黄仁勋：4 万亿美元的英伟达与 AI 革命

Lex Fridman

下面是和英伟达 CEO 黄仁勋的对话。英伟达是驱动 AI 革命的引擎，很大一部分成功可以直接归因于他作为领导者、工程师和创新者的意志和那些下注。这是 Lex Fridman Podcast。亲爱的朋友们，黄仁勋。

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. 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. This is Lex Fridman Podcast. And now dear friends, here’s Jensen Huang.

Lex Fridman

你把英伟达推进到一个新阶段：从芯片级设计走到机架级。很长时间里赢意味着做出最好的 GPU，你们现在仍然做。但你把它扩成 GPU、CPU、内存、网络、存储、供电、冷却、软件、机架本身、你们宣布的 pod、甚至数据中心的极限共设计。这么多复杂组件，共设计最难的部分是什么？

You’ve propelled NVIDIA into a new era in AI, moving beyond its focus on chip-scale design to now rack-scale design. Winning for NVIDIA for a long time used to be about building the best GPU possible, 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 center. What is the hardest part of co-designing a system with that many complex components and design variables?

Jensen Huang

极限共设计之所以必要，是因为问题已经装不进一台计算机、一块 GPU 去加速。你想比你加进去的计算机数量更快：加了 1 万台计算机，你希望快 100 万倍。于是你得把算法切开、重构，把管线、数据、模型都切分。一旦这样分布式，而不只是把问题放大，所有东西都会挡路。

The reason why extreme co-design is necessary is because the problem no longer fits inside one computer to be accelerated by one GPU. You would like to go faster than the number of computers that you add. You added 10,000 computers, but you would like it to go a million times faster. Then you have to take the algorithm, break it up, refactor it, shard the pipeline, shard the data, shard the model. When you distribute the problem this way, not just scaling up, everything gets in the way.

这就是 Amdahl 定律：加速取决于它占总工作量多少。如果计算只占 50%，你把计算加快一百万倍，总工作量也只快两倍。于是你不但要分布式计算，还要切管线，还要解决网络——所有这些计算机连在一起。这个规模的分布式计算，CPU 是问题，GPU 是问题，网络是问题，交换是问题，把工作负载铺开也是问题。这是一个极其复杂的计算机科学问题。否则我们只能线性扩展，或按摩尔定律的能力扩展——Dennard scaling 已经大幅放慢。

This is the Amdahl’s law problem. If computation represents 50% of the problem, and I sped up computation a million times, I only sped up the total workload by a factor of two. Now you also have to solve the networking problem because you’ve got all of these computers connected together. 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 is a problem. It’s just a massively complex computer science problem. Otherwise we scale up linearly or based on Moore’s Law, which has largely slowed because Dennard scaling has slowed.

Jensen Huang

极限共设计是在从架构到芯片、系统、系统软件、算法、应用的整栈上优化。然后还要越过 CPU、GPU、网络芯片、scale-up 和 scale-out 交换机，把供电和冷却都算进去，因为这些计算机极度吃电。公司的架构该反映它要产出什么、它存在的环境。很多公司的组织图看起来都一样，汉堡组织图、车公司组织图，对我没意义。

Extreme co-design is optimizing across the entire stack of software from architectures to chips, to systems, to system software, to the algorithms, to the applications. Then beyond CPUs and GPUs and networking chips and scale-up switches and scale-out switches, you gotta include power and cooling. When you’re designing a company, you should first think about what you want the company to produce. I see a lot of companies’ organization charts, and they all look the same. Hamburger organization charts, car company organization charts. They all look the same. It doesn’t make any sense to me.

我的直接下属是 60 人。我跟他们不做一对一会，因为不可能。你要干活，不能有 60 人还一对一。几乎所有人都有一只脚在工程里：内存专家、CPU 专家、光学、GPU、架构、算法、设计。

My direct staff is 60 people. I don’t have one-on-ones with them because it’s impossible. You can’t have 60 people on your staff if you’re gonna get work done. Almost all of them have a foot in engineering. There’s experts in memory, there’s experts in CPUs, there’s experts in optical, GPUs, architecture, algorithms, design.

Lex Fridman

这需要你同时是工程师、管理者、管理者的管理者，还要当心理学家。

That requires you to be an engineer, a manager, a manager of managers, and a psychologist.

Jensen Huang

是啊。我们做极限共设计，是因为没有别的路。它很难，但别无他法。过去十年摩尔定律大约把计算推进了 100 倍。我们靠共设计把计算放大了大约 100 万倍。

Yeah. We do extreme co-design because we have no choice. It’s hard, but there’s no other way. In the last 10 years, Moore’s Law would have progressed computing about 100 times. We scaled computing about a million times.

Lex Fridman

你很长时间都相信广义的缩放律。现在还信吗？

You’ve been a believer in scaling laws, broadly defined, for a long time. Are you still a believer?

Jensen Huang

信。现在缩放律更多了。

Yeah. We have more scaling laws now.

Lex Fridman

你讲过四条：预训练、后训练、测试时、智能体缩放。往前看，最让你夜里睡不着、必须跨过去才能继续缩放的障碍是什么？

You’ve outlined four of them: pre-training, post-training, test time, and agentic scaling. Looking forward, what blockers keep you up at night that you have to overcome to keep scaling?

Jensen Huang

可以回头看看大家以为的障碍。一开始是预训练缩放律：高质量数据会限制我们能达到的智能。模型越大、数据相应越多，AI 就越聪明。Ilya Sutskever 说「数据用完了」或「预训练结束了」之类的话，行业慌了，以为这是 AI 的尽头。当然不是。

We can reflect on what people thought were blockers. First, pre-training: people thought, rightfully, that high-quality data would limit the intelligence we achieve. The larger the model, the correspondingly more data, the smarter the AI. Ilya Sutskever said “we’re out of data” or “pre-training is over.” The industry panicked that this was the end of AI. Of course that’s not true.

