:54:33] it just allows me to replay the game

[01:54:35] over and over and it get it makes you

[01:54:38] realize you can reexperience in a

[01:54:41] totally new way the world you already

[01:54:44] love.

[01:54:44] >> So I I do that all the time. One of my

[01:54:46] favorite games just walk around Skyrim.

[01:54:48] We created this thing called RTX Mod.

[01:54:50] >> Uhhuh.

[01:54:51] >> Yeah. It's a modding tool.

[01:54:52] >> Awesome.

[01:54:53] >> And allows it allows the community to

[01:54:55] inject the latest technology into an old

[01:54:59] game.

[01:55:00] >> Of course, like what makes a great video

[01:55:01] game is not just graphics. It's also

[01:55:03] story and character development. But

[01:55:06] that's right. Beautiful graphics can add

[01:55:08] to the the immersion, the the feeling

[01:55:11] like it's another place you're

[01:55:13] transported to.

[01:55:16] uh what's uh you said I think accurately

[01:55:18] that the AGI timeline

[01:55:22] question rests on your definition of

[01:55:24] AGI.

[01:55:26] So let let's let me ask you about a

[01:55:29] possible timelines here. Let's this

[01:55:32] ridiculous definition perhaps of what

[01:55:34] AGI is but a an AI system that's able to

[01:55:39] essentially do your job. So run, no

[01:55:44] start,

[01:55:46] grow and run a successful technology

[01:55:50] company that's worth

[01:55:52] >> a good one or A1.

[01:55:54] >> No, it has to it has to be worth more

[01:55:56] than a billion

[01:55:58] more more than a billion dollars.

[01:56:01] So you know, you know how hard it is to

[01:56:05] do all those components. So how far are

[01:56:07] we away from that? So we're talking

[01:56:10] about open claw that does all the

[01:56:14] incredibly complex stuff that are

[01:56:16] required to to first of all innovate to

[01:56:20] find customers to sell to them to to

[01:56:22] manage to build a team of some agents

[01:56:26] some humans all that kind of stuff. Is

[01:56:28] this 5 10 15 20 years away?

[01:56:31] >> I think it's now I think we've achieved

[01:56:33] AGI.

[01:56:34] >> You think you can have a company run by

[01:56:36] an AI system like this?

[01:56:37] >> Possible. And the reason for that is

[01:56:39] this. You said a billion and you didn't

[01:56:41] say forever and and so for example uh it

[01:56:46] is not out of the question that

[01:56:50] uh a claw was able to create a web

[01:56:53] service some interesting little app that

[01:56:59] all of a sudden you know a few billion

[01:57:03] people used for 50 and then it went out

[01:57:08] of business again shortly after. Now, we

[01:57:09] saw a whole bunch of those type of

[01:57:10] companies during the internet era and

[01:57:13] most of the those websites were not

[01:57:15] anything more sophisticated than what

[01:57:19] Open Claw could generate today.

[01:57:20] >> Interesting. Achieve virality and

[01:57:22] monetize that virality.

[01:57:24] >> Yeah. It's just that I don't know what

[01:57:25] it is, but I I couldn't have predicted

[01:57:27] any of those companies at the time

[01:57:28] either. You know,

[01:57:30] >> you're going to get a lot of people

[01:57:31] excited with that statement.

[01:57:32] >> Yeah. It's like, what do you mean? I can

[01:57:34] I can just uh launch an agent and u make

[01:57:38] a lot of money? Well, by the way, it's

[01:57:39] happening right now, right? You know

[01:57:40] that when when you go to China, uh

[01:57:42] you're going to see you're going to see

[01:57:44] um a whole bunch of people uh teaching

[01:57:47] their getting their claws to try to go

[01:57:48] out and look for jobs and, you know, do

[01:57:51] work, make money. And and I'm not I'm

[01:57:55] not actually I wouldn't be surprised if

[01:57:57] some social thing happened or somebody

[01:57:59] created a a digital influencer, super

[01:58:02] super cute. um or some social

[01:58:05] application that you know feeds your

[01:58:08] little tomagotchi or something like that

[01:58:09] and and it become an out of the blue an

[01:58:13] instant success. A lot of people use it

[01:58:16] for a couple of months and it kind of

[01:58:17] dies away. Now the odds of of of

[01:58:22] you know 100,000 of those agents um

[01:58:25] building Nvidia 0%.