用来训练的数据会继续涨，很多大概是合成的。这把人搞糊涂了。大家忘了：我们用来教彼此的大多数数据本来就是合成的——不是从自然里长出来的，是人创造的。我消费、修改、增强、再生成，别人再消费。AI 已经能拿着 ground truth 去增强，合成海量数据。

We’re going to keep scaling the data we train with. A lot of it will be synthetic, and that confused people. They’ve forgotten that most of the data we teach each other with is synthetic. It didn’t come out of nature. You created it. I consume it, modify it, augment it, regenerate it; somebody else consumes it. AI can now take ground truth, enhance it, and synthetically generate an enormous amount of data.

后训练会继续缩放。人类生成的数据占比会越来越小。训练不再被数据限制，而是被算力限制——因为大多数数据是合成的。下一阶段是测试时。我还记得有人说：推理很容易，预训练才难。推理芯片会是小小的芯片，不像英伟达那些又复杂又贵的。他们说未来推理是最大市场、很容易、会商品化，谁都能自己做芯片。这对我一直不合逻辑：推理就是思考，思考很难，比阅读难得多。

Post-training continues to scale. Human-generated data will be a smaller and smaller share. Training is no longer limited by data; it’s limited by compute, because most of the data is synthetic. Then test time. I still remember people saying inference is easy, pre-training is hard. Inference chips would be little tiny chips, not like NVIDIA’s complicated expensive ones. Inference would be the biggest market, easy, commoditized; everybody could build their own chips. That was always illogical to me, because inference is thinking, and thinking is hard — way harder than reading.

预训练是记忆和泛化，找模式。测试时缩放是思考、推理、规划、搜索。怎么可能算力很轻？我们完全说对了：测试时缩放极度吃算力。再往后呢？我们已经做出一个智能体，测试时它去研究、敲数据库、用工具，最重要的一件事是派生出一堆子智能体——等于在组建大团队。给英伟达加人，比把我自己放大容易得多。下一条缩放律就是智能体缩放：把 AI 乘起来，你想派生多少智能体就派生多少。

Pre-training is memorization and generalization, looking for patterns. Test-time scaling is thinking, reasoning, planning, search. How could that be compute-light? We were absolutely right: test-time scaling is intensely compute-intensive. What’s beyond that? We’ve created one agentic person. During test time that system does research, bangs on databases, uses tools, and one of the most important things it does is spawn sub-agents — creating large teams. It’s so much easier to scale NVIDIA by hiring more employees than to scale myself. The next scaling law is agentic scaling: multiplying AI. We can spin off agents as fast as we want.

智能体系统会制造大量数据和经验。有些我们会说「这真好，该记住」。这套数据回到预训练去记忆和泛化，再进后训练微调，再用测试时增强，智能体系统再放到行业里。这个循环会一直转。智能最终靠一件事缩放：算力。

As we use agentic systems they’ll create a lot more data and experiences. Some of it we’ll say, “this is really good, we ought to memorize this.” That data set comes all the way back to pre-training. We memorize and generalize it, refine it in post-training, enhance it with test time, and agentic systems put it out to the industry. This loop is going to go on and on. Intelligence is going to scale by one thing: compute.

Lex Fridman

棘手的是你得提前预判：有些组件要用不同硬件才最优。你得预判 AI 创新会往哪走。

There’s a tricky thing: some of these components require different hardware to do it optimally. You have to anticipate where the AI innovation is going to lead.

Lex Fridman

我们有很长一段「以为会卡、后来跨过去了」的历史。现在智能体马上到处都是，显然需要算力。接下来卡缩放的会是什么？

We have a long history of blockers we thought were going to be blockers, and we overcame them. Now that it’s clear agents will be everywhere, we’re going to need compute. What is going to be the blocker for that scaling?

Jensen Huang

电是担心，但不是唯一担心。所以我们才拼命做极限共设计，每年把每瓦每秒 token 数提高几个数量级。过去十年摩尔定律大约 100 倍，我们把计算放大了 100 万倍。能效、perf per watt 直接决定一家公司、一座工厂的收入。我们要把这个推到极限，尽快把 token 成本打下来。

Power is a concern, but it’s not the only concern. That’s why we’re pushing so hard on extreme co-design, so we can improve tokens per second per watt orders of magnitude every year. In the last 10 years Moore’s Law would have progressed computing about 100 times. We scaled computing a million times. Energy efficiency, perf per watt, completely affects the revenues of a company and a factory. We’re going to push that to the limit so we can keep driving token costs down as fast as we can.

我们计算机的价格在涨，但 token 生成效率涨得更快，token 成本就是在降，大约每年一个数量级。

Our computer price is going up, but our token generation effectiveness is going up so much faster that token cost is coming down — an order of magnitude every year.

Lex Fridman

绕过电这个卡点的办法，是把每瓦每秒 token 做得更高效。当然还有怎么拿到更多电。

The way to get around the power blocker is tokens per second per watt — make it more efficient. Of course there’s also how we get more power.

Jensen Huang

我们也该拿到更多电。

We should also get more power.

Lex Fridman

这很复杂。你讲过小型模块化核电。AI 供应链的瓶颈——ASML 的 EUV、TSMC 的先进封装比如 CoWoS、SK 海力士的 HBM——让你夜里睡不着到什么程度？

That’s a really complicated one. You’ve talked about small modular nuclear power plants. How much do bottlenecks in the AI supply chain keep you up at night — ASML with EUV, TSMC with advanced packaging like CoWoS, SK Hynix with high-bandwidth memory?

Jensen Huang

一直在想，一直在做。历史上没有哪家公司以我们这个规模增长，同时还在加速增长。在整个 AI 计算世界里我们还在提高份额。上游下游供应链对我们极其重要。我花大量时间告诉一起干活的 CEO：什么动力会让增长继续甚至加速。主题演讲右侧几乎是整个 IT 上游和基础设施下游的 CEO，好几百个。我不觉得以前有哪个 keynote 会来好几百个 CEO。我告诉他们现在的业务、近未来的增长驱动、下一步往哪走，好让他们决定怎么投。我像告诉自己员工那样告诉他们。

All the time, and we’re working on it all the time. No company in history has ever grown at a scale that we’re growing while accelerating that growth. In the overall world of AI computing, we’re increasing share. Supply chain, upstream and downstream, are really important. I spend a lot of time informing all the CEOs I work with: what dynamics will cause growth to continue or even accelerate. To the entire right-hand side of me were CEOs of practically the entire IT industry upstream and the infrastructure industry downstream. Several hundred CEOs. I don’t think there’s ever been keynotes where several hundred CEOs show up. I’m telling them our business condition now, the growth drivers in the very near future, and where we’re going next so they can inform how they invest. I inform them that way like I inform my own employees.

Jensen Huang

大约三年前我去说服几位 DRAM CEO：当时 HBM 几乎只给超算用，用量很少，但这会成为未来数据中心的主流内存。起初听起来荒唐，几位 CEO 信了，决定投建 HBM。另一种放进数据中心很怪的内存，是手机用的低功耗内存。我们想把它改造成数据中心里的超算内存。他们说「手机内存给超算？」我解释为什么。看看 LPDDR5、HBM4，出货量惊人。三家都创了历史纪录年，而这些是做了 45 年的公司。我的工作一部分就是告知、塑造、激励。

About three years ago I was able to convince several DRAM CEOs that even though HBM was used quite scarcely, barely by supercomputers, this was going to be a mainstream memory for data centers. At first it sounded ridiculous, but several believed me and decided to invest in building HBM. Another memory was rather odd to put in a data center: the low-power memories we use for cell phones. We wanted them adapted for supercomputers. They go, “cell phone memory for supercomputers?” I explained why. Look at LPDDR5, HBM4. The volumes are incredible. All three of them had record years in history, and these are 45-year companies. That’s part of my job: inform, shape, inspire.

Lex Fridman

所以你不只是在英伟达内部显化未来，你在显化未来的供应链。你跟 TSMC、ASML 谈话。

You’re not just manifesting the future inside NVIDIA. You’re manifesting the supply chain of the future. Conversations with TSMC, with ASML.

Jensen Huang

上游，下游。GEV、Caterpillar。那是我们的下游。Vera Rubin 机架大约 130 万个零件，也可以说一百五十万。大约 200 家供应商。

Upstream, downstream. GEV, Caterpillar. That’s downstream from us. Each rack is 1.3, one and a half million components. There are 200 suppliers across the Vera Rubin rack.

系统架构从你记得的 DGX-1 变成 NVLink-72 机架级计算，意味着软件、工程、设计测试、供应链都变了。其中一件事是：我们把数据中心里的超算集成，挪到了供应链里的超算制造。如果你想同时跑 50 吉瓦超算，而制造这 50 吉瓦只要一周，那供应链里每周都需要一吉瓦电来建造和测试，然后我才发货。

We changed the system architecture from the original DGX-1 you remembered to NVLink-72 rack-scale computing. What does that mean to software, engineering, how we design and test, the supply chain? We moved supercomputer integration at the data center into supercomputer manufacturing in the supply chain. If you’d like 50 gigawatts of supercomputers running simultaneously, and it takes one week to manufacture that 50 gigawatts, then each week in the supply chain the supercomputers are going to need a gigawatt of power to build and test before I ship.

Jensen Huang

我想把这句话讲出去：电网是按最坏情况再加一点余量设计的。最坏情况是冬天几天、夏天几天、极端天气。99% 的时间我们根本不在最坏情况，大概跑在峰值的 60% 左右。所以 99% 的时间电网有闲置电力，它们必须空着待命，因为万一医院、基础设施、机场要电。问题是：能不能签合同、把计算机架构和数据中心设计成——社会基础设施要最大电时，数据中心少拿一点。

Our power grid is designed for the worst-case condition with some margin. The worst case is a few days in winter, a few days in summer, extreme weather. 99% of the time we’re nowhere near that, probably running around 60% of peak. So 99% of the time the grid has excess power sitting idle, because hospitals, infrastructure, airports have to be powered just in case. The question is whether we can create contractual agreements and design computer architecture and data centers such that when they need maximum power for infrastructure, the data centers would get less.

那种时刻本来就很少。那时要么用一点备用发电，要么把负载挪走，要么计算机跑慢一点：降性能、降功耗，回答延迟稍长一点。现在这些合同要求 100% 在线，给电网很大压力，逼他们从已经是最大的能力再往上加。我只想用他们空着的那部分。

That’s a rare instance anyway. During that time we either have a backup generator for that little part, or we shift the workload somewhere else, or we just run slower — degrade performance, reduce power, slightly longer latency. Instead of expecting 100% uptime, these really rigorous contracts put a lot of pressure on the grid to increase from their maximum. I just want to use their excess. It’s just sitting there.

Lex Fridman

拦着的是监管还是官僚？

What’s stopping there? Is it regulation? Is it bureaucracy?