[01:58:28] And and then and then the the one part

[01:58:30] that I will I will do um and I and I I

[01:58:34] want to make sure we all do is to

[01:58:36] recognize that people are really worried

[01:58:38] about their jobs

[01:58:40] and and um I just want to remind them

[01:58:43] that the purpose of your job and the

[01:58:47] tasks and the tools that you use to do

[01:58:50] your job are related, not the same. I've

[01:58:53] been doing my job for 33 years. I'm the

[01:58:55] longest running tech CEO in the world.

[01:58:57] 34 years and the tools that I've used to

[01:59:00] do my job has changed

[01:59:03] continuously in the last 34 years and

[01:59:06] sometimes quite dramatically you know

[01:59:09] over the course of a couple two three

[01:59:10] years and and the the the one story that

[01:59:13] I I I really want to make sure that

[01:59:14] everybody hears is the story the the

[01:59:18] first job that every that computer

[01:59:20] scientists said AI researchers said was

[01:59:22] going to go away was radiology

[01:59:25] because computer vision was going to

[01:59:26] achieve superhuman levels and it did. CV

[01:59:31] computer vision was superhuman in 2019

[01:59:35] 20 maybe maybe a little bit later 2020.

[01:59:38] >> Mhm.

[01:59:39] >> Okay. And so it's been a long time since

[01:59:41] computer vision has been superhuman. And

[01:59:43] so the prediction was radiologists would

[01:59:45] go away because studying radiology scans

[01:59:48] was thing of the past. AI will do that.

[01:59:50] Well, they were absolutely right.

[01:59:54] Computer vision is completely

[01:59:56] superhuman. Every radiology platform and

[01:59:59] package today is driven by AI.

[02:00:02] And yet the number of radiologists grew.

[02:00:06] And so the question is why? And we now

[02:00:08] have a shortage of radiologists in the

[02:00:10] world. And so one the alarmist

[02:00:15] warning went too far and has scared

[02:00:17] people from

[02:00:20] doing this profession that is so

[02:00:21] important to society. And so it did

[02:00:24] harm. Now why was it wrong? The reason

[02:00:27] why is because the purpose of a

[02:00:30] radiologist, the purpose is to diagnose

[02:00:32] disease and help patients and doctors

[02:00:36] diagnose disease.

[02:00:38] And because we're able to study scans so

[02:00:41] much faster now, you could study more

[02:00:43] scans. You could diagnose better. You

[02:00:46] could you could um impatient faster. We

[02:00:50] can see people more. the hospitals are

[02:00:53] making more money. You have more

[02:00:54] patients in the hospital. You need more

[02:00:56] radiologists. I mean the the amazing

[02:00:58] thing is it's so obvious this was going

[02:01:02] to happen. The number of software

[02:01:04] engineers at NVIDIA is going to grow,

[02:01:05] not decline.

[02:01:08] And the reason for that is because the

[02:01:10] purpose of a software engineer and the

[02:01:12] task of a software engineer for coding

[02:01:14] are related, not the same. I wanted my

[02:01:17] software engineers to solve problems. I

[02:01:19] didn't care how many lines of code they

[02:01:20] wrote.

[02:01:22] You know, but their job, their purpose

[02:01:24] of their job didn't change. Solving

[02:01:26] problems, working as a team, diagnosing

[02:01:28] problems, evaluating the result, looking

[02:01:32] for new problems to solve innovation,

[02:01:35] connecting dots, you know, none of that

[02:01:38] stuff is going to go away.

[02:01:39] >> So, you think it's possible that let's

[02:01:41] even take coding, you think the number

[02:01:43] of programmers in the world might

[02:01:44] increase, not decrease?

[02:01:47] >> And the reason for that is this. What is

[02:01:49] the definition of coding?

[02:01:52] I believe that is the definition coding

[02:01:54] as of today is simply specifying

[02:01:58] specification and maybe if you want to

[02:02:01] be rather directive you could even give

[02:02:04] it an architecture of the software that

[02:02:06] you're you wanted to write. So the

[02:02:08] question is how many people could do

[02:02:10] that? Describe a specification for a

[02:02:13] computer to go telling the computer what

[02:02:15] to go build. How many people? I think we

[02:02:18] just went from 30 million to probably 1

[02:02:20] billion.

[02:02:22] And so every every carpenter in the

[02:02:25] future will be a coder. Except a

[02:02:28] carpenter with AI is also an architect.

[02:02:33] They just increased the value that they

[02:02:34] could deliver to the customer. Their

[02:02:37] their

[02:02:39] artistry just elevated tremendously.

[02:02:43] I believe that every accountant is, you

[02:02:45] know, also your financial analyst, also

[02:02:47] your financial adviser. So all of these

[02:02:50] professions have just been elevated and

[02:02:53] if I were a carpenter, I sees a I see

[02:02:55] AI, I would just completely go berserk.

[02:02:58] You know, the services I can bring to my

[02:03:00] clients, if I were a plumber, completely

[02:03:03] go berserk. and the the people that are

[02:03:05] currently programmers and software

[02:03:07] engineers, I think they're at the

[02:03:09] cutting edge of understanding

[02:03:11] intuitively how to communicate

[02:03:15] with the agents using natural language

[02:03:17] in order to design the best kind of

[02:03:19] software.

[02:03:20] >> That's right. So over time they'll

[02:03:22] converge but I think uh there's still

[02:03:25] value in getting I think uh learning how

[02:03:27] to program like learning what

[02:03:29] programming languages are uh the old the

[02:03:32] old kind of programming uh what what are

[02:03:35] good practices for programming languages

[02:03:37] what are design principles for

[02:03:39] programming languages for large software

[02:03:42] systems

[02:03:43] >> and and the reason for that lex and you

[02:03:46] know I just say for the audience I think

[02:03:49] the goal of the goal of specification,

[02:03:52] the artistry of specification, the goal

[02:03:55] and the artistry of it um is going to

[02:03:58] depend on what problem you're trying to

[02:04:00] solve. when I'm thinking when I'm

[02:04:02] thinking about giving the company

[02:04:04] strategies and um formulating corporate

[02:04:07] directions and things that we should do

[02:04:10] um I describe it at a level that is

[02:04:14] sufficiently

[02:04:16] specific that people generally

[02:04:19] understand the direction and it's

[02:04:21] actionable they it's so specific enough

[02:04:24] that they can take action on it but I

[02:04:26] underspecify it on purpose so that

[02:04:30] enable 43 3,000 amazing people to make

[02:04:33] it even better than I imagined.

[02:04:36] And so when I'm working with engineers,

[02:04:39] when I'm working with people, um, I

[02:04:41] think about who what problem am I trying

[02:04:43] to solve? Who am I working with?

[02:04:47] And the level of specification, the

[02:04:51] level of architecture definition

[02:04:54] relates to that. And and so

[02:04:59] everybody's going to have to learn how

[02:05:01] where in the spectrum of coding they

[02:05:03] want to be. Writing a specification is

[02:05:05] coding. And so you might decide to be

[02:05:08] quite prescriptive because there's a

[02:05:10] very specific outcome you're looking

[02:05:11] for. You might decide that you know this

[02:05:14] is an area you want to be much more

[02:05:16] exploratory. And so you might

[02:05:18] underspecify and enable you to go back

[02:05:21] and forth with the AI to even push your

[02:05:23] own boundaries of creativity. And so

[02:05:26] this artistry of where you are in the

[02:05:28] spectrum, this is the future of coding.

[02:05:31] >> But just to linger on it, outside of

[02:05:32] coding, I think a lot of people

[02:05:34] rightfully so

[02:05:36] uh are worried about their jobs, have a

[02:05:38] lot of anxiety about their jobs,

[02:05:40] especially in the white collar sector.

[02:05:43] Um I don't think any of us know

[02:05:47] what to do

[02:05:50] uh with tumultuous times that always

[02:05:52] come when automations and new technology

[02:05:54] arrives. And I just

[02:05:57] first of all I think um

[02:06:01] we all need to have compassion and the

[02:06:03] responsibility to feel sort of the

[02:06:05] burden of what the actual suffering

[02:06:07] feels like for individual people and

[02:06:09] families that lose their job. I think

[02:06:11] whenever you have transformative

[02:06:13] technology like that's coming with with

[02:06:15] artificial intelligence, there's going

[02:06:17] to be a lot of pain and I don't know

[02:06:19] what to do about that uh pain.

[02:06:21] Hopefully, it creates much more

[02:06:22] opportunities for those same people uh

[02:06:26] for the same kind of job as uh the

[02:06:30] tooling evolves and makes them more

[02:06:32] productive and makes it more fun

[02:06:34] hopefully as it does in the programming.