Jensen Huang

三方问题。先从终端客户：他们要求数据中心永远不能不可用，期望完美。要交付完美，就要备用发电加电网都完美，人人都要六个九。我打赌 CEO 不知道这些合同。合同谈判两边都要最好条款，CSP 再去找公用事业要六个九。第一件事是让客户和 CEO 明白自己在要什么。第二，数据中心必须能优雅降级：电网说退到大约 80%，我们说没问题，挪负载、数据不丢、降计算速率、少用电。关键负载立刻挪到仍有 100% 的地方。第三，公用事业要意识到这是机会：别说「扩电网要五年」；如果你愿意接受这种保证等级的电，我下个月就能给你、按这个价。电网现在浪费太多，该去拿。

It’s a three-way problem. It starts with the end customer: they put requirements that data centers can never not be available. They expect perfection. To deliver that you need backup generators and the grid supplier to deliver perfection. Everybody’s gotta have six nines. I bet the CEO doesn’t know this. Contract negotiators on both sides want the best contract; CSPs then expect six nines from utilities. First, make sure CEOs and customers realize what they’re asking for. Second, we have to build data centers that gracefully degrade. If the grid says we’re backing you down to about 80%, that’s no problem: move the workload, data’s never lost, reduce the computing rate, use less energy. Critical workloads I shift somewhere that still has 100% uptime. Third, utilities need to recognize this as an opportunity: instead of “it’ll take five years to increase my grid,” if you’re willing to take power of this level of guarantee, I can make it available next month at this price. There’s way too much waste in the grid. We should go after it.

Lex Fridman

数据中心里这种智能、动态的电力分配，工程上有多难？

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

Jensen Huang

一旦能规格化，就能工程化。只要服从物理定律和第一性原理，我觉得没问题。

As soon as you could specify, you could engineer it. So long as it obeys the laws of physics on first principles, I think we’re good.

Lex Fridman

你高度赞扬 Elon 和 xAI 在孟菲斯建成 Colossus。大概四个月，现在大约 20 万 GPU，还在很快增长。他的工程和管理方式，有什么值得所有建数据中心的人学？

You’ve highly lauded Elon and xAI’s accomplishment in Memphis, building Colossus in just four months. It’s now at 200,000 GPUs and growing very quickly. Is there something about his approach that’s instructive to data center creators — engineering, construction, management?

Jensen Huang

首先 Elon 在很多题目上都很深，同时又是很好的系统思考者，能跨学科推。他质疑一切：一，有必要吗？二，必须这样做吗？三，必须这么久吗？他把东西砍到不能再少、必要能力还在。他是你能想象的最极简，而且是系统尺度上的极简。我也喜欢他出现在行动点上：有问题他就去，「给我看问题」。这些合在一起，能压过很多「我们一直这么干」「我在等他们」。每个人都有借口。最后：你自己带着那么大的紧迫感行动，会让别人也紧迫。每个供应商都有很多客户、很多项目，他让自己成为别人项目的最高优先级，靠示范。

First, Elon is deep in so many topics, yet he’s a really good systems thinker, able to think through multiple disciplines. He questions everything: one, is it necessary? Two, does it have to be done this way? Three, does it have to take this long? Everything down to its minimal amount that’s necessary, and the necessary capabilities remain. He is as minimalist as you could imagine, at a system scale. I also love that he is present at the point of action. If there’s a problem, he’ll go there: “show me the problem.” When you do all of this in combination, you overcome a lot of “this is just the way we do it,” “I’m waiting for them.” Everybody has a lot of excuses. And when you act personally with so much urgency, it causes everybody else to act with urgency. Every supplier has a lot of customers and projects. He makes it his business that he’s the top priority of everybody else’s projects, by demonstrating it.

Lex Fridman

我参加过那些会。不够人问「能不能快很多，怎么快，为什么必须这么久」。这常常变成工程问题。有一次他在过整根电缆怎么插进机架，跟现场工程师一起，想搞清楚流程，让它更少出错。从每一个任务建立直觉，细节尺度和系统尺度上的低效都会露出来。再加上那把大锤：「咱们彻底换一种做法，拿掉所有可能的拦路石。」

I’ve been in a bunch of those meetings. Not enough people ask, “can this be done a lot faster, and how? Why does it have to take this long?” That becomes an engineering question. One of the times I was hanging out with him, he was going through the entire process of how to plug cables into a rack, working with an engineer on the ground, trying to understand the process so it can be less error-prone. You build intuition from every single task. You get a sense at the detailed scale and the systems scale of where the inefficiencies are. Plus you have the big hammer of “let’s do it totally different and remove all possible blockers.”

Lex Fridman

英伟达的极限系统共设计，和 Elon 做系统工程，有平行之处吗？

Are there parallels in NVIDIA’s extreme systems co-design and the way Elon approaches systems engineering?

Jensen Huang

共设计就是终极系统工程问题，我们从第一性原理出发。另一件事是我 30 年前开始的一种心态，叫「光速」。光速不只是快，是物理能做到的极限的简称。我们做的每件事都拿去跟光速比：内存速度、数学速度、功耗、成本、时间、人力、制造周期。

Co-design is an ultimate systems engineering problem. We approach the work from that first principle. The other thing is a method I started 30 years ago, called the speed of light. The speed of light is not just about speed. It’s my shorthand for what’s the limit of what physics can do. Every single thing we do is compared against the speed of light: memory speed, math speed, power, cost, time, effort, number of people, manufacturing cycle time.

低延迟系统和高吞吐系统架构根本不同，你要知道各自的光速，再在总系统里做权衡。我逼每个人在动手前先想第一性原理、物理极限。我不喜欢「持续改进」那种进场方式。有人说「现在要 74 天，我们可以给你做到 72 天」。我宁可拆回零：「先解释为什么一开始是 74 天。如果完全从零建，今天可能多久？」常常会惊讶，也许是 6 天。剩下那 74 天里很多可以是有理由的妥协和降本，但至少你知道它们是什么。知道 6 天可能，从 74 谈到 6，出奇地有效。

A system that achieves extremely low latency versus one that achieves very high throughput are architected fundamentally differently. You want to know the speed of light of each, then make trade-offs in the total system. I force everybody to think about first principles, the physical limits, before we do anything. I don’t love continuous improvement as a way into a problem. Somebody says it takes 74 days today and we can do it in 72. I’d rather strip it all back to zero: explain why 74 days in the first place. If I were to build it completely from scratch, how long would it take? Oftentimes you’d be surprised. It might come to six days. The rest of the 74 could be well-reasoned compromises and cost reductions, but at least you know what they are. Now that you know six days is possible, the conversation from 74 to six is surprisingly much more effective.

Lex Fridman

Vera Rubin pod：七种芯片、五种专用机架、40 个机架、1.2 千万亿晶体管、近 2 万颗英伟达 die、1100 多颗 Rubin GPU、60 exaflops、每秒 10 PB 的 scale 带宽。这只是一个 pod。NVL72 机架单独就 130 万零件。这么复杂，简洁还是启发式吗？

The Vera Rubin pod: seven chip types, five purpose-built rack types, 40 racks, 1.2 quadrillion transistors, nearly 20,000 NVIDIA dies, over 1,100 Rubin GPUs, 60 exaflops, 10 petabytes per second of scale bandwidth. That’s just one pod. The NVL72 rack alone is 1.3 million components. In such complexity, is simplicity sometimes a good heuristic?

Jensen Huang

那只是一个 pod。换个角度看，我们大概每周要交出大约 200 个这样的 pod。

That’s just one pod. We’re probably gonna have to crank out about 200 of these pods a week, just to put it in perspective.

Jensen Huang

先讲事实。大约 50% 的全球 AI 研究者是中国人，加减一点，而且大多数还在中国。我们这里也有很多，但中国仍有出色的研究者。他们的科技产业正好在移动云时代出现，贡献方式是软件。这是一个科学和数学教育很好的国家。中国不是一个铁板一块的经济体，很多省、市、市长互相竞争。所以有那么多 EV 公司、那么多 AI 公司、你能想到的每种公司都有人做。内部竞争疯狂，留下来的是非常强的公司。

Let’s start with some facts. 50% of the world’s AI researchers are Chinese, plus or minus, and they’re mostly in China still. We have many of them here, but there are amazing researchers still in China. Their tech industry showed up at precisely the right time, the mobile cloud era; their way of contributing was software. This is a country with kids incredibly well educated in science and math. China is not one giant economic country. It’s got many provinces and cities with mayors all competing with each other. That’s why there are so many EV companies, so many AI companies, every company you could imagine. They have insane competition internally. What remains is an incredible company.

他们的社会文化是家庭第一、朋友第二、公司第三。来回交流很多，本质上一直在开源。他们给开源贡献更多很合理：「我们在保护什么？」工程师的兄弟在那家公司，朋友在那家公司，都是同学。同学概念几乎是一辈子的兄弟。知识传得很快，没必要把技术藏着，不如开源。开源社区再放大、加速创新。顶尖人才、开源带来的快速创新、朋友关系、疯狂竞争——合在一起，这是今天世界上创新最快的国家。这些都根植于孩子怎么长大、教育、父母要他们学业好。他们正好赶上技术指数增长的时刻。

They have a social culture where it’s family first, friends second, company third. The amount of conversation that goes back and forth — they’re essentially open source all the time. That they contribute more to open source is so sensible. “What are we protecting?” My engineers’ brothers are in that company, their friends, they’re all schoolmates. Schoolmates, you’re brothers for life. They share knowledge very quickly. There’s no sense keeping technology hidden. You might as well put it on open source. The open source community then amplifies, accelerates innovation. Great talent, rapid innovation because of open source and the nature of friends, and insane competition. This is the fastest innovating country in the world today. Everything I just said is fundamental to how the kids were grown, excellent education, parents wanting them to do well in school. They showed up at precisely the time when technology is going through that exponential.

Lex Fridman

再加上文化上，当工程师挺酷。

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

Jensen Huang

这是一个建造者国家。我们国家的领导人很了不起，但大多是律师——要保护我们、法治。他们的国家从贫困里建起来，大多数领导人是非常强的工程师，一些最聪明的头脑。

It’s a builder nation. Our country’s leaders are incredible, but they’re mostly lawyers — trying to keep us safe, rule of law. Their country was built out of poverty. Most of their leaders are incredible engineers. Some of the brightest minds.

Lex Fridman

开源这条，你长期看好 Perplexity，也谢谢你们放出开源的 Nemotron 3 Super，1200 亿参数开权 MoE。DeepSeek、MiniMax 这些公司在推开源 AI，英伟达也在做接近前沿的开源大模型。你对开源的愿景是什么？

You have been a fan of Perplexity a long time. Thank you for releasing open source Nemotron 3 Super, a 120 billion parameter open-weight MoE model. China with DeepSeek and MiniMax is pushing open source AI, and NVIDIA is leading with close to state-of-the-art open source LLMs. What’s your vision with open source?

Jensen Huang

首先，要做伟大的 AI 计算公司，必须理解模型怎么演化。Nemotron 3 不只是纯 Transformer，是 Transformer 加 SSM。我们很早做条件 GAN、progressive GAN，一步步走到扩散。在模型架构和不同领域做基础研究，能看见未来模型需要什么样的计算系统。这是极限共设计策略的一部分。第二，一方面我们要世界级模型当产品，它们该是专有的；另一方面我们要 AI 扩散进每个行业、每个国家、每个研究者、每个学生。如果全是专有的，很难在上面做研究、创新。开源对很多行业加入 AI 革命是根本必要的。英伟达有规模、有动机，只要我们还在，就该继续做这些模型，激活每个行业、每个研究者、每个国家。

First, if we’re going to be a great AI computing company, we have to understand how AI models are evolving. One of the things I love about Nemotron 3 is it’s not just a pure transformer — transformer and SSMs. We were early in developing conditional GANs, progressive GANs, which led step by step to diffusion. Basic research in model architecture and different domains gives us visibility into what computing systems would do a good job for future models. That’s part of extreme co-design. Second, we want world-class models as products, and they should be proprietary. On the other hand we want AI to diffuse into every industry, every country, every researcher, every student. If everything is proprietary, it’s hard to do research and innovate on top of. Open source is fundamentally necessary for many industries to join the AI revolution. NVIDIA has the scale and the motives to build these AI models for as long as we shall live. We can activate every industry, every researcher, every country.