[02:06:36] I've I haven't I've been having so much

[02:06:37] fun programming, I have to say. like

[02:06:39] I've never had this much fun. So

[02:06:41] hopefully it makes their job automates

[02:06:42] the boring parts and makes the creative

[02:06:45] parts uh the ones that the the human

[02:06:48] beings are responsible for. But still

[02:06:49] there's going to be a lot of pain and

[02:06:51] suffering. So my first recommendation

[02:06:53] before and this is now how I deal with

[02:06:56] anxiety. In fact, we just talked about

[02:06:57] it earlier.

[02:06:58] >> Mhm.

[02:06:59] >> Enormous anxiety about the future,

[02:07:01] enormous anxiety about the pressure,

[02:07:02] enormous anxiety about uncertainty.

[02:07:05] I first break it down and then I'm going

[02:07:08] to tell myself,

[02:07:10] okay, there are some things you can do

[02:07:12] something about. There are some things

[02:07:13] you can't do anything about, but for the

[02:07:15] stuff that you can do something about,

[02:07:17] let's reason reason about it and let's

[02:07:19] go do it.

[02:07:20] >> If we were to hire a new college

[02:07:21] graduate today and I have a choice

[02:07:24] between two, one that have that is no

[02:07:28] clue what AI is and one that is expert

[02:07:32] in using AI. I would hire the one who's

[02:07:35] expert in using AI. If I had an

[02:07:38] accountant, a marketing person, the one

[02:07:42] that is expert in using AI, supply

[02:07:45] chain, customer service, a salesperson,

[02:07:48] business development, a lawyer,

[02:07:51] I would hire the one who is expert in

[02:07:54] using AI. And so I would I would advise

[02:07:57] that every college student, every every

[02:08:00] teacher should encourage their student

[02:08:02] to to go use AI. Every college student

[02:08:06] should graduate and be an expert in AI.

[02:08:08] And every everybody, if you're a

[02:08:10] carpenter, if you're, you know,

[02:08:12] electrician, go use AI. Go see what it

[02:08:16] can do to transform your current job.

[02:08:19] Elevate yourself. If I were a farmer, I

[02:08:22] would absolutely use AI. If I were a

[02:08:24] pharmacist, pharmacist, I would use AI.

[02:08:27] I want to see how what it could do to

[02:08:28] elevate my job so that I could be the I

[02:08:31] could be the innovator to revolutionize

[02:08:34] this industry myself.

[02:08:36] >> And so that that would be the first

[02:08:37] thing that I would do. And and then I

[02:08:39] would also I would also help them. Um it

[02:08:44] is the case that the technology will

[02:08:46] dislocate and will eliminate many tasks.

[02:08:52] If and because it will automate it. If

[02:08:54] your job is the task if your job is the

[02:08:58] task then you're very highly going to be

[02:09:01] disrupted.

[02:09:02] If your

[02:09:04] job's purpose includes you certain

[02:09:07] tasks. Mhm.

[02:09:08] >> Then it it's vital that you go learn how

[02:09:10] to use AI to automate those tasks. And

[02:09:12] then there's the world of spectrum in

[02:09:14] between.

[02:09:14] >> And by the way, the beautiful thing

[02:09:16] about AI, so the the the chatbot

[02:09:19] versions

[02:09:21] is you can break down you have anxiety

[02:09:24] and you can break down the problem by

[02:09:26] talking to it. Like I've I've recently

[02:09:29] it's really just incredible how much you

[02:09:31] can think through your life's problems

[02:09:33] and through and I don't mean like

[02:09:34] therapy problems. I mean like very

[02:09:36] practically, okay, I'm worried about my

[02:09:39] literally I'm worried about my job. What

[02:09:40] are the skills? What are the steps I

[02:09:42] need to take? How do I get better at AI?

[02:09:44] Everything you just said, you can

[02:09:45] literally ask and it's going to give you

[02:09:47] a point by point. I mean, it's just a

[02:09:50] great life coach. Period. This

[02:09:52] >> I don't know how to use AI. And the AI

[02:09:53] goes, well, let me show you.

[02:09:54] >> Exactly. It's very meta, but it's

[02:09:57] >> it's kind of incredible. So, people

[02:09:59] definitely should.

[02:10:00] >> You can't walk up to Excel and say, I

[02:10:01] don't know how to use Excel. You're

[02:10:02] done. I mean that's really what AI has

[02:10:05] done for me in all walks of life is that

[02:10:07] initial friction of being a beginner of

[02:10:10] using a thing for the first time. I can

[02:10:12] literally ask about any single thing.

[02:10:14] >> What are the first steps I need to take?

[02:10:16] >> That's right.