第三，AI 不只是语言。这些 AI 会用工具、模型和在其他模态上训练的子智能体：生物、化学、物理定律、流体、热力学，不都是语言结构。得有人把天气预报、生物 AI、物理 AI 推到极限。我们不造车，但要让每家车企拿到好模型。我们不发现药，但要让礼来有世界上最好的生物 AI 去做药。三件事：AI 很宽、不只语言；让所有人进来；以及 AI 的共设计。

Third, AI is not just language. These AIs will likely use tools and models and sub-agents trained on other modalities: biology, chemistry, laws of physics, fluids and thermodynamics. Not all of it is in language structure. Somebody has to make sure weather prediction, AI for biology, physical AI, can be pushed to the frontier. We don’t build cars, but we want every car company to have access to great models. We don’t discover drugs, but I want Lilly to have the world’s best biology AI systems. Three fundamental reasons: AI is really broad, not just language; we want to engage everybody; and co-design of AI.

Jensen Huang

对 TSMC 最深的误解是以为他们只有技术：有很好的晶体管，别人拿出另一个晶体管就游戏结束。技术当然重要——晶体管、金属化、封装、3D 封装、硅光，所有这些让公司特别。但他们还能编排全世界几百家公司的动态需求：上量、下量、推迟、提前、客户之间切换、紧急开晶圆、停晶圆。世界一直在变形，他们工厂还是高吞吐、高良率、很好的成本、出色的客户服务。他们把承诺当真。晶圆答应哪天到就哪天到，好让你能把公司运转起来。这套制造系统我觉得完全是奇迹。

The deepest misunderstanding about TSMC is that their technology is all they have. Somehow they have a really great transistor, and if somebody shows up with another transistor, game over. The technology — not just the transistor, the metallization, packaging, 3D packaging, silicon photonics — is really what makes the company special. But their ability to orchestrate the dynamic demands of hundreds of companies in the world as they’re increasing, decreasing, pushing out, pulling in, changing from customer to customer, emergency wafer starts, wafer stopping. The world is shape-shifting all the time, and somehow they’re running a factory with high throughput, high yields, really great costs, excellent customer service. They take their promises seriously. When the wafers were promised to show up, the wafers show up, so you can run your company. Their manufacturing system is completely miraculous.

第二是文化：一边推进技术，一边面向客户服务。很多公司很会服务客户，但不在技术刀刃上；很多在刀刃上，客户服务不是最好。他们两边都世界级。第三，我最看重他们创造的那种无形的东西：信任。我把公司压在他们上面，这是大事。

Second is their culture. Simultaneously technology-focused, advancing technology; simultaneously customer-service oriented. A lot of companies are very customer-service oriented but not at the bleeding edge. A lot are at the bleeding edge but not the best at customer service. They’ve balanced these two and they’re world-class at both. Third, the technology I most value that they created is this intangible called trust. I trust them to put my company on top of them. That’s a very big deal.

Lex Fridman

这种信任建立在多年表现上，也有人际关系。

That trust is established based on many years of performance, but there are human relationships involved as well.

Jensen Huang

三十年。我不知道我们经他们做了多少百亿、多少千亿美元生意，我们没有合同。这挺了不起。

Three decades. I don’t know how many tens, hundreds of billions of dollars of business we’ve done through them, and we don’t have a contract. That’s pretty great.

Lex Fridman

有个故事：2013 年台积电创办人张忠谋请你当 TSMC 的 CEO，你说你已经有工作了。是真的吗？

There’s this story that in 2013 the founder of TSMC, Morris Chang, offered you the chance to become TSMC’s chief executive, and you said you already had a job. Is this story true?

Jensen Huang

故事是真的。我没有轻描淡写。我深感荣幸。当时和现在我都知道 TSMC 是历史上最重要的公司之一，Morris 是我生命里最受尊敬的高管、商业和个人朋友之一。他开口，我谦卑、真的荣幸。但我在这里做的工作很重要。我脑子里看见过英伟达会成为什么、我们能有什么影响。这是我的责任，唯一的责任，让它发生。所以我拒绝了，不是因为那不是不可思议的邀请，而是我根本不能接。

Story is true. I didn’t dismiss it. I was deeply honored. I knew then as I know now, TSMC is one of the most consequential companies in history, and Morris is one of the highest regarded executives and business and personal friends I’ve had in my life. For him to ask, I was humbled and really honored. But the work I’m doing here is really important. I’ve seen in my mind’s eye what NVIDIA was going to be and the impact we could have. It was my responsibility, my sole responsibility, to make this happen. I declined it not because it wasn’t an incredible offer — it’s an unbelievable offer — but I simply couldn’t take it.

Lex Fridman

英伟达和台积电都是人类文明史上最伟大的公司之二，管任何一家都得真正 all in。

NVIDIA and TSMC are two of the greatest companies in the history of human civilization. Running either one, you have to truly be all in.

Jensen Huang

所以现在我可以帮两家公司。

So now I can help both companies.

Lex Fridman

英伟达现在是世界上最有价值的公司。最大的护城河是什么？

NVIDIA is now the most valuable company in the world. What is NVIDIA’s biggest moat — the edge that protects you from the competition?