[02:10:17] >> And and that that handholding that it

[02:10:18] does removing the friction of uh all the

[02:10:22] experiences that the world offers is you

[02:10:25] know like like I mentioned to you

[02:10:26] offline you mentioned I'm I'm going to

[02:10:28] China and Taiwan.

[02:10:30] >> So awesome.

[02:10:32] for you. Where do I go? What where do I

[02:10:34] go? How do I all those questions

[02:10:35] immediately answered and it's beautiful?

[02:10:37] >> Well, when you when you go to Taiwan,

[02:10:39] just ask AI, what are Jensen's favorite

[02:10:42] restaurants in Taiwan?

[02:10:44] >> Yeah.

[02:10:44] >> And it'll actually Oh, yeah. Yeah.

[02:10:46] >> Is it accurate? Okay.

[02:10:46] >> Yeah. Yeah. All right.

[02:10:47] >> It's all over all over Taiwan.

[02:10:50] >> Well, you're you're a rock star over

[02:10:52] there and um and like we also mentioned

[02:10:54] offline, maybe our paths will cross,

[02:10:56] which would be really wonderful in

[02:10:57] Computex GTC Taiwan.

[02:11:01] Uh do you think there are some things

[02:11:04] about human nature about human

[02:11:06] consciousness

[02:11:08] that is

[02:11:10] fundamentally non-computational

[02:11:12] maybe something a chip no matter how

[02:11:14] powerful uh can never replicate? I don't

[02:11:18] know if the chip will ever get nervous

[02:11:20] and that's the you know of course the

[02:11:22] conditions by which uh that causes

[02:11:26] anxiety or nervousness or whatever

[02:11:28] emotion um I believe that AI will be

[02:11:34] able to recognize those and understand

[02:11:37] those. I don't think my chips will feel

[02:11:40] those and therefore the how how that

[02:11:44] anxiety, how that feeling, how that

[02:11:46] excitement, how that how that you know

[02:11:50] all of those feelings manifest in human

[02:11:53] performance for example extremely

[02:11:56] amazing human performance, athletic

[02:11:57] performance, you know, average or lesser

[02:12:00] than average. um that that entire

[02:12:02] spectrum of human performance that comes

[02:12:05] out of exactly the same circumstances

[02:12:08] for different people manifesting in

[02:12:11] different outcome

[02:12:13] manifesting in different performance. I

[02:12:16] I don't think there's anything about

[02:12:19] anything that we're building that would

[02:12:20] suggest that two different computers

[02:12:24] being presented with all of exactly the

[02:12:26] same context would per of course it

[02:12:29] would produce statistically different

[02:12:31] outcomes but it's not because it felt

[02:12:33] different.

[02:12:34] >> Yeah. The subjective

[02:12:36] boy there's something truly special

[02:12:38] about the subjective experience

[02:12:41] that we humans feel. Like I mentioned to

[02:12:43] you, I was I was I was pretty nervous

[02:12:46] talking to you like I mentioned to you

[02:12:48] that the hope the fear the anxiety and

[02:12:51] just life itself the richness of life

[02:12:54] how amazing everything is how deeply we

[02:12:56] fall in love how deeply our hearts get

[02:12:58] broken how afraid we are of death and

[02:13:01] how much pain we feel when our loved

[02:13:03] ones pass away all of that the whole

[02:13:06] thing I don't it's very hard to think AI

[02:13:10] being able to a computational device

[02:13:12] being able to do that but there's so

[02:13:14] many mysteries about this whole thing

[02:13:16] that we're yet to uncover that I am open

[02:13:19] to be surprised.

[02:13:21] >> I've been surprised a lot over the past

[02:13:23] >> few months and few years. Scaling can

[02:13:26] create some incredible miracles in the

[02:13:28] space of intelligence

[02:13:31] >> has been truly marvelous to watch. So

[02:13:32] I'm open to surprise

[02:13:34] >> and and it's just really important to to

[02:13:37] break down what is intelligence and the

[02:13:39] word that word we use all the time. It's

[02:13:41] not a mysterious word. Intelligence has

[02:13:44] a meaning, you know,

[02:13:46] >> and it's a system that, you know, it's

[02:13:49] it it's something that we do that in

[02:13:51] includes perception and understanding

[02:13:54] and reasoning and the ability to do plan

[02:13:56] and you know that that loop that loop is

[02:14:00] is um the fundamentally what

[02:14:02] intelligence is. Intelligence is not one

[02:14:05] word that is exactly equal to humanity.