Jensen Huang

公司最重要的财产是计算平台的安装基数。今天最重要的是 CUDA 的安装基数。二十年前当然没有安装基数。如果有人搞出个 GUDA 或 TUDA，也没什么差别，因为从来不只是技术。技术当然了不起、有远见。但公司对它投入、坚持、扩大覆盖。不是三个人做成 CUDA，是 4.3 万人做成 CUDA，加上几百万开发者相信我们会把 CUDA 从 1、2、3 做到 13，把他们那座软件山迁上来。

Our single most important property as a company is the install base of our computing platform. The single most important thing today is the install base of CUDA. 20 years ago there was no install base. If somebody came up with a GUDA or TUDA, it wouldn’t make any difference, because it’s never been just about the technology. The technology of course was incredible, visionary. But the company was dedicated to it, stuck with it, expanded its reach. It wasn’t three people that made CUDA successful. It was 43,000 people, and the several million developers that believed in us, that trusted we were going to continue to make CUDA 1, 2, 3, 13, that decided to port and dedicate their mountain of software on top of it.

安装基数是第一优势。再叠上我们这种速度、这种复杂度——历史上没人建过这么复杂的系统，一年建一代几乎不可能。开发者心里是：我支持 CUDA，平均等大约六个月就会好 10 倍。而且我在 CUDA 上开发，能触达几亿台计算机，每个云、每家计算机公司、每个行业、每个国家。开源包先上 CUDA，两种属性同时拿到。而且我 100% 相信英伟达会把 CUDA 一直维护、改进、优化库，只要他们还在。信任。合在一起，如果我今天是开发者，我会先打 CUDA，主要打 CUDA。

The install base is the number one most important advantage. Amplify it with the velocity of our execution at this scale. No company in history had ever built systems of this complexity, and to build it once a year is impossible. From the developer’s perspective, if I support CUDA, tomorrow it’ll be 10 times better. I just have to wait six months on average. If I develop it on CUDA, I reach a few hundred million computers. I’m in every cloud, every computer company, every industry, every country. If I create an open source package and put it on CUDA first, I get both attributes simultaneously. And I trust 100% that NVIDIA is going to keep CUDA around and maintain it and keep optimizing the libraries for as long as they shall live. Trust. If I were a developer today, I would target CUDA first. I would target CUDA most.

第二是生态。我们把极其复杂的系统垂直集成，又水平集成进每一家公司的计算机：Google Cloud、Amazon、Azure，AWS 现在上得很快，CoreWeave、Nscale 这类新公司，礼来的超算，企业计算机，边缘的无线基站。一种架构进所有这些系统：车、机器人、卫星、太空。生态几乎覆盖世界上每个行业。

Our second one is our ecosystem. We vertically integrated this incredibly complex system, but we integrate it horizontally into every single company’s computers. Google Cloud, Amazon, Azure. We’re ramping up AWS like crazy. New companies like CoreWeave and Nscale. Supercomputers at Lilly. Enterprise computers. The edge in radio base stations. One architecture is in all these different systems. Cars, robots, satellites, space. The ecosystem is so broad it basically covers every single industry in the world.

Lex Fridman

CUDA 安装基数未来会演变成 AI 工厂这种护城河吗？未来的英伟达会不会全是 AI 工厂？

How does the CUDA install base evolve into the future with AI factories as a moat? Do you think NVIDIA of the future is all about the AI factory?

Jensen Huang

计算单元对我们曾经是 GPU，然后是一台计算机，然后是集群，现在是整座 AI 工厂。以前我脑子里看见芯片。宣布新一代，「女士们先生们，今天我们宣布 Ampere」，我把芯片举起来。那是我的心智模型。今天把芯片举起来还是挺可爱的，但它不是我在做的事的心智模型。我的心智模型是这个吉瓦级的大家伙：发电连着电网，冷却系统，网络是怪物，里面一万人在装，几百个网络工程师，后面几千人在上电。给一座工厂上电不是有人说「开了」，要几千人。

The unit of computing used to be GPU to us. Then it became a computer, then a cluster. Now it’s an entire AI factory. In the old days I visualized the chip. When I announced a new generation, “ladies and gentlemen, we’re announcing Ampere today,” I’d pick up the chip. That was my mental model. Today, picking up the chip is still adorable. It’s adorable. It’s not my mental model of what I’m doing. My mental model is this giant gigawatt thing that has power generations connected to the grid, cooling systems, networking of incredible monstrosity. 10,000 people in there trying to install it, hundreds of networking engineers, thousands of engineers behind it trying to power it up. Powering up one of those factories is not somebody going “it’s on now.” It takes thousands of people.

整套基础设施。我希望下一次点击是行星尺度。那会是下一跳。

Entire infrastructure. I’m hoping my next click is when I’m thinking about building computers, it’s planetary scale. That’ll be the next click.

Lex Fridman

Elon 讲过在太空做计算，让扩能源容易一些。你怎么看太空这条线？

What do you think about the space angle Elon has talked about, doing compute in space for solving some of the energy issues in scaling?

Jensen Huang

冷却不容易。我们已经在那儿了。英伟达 GPU 是太空里的第一批 GPU。我都没意识到，不然也许会宣布：我们在太空了，给 GPU 穿个小小航天服。太空很适合做大量成像。卫星分辨率很高，现在持续扫地球。你想要厘米级、持续的全球成像，等于全世界的实时遥测。你不想把那东西打回地球，那是一个又一个 PB。得在边缘做 AI，把你不需要的、见过的、没变的扔掉，只留需要的。所以 AI 必须在边缘做。放在极地可以 24/7 太阳能。但没有传导、没有对流，几乎只有辐射。太空很大，大概会把巨大的散热器摆出去。

Cooling issues is not easy. We’re already there. NVIDIA GPUs are the first GPUs in space. I didn’t realize it. I would have declared it maybe. We’re in space. Little astronaut suit on one of our GPUs. We’ve been in space. It’s the right place to do a lot of imaging. Those satellites have really high-resolution imaging systems, sweeping the Earth continuously. You want centimeter-scale imaging done continuously for the world, real-time telemetry of everything. You don’t want to beam that back down to Earth. Petabytes and petabytes of data. You gotta do AI right there at the edge, throw away everything you don’t need, you’ve seen before, didn’t change, keep the stuff you need. AI had to be done at the edge. Obviously we have 24/7 solar if we put it at the polars. But there’s no conduction, no convection. You’re pretty much just radiation. Space is big. I guess we’re just gonna put big, giant radiators out there.

Lex Fridman

这有多疯狂？五年、十年、二十年？我们在谈缩放的卡点。

How crazy of an idea do you think it is? Five years out, 10, 20? We’re talking about blockers for AI scaling.

Jensen Huang

我实际得多。我先找下一桶机会。同时我在培养太空：派工程师去做，学辐射怎么处理、性能怎么退化、缺陷怎么持续测试和 attestation、冗余怎么做、怎么优雅降级。软件怎么想冗余和性能。让计算机永不坏，只是变慢。可以先做大量工程探索。与此同时我最喜欢的答案是消灭浪费。地上那些闲置电力，我想尽快抽干。

I’m just so much more practical. I look for where my next bucket of opportunities are first. Meanwhile I’m cultivating space. I send engineers to go work on the problem. We’re learning a lot: how do we deal with radiation, degrading performance, continuous testing and attestation of defects, redundancy, degrade gracefully. What about software — redundancy and performance out in space? Make it so the computer never breaks, it just gets slower. We could start doing a lot of engineering exploration upfront. In the meantime, my favorite answer is eliminate waste. We’ve got all that idle power. I want to evacuate it as fast as possible.

Lex Fridman

你觉得英伟达有朝一日会值 10 万亿吗？换个问法：什么样的未来世界会让这成为真的？

Do you think NVIDIA may be worth 10 trillion at some point? Let’s ask it this way. What does the future of the world look like where that’s true?

Jensen Huang

我觉得英伟达的增长极其可能，在我心里是必然。我们是历史上最大的计算机公司，这本身就该问为什么。两个基础技术原因。第一，计算从检索式、文件检索系统变成生成式。几乎所有东西都是文件：我们预先写、预先录、画好，放到网上、放进文件，用推荐系统、聪明的过滤器给你检索。那是人类预录加文件检索。现在 AI 计算机有情境意识，必须实时处理和生成 token。从检索式计算到生成式计算。新世界需要的处理比旧世界多得多。旧世界需要大量存储，新世界需要大量计算。

I think NVIDIA’s growth is extremely likely, and in my mind, inevitable. We’re the largest computer company in history. That alone should beg the question, why? Two foundational technical reasons. First, computing went from being a retrieval-based file retrieval system. Almost everything is a file. We pre-write, pre-record, draw something, put it on the web, put it in a file, use a recommender, some smart filter, to figure out what to retrieve. We were a human pre-recording and file-retrieving system. Now AI computers are contextually aware, which means they have to process and generate tokens in real time. From a retrieval-based computing system to a generative-based computing system. We’re going to need a lot more processing in this new world. We needed a lot of storage in the old world. We need a lot of computation in this new world.

这种按情境、按情境意识、在生成前用新洞察落地的计算，只有在它无效时才会退回去。做深度学习这 10、15 年，如果任何一刻我得出「这走不通、死胡同、缩不了、解不了这个模态、用不到这个应用」，我会完全不同地看这件事。但过去五年给我的信心比之前十年还多。

This computation-intensive way of generating information that’s contextually relevant, situationally aware, grounded on new insight before it generates, would only go back if it’s not effective. For the last 10, 15 years working on deep learning, if at any single moment I would have come to the conclusion this is not going to work out, a dead end, not going to scale, not going to solve this modality, not going to be used in this application, I would feel very differently. The last five years has given me more confidence than the previous ten years.

第二个想法：计算机曾经是仓储系统，很大程度上是仓库。我们现在在建工厂。仓库赚不了多少钱。工厂直接对应公司收入。计算机不但做事方式变了，它在世界上的目的变了。它不再是计算机，是工厂，用来产生收入。我们看到工厂在生产人们想消费的产品、商品，商品对那么多受众那么有价值，token 开始分层，像 iPhone：免费 token、高级 token、中间好几档。智能原来是可缩放的产品。有极高智能、专门用途的 token，人愿意付。每百万 token 一千美元，不是会不会，只是何时。

The second idea: computers, because they were a storage system, were largely a warehouse. We’re now building factories. Warehouses don’t make much money. Factories directly correlate with the company’s revenues. Not only did the computer change the way it did it, its purpose in the world changed. It’s no longer a computer, it’s a factory, used for generation of revenues. We’re seeing this factory generating products, commodities people want to consume, so interesting, so valuable to so many audiences that the tokens are starting to segment, like iPhones. Free tokens, premium tokens, several in the middle. Intelligence, as it turns out, is a scalable product. Extremely high-intelligence products, tokens used for specialized things, people will be willing to pay. Somebody willing to pay $1000 per million tokens is just around the corner. It’s not if, it’s only when.

世界需要多少这样的工厂？需要多少 token？社会愿意为这些 token 付多少？生产率大幅提高后世界经济会怎样？会发现新药、新产品、新服务吗？合在一起，我绝对确定世界 GDP 增长会加速。我绝对确定 GDP 里用于计算的比例会比过去高 100 倍——因为它不再是存储单元，是产品生产单元。再倒推英伟达在这个新经济里能吃到多少，我觉得我们会大很多。近未来英伟达有没有可能成为 3 万亿美元收入的公司？当然有。我看不到任何物理极限说 3 万亿不可能。供应链的负担由 200 家公司分担。问题是有没有能源。我们肯定会有。那个数字只是一个数字。

How many of these factories does the world need? How many tokens? How much is society willing to pay? What would happen to the world’s economy if productivity improved so substantially? Are we going to discover new drugs, new products, new services? Taken in combination