[02:14:11] And that's I think it's really important

[02:14:12] to separate the two. We have two words

[02:14:14] for that. I'm not I don't over fantasize

[02:14:18] about and I don't over romanticize about

[02:14:21] intelligence. Intelligence is and people

[02:14:25] have heard me say it before. I actually

[02:14:27] think intelligence is a commodity.

[02:14:29] I'm surrounded by intelligent people.

[02:14:33] And I'm surrounded by intelligent people

[02:14:34] more intelligent than I am in each one

[02:14:36] of the spaces that they're in. And yet I

[02:14:40] have a role in that circle. It's

[02:14:43] actually kind of interesting.

[02:14:45] They're more educated than I am.

[02:14:49] They went to better schools than I did.

[02:14:51] They're deeper than in any in this field

[02:14:54] that they're in. All of them. I have 60

[02:14:56] of them. They're all superhuman to me.

[02:14:59] >> And somehow I'm sitting in the middle

[02:15:01] orchestrating all 60 of them. And so you

[02:15:03] got to ask yourself,

[02:15:05] what is what is it about a dishwasher

[02:15:09] that allows that dishwasher to sit in

[02:15:11] the middle of superhumans?

[02:15:13] Does that make sense?

[02:15:15] >> And so, but that's my point. My point is

[02:15:18] intelligence is a is a functional thing.

[02:15:22] Humanity is not a not specified

[02:15:25] functionally.

[02:15:26] It's a much much bigger word. and and

[02:15:29] our life experience, our tolerance for

[02:15:33] pain, our determination,

[02:15:36] those are those are different words in

[02:15:37] intelligence. And so the the thing that

[02:15:40] I I want to help the audience

[02:15:42] understand, if I could give them one

[02:15:44] thing is is intelligence is a word that

[02:15:47] we've elevated to very high form over

[02:15:50] time. the the word we should really

[02:15:52] elevate is humanity, character,

[02:15:54] humanity, all of those things,

[02:15:57] compassion, generosity,

[02:16:00] all of the things that you say just now,

[02:16:03] >> I believe those are superhuman powers

[02:16:06] and that now intelligence is going to be

[02:16:08] commoditized because we've spoken about

[02:16:10] it. The most important thing is your

[02:16:12] education. The most now even even when

[02:16:15] they said the most important thing is

[02:16:16] your education. when you went to school,

[02:16:18] there's more than just knowledge that

[02:16:20] you gained.

[02:16:22] >> And so, but unfortunately, our society

[02:16:25] had put everything into one single word.

[02:16:28] And life is more than one word. And I'm

[02:16:31] just telling you, my life would suggest

[02:16:34] that being lower

[02:16:36] on the intelligence curve than everybody

[02:16:39] around me doesn't change the fact I'm

[02:16:42] the most successful. And so and and I

[02:16:45] think I think that that kind of is I'm

[02:16:47] trying hopefully to inspire everybody

[02:16:49] else that don't let this de

[02:16:51] democratization of intelligence, this

[02:16:53] commoditization of intelligence,

[02:16:57] you know, cause you anxiety. You should

[02:16:59] be inspired by that.

[02:17:00] >> Yeah. I I I think uh AI will help us

[02:17:02] celebrate humans more. And I'm certainly

[02:17:07] humanity and human first. And I I think

[02:17:11] what makes this world incredible is

[02:17:13] humans forever will be so. And just AI

[02:17:16] is this incredible tool that makes us

[02:17:18] >> That's exactly right.

[02:17:19] >> Humans more powerful.

[02:17:20] >> That's exactly right.

[02:17:21] >> Uh so much of the success of Nvidia

[02:17:25] and um the lives of millions of people

[02:17:28] that I mentioned uh depend on you.

[02:17:31] Uh but you're just one human like we

[02:17:33] mentioned u mortal like all of us. Do

[02:17:36] you think about your mortality? Are you

[02:17:39] afraid of death?

[02:17:42] >> I really don't want to die. Um, I have a

[02:17:45] great life. I have a great family.

[02:17:48] I have really important work.

[02:17:53] Uh,

[02:17:54] this is this is not a once in a once in

[02:17:58] a lifetime experience suggests that it

[02:18:02] has been experienced by many people just

[02:18:05] not one person. Uh this is a once in a

[02:18:08] humanity experience what I'm going

[02:18:10] through. Uh Nvidia is one of the most

[02:18:13] consequential technology companies in

[02:18:14] history. We're doing very important

[02:18:16] work. I take it very seriously.

[02:18:19] Um

[02:18:20] and and so some of the some of the

[02:18:22] things that that of course are are

[02:18:24] practical things like how do we think

[02:18:27] about succession planning and and um I

[02:18:32] I'm famous in saying that I don't

[02:18:33] believe in succession planning

[02:18:36] and and the reason the reason for that

[02:18:38] the reason for that isn't because I'm

[02:18:40] immortal. Um the reason for that is

[02:18:42] because if you're worried about

[02:18:46] succession planning, if you're worried

[02:18:49] all that anxiety of succession planning,

[02:18:50] then what should you do about it? Then

[02:18:52] you break it all the way back down. The

[02:18:54] most important thing you should do today

[02:18:56] if you care about the future of your

[02:18:57] company post you is to pass on

[02:19:01] knowledge, information, insight, skills,

[02:19:05] experience as often and continuously as

[02:19:07] you can. which is the reason why I

[02:19:09] continuously reason about everything in

[02:19:11] front of my team. Every single meeting

[02:19:14] is about a reasoning meeting. Every

[02:19:17] moment I spend inside a company, outside

[02:19:19] the company is about passing on

[02:19:21] knowledge to people as fast as I can.

[02:19:23] Nothing I learn ever sits on my desk

[02:19:27] longer than, you know, a fraction of a

[02:19:29] second. I'm passing that information,

[02:19:31] that know. Oh my gosh, this is cool.

[02:19:33] Before I even finish learning all of it

[02:19:35] myself, I've already pointing it to

[02:19:37] somebody else. get on this. This is so

[02:19:38] cool. You're going to want to you're

[02:19:40] going to want to learn this. And so I'm

[02:19:42] constantly passing knowledge, empowering

[02:19:45] people, elevating the capability of

[02:19:48] everybody around me so that

[02:19:51] um the outcome that I that I seek that I

[02:19:55] hope for is that I die on the job, you

[02:19:58] know, and and hopefully I die on the job

[02:20:00] instantaneously. You

[02:20:03] and there's no long periods of

[02:20:04] suffering, you know. Well, from a fan

[02:20:07] perspective, given your your uh

[02:20:10] extremely

[02:20:13] um your enormous positive impact on on

[02:20:16] civilization, of course, I hope you keep

[02:20:18] going, but also it's just fun to watch

[02:20:20] what is doing. You're, you know, it's

[02:20:23] just the rate of innovation and I'm a

[02:20:25] huge fan of engineering. It's so much

[02:20:27] incredible engineering is continuously

[02:20:29] being done by Nvidia. It's just fun to

[02:20:31] watch. It's a celebration of humanity.

[02:20:33] is a celebration of great builders, a

[02:20:35] celebration of great engineering. So it

[02:20:37] represents something special. Uh so I

[02:20:40] hope uh you and Nvidia keep going. What

[02:20:42] gives you hope about this whole thing we

[02:20:45] got going on about humanity? About the

[02:20:47] future of humanity when you look out and

[02:20:49] you think about the future quite a bit

[02:20:51] when you look out 10, 20, 50, 100 years

[02:20:53] from now, what gives you hope? I I've

[02:20:56] always had I've always had uh uh great

[02:20:59] confidence in in the in the kindness

[02:21:06] uh the generosity

[02:21:09] uh

[02:21:10] um the compassion, the human capacity.

[02:21:14] I've always been extremely confident of

[02:21:18] that. sometimes um

[02:21:23] more so than I should and and I I get

[02:21:26] taken advantage of. But it doesn't it

[02:21:28] doesn't ever cause me not to. I start

[02:21:32] with always

[02:21:34] uh that that people want want to do

[02:21:36] good. People want to um uh help others

[02:21:41] and

[02:21:43] uh vastly I am proven right,

[02:21:48] constantly proven right and and often

[02:21:53] uh exceeds my expectations

[02:21:56] and and so I have complete confidence in

[02:21:59] the human capacity.

[02:22:01] I think the the the thing that the

[02:22:04] things that give me incredible hope

[02:22:07] is what I see as as I extrapolate as I

[02:22:11] what I see now is possible and as I

[02:22:15] extrapolate

[02:22:16] um based on the things that we're doing

[02:22:18] what will very likely happen

[02:22:22] >> and and um and that there's so many

[02:22:25] things that we want to solve there's so

[02:22:28] many problems we want to solve there's

[02:22:29] so any things that we want to build.

[02:22:32] There's so many good things that we want

[02:22:34] to do that are now within our reach and

[02:22:37] within the reach of my my lifetime. You

[02:22:40] just can't possibly not be romantic

[02:22:44] about that. You know what I'm saying?

[02:22:46] >> Yeah. What an exciting time to be alive.

[02:22:48] >> Yeah.

[02:22:48] >> Like truly truly. So

[02:22:50] >> how can you not be romantic about about

[02:22:52] about that? the the the fact that that

[02:22:56] there is a there it's a reasonable thing

[02:22:59] to expect the end of disease. It's a

[02:23:02] reasonable thing to expect. It's a

[02:23:04] reasonable thing to expect that

[02:23:06] pollution will be drastically reduced.

[02:23:09] It's a reasonable thing to expect that

[02:23:12] traveling at the speed of light is

[02:23:15] actually in our future. And then you

[02:23:17] know not not for long distances but

[02:23:20] short distances you know you people ask

[02:23:22] me how you well first of all very soon

[02:23:24] I'm going to put a humanoid on a

[02:23:26] spaceship and it's going to be you know

[02:23:28] my humanoid and and we're going to send

[02:23:30] it out as soon you know as soon as

[02:23:32] possible and it's going to keep

[02:23:34] improving and enhancing along the flight

[02:23:36] >> and then when it's time

[02:23:39] all of the all of my consciousness has

[02:23:41] already been you know so much of my life

[02:23:43] has been uploaded in the internet take

[02:23:45] all my inbox take everything that I've

[02:23:46] done, everything I've said, you know,

[02:23:48] it's been collect and be becoming my AI

[02:23:51] and um I'm just, you know, when the time

[02:23:53] comes, you know, we just send that at

[02:23:55] the speed of light, catch up with my

[02:23:56] robot.

[02:23:59] Oh, that's brilliant. I mean, but for

[02:24:02] me, that's sort of application focused,

[02:24:04] but also for me the curiosity

[02:24:07] uh maxing perspective, I just all of

[02:24:10] those mysteries. It's so much

[02:24:12] fascinating scientific questions there.

[02:24:14] Understanding the biological machine is

[02:24:16] is right around the corner. It's it's

[02:24:18] not 10 years. It's 5 years probably.

[02:24:20] >> And then your biological machine, the

[02:24:22] the human mind and cracking physics,

[02:24:24] theoretical physics open. It's so

[02:24:25] exciting.

[02:24:26] >> Explaining consciousness, that one would

[02:24:28] be awesome

[02:24:29] >> and it's all within our reach.

[02:24:30] >> Yeah.

[02:24:31] >> Uh Jensen, thank you so much for

[02:24:32] everything you've done over the years.

[02:24:34] Thank you for everything you're doing

[02:24:35] for the world. Thank you for being who

[02:24:37] you are. Uh, I can tell you're a great

[02:24:40] human being and uh, I wish you

[02:24:44] incredible success this year. I can't

[02:24:46] wait as a fan. I can't wait to see what

[02:24:48] you do next and hopefully I'll see you

[02:24:49] in Taiwan. And thank you so much for

[02:24:51] talking today.

[02:24:52] >> Thank you, Lex. I had a great time and

[02:24:54] and also if I could just say one more

[02:24:56] thing

[02:24:56] >> and thank you for all the interviews

[02:24:58] that you do, the depth, the the respect

[02:25:02] that you go through with and the

[02:25:04] research that you do uh to reveal, you

[02:25:07] know, for all of us, uh the the amazing

[02:25:10] people that you've interviewed over the

[02:25:12] years. I've enjoyed I I've enjoyed them

[02:25:14] immensely and and and as an innovator to

[02:25:18] have created this long form unbelievable

[02:25:22] and and yet you know it's just

[02:25:24] captivating. So anyways, thank you for

[02:25:25] everything you do.

[02:25:26] >> It means the world. Thank you, Jess.

[02:25:27] >> Thank you, Lex. Thank you for listening

[02:25:30] to this conversation with Jensen Kuang.

[02:25:32] To support this podcast, please check

[02:25:34] out our sponsors in the description

[02:25:36] where you can also find links to contact

[02:25:38] me, ask questions, give feedback, and so

[02:25:41] on. And now let me leave you with some

[02:25:44] words from Alan K.

[02:25:46] The best way to predict the future is to

[02:25:49] invent it.

[02:25:51] Thank you for listening and hope to see

[02:25:53] you next time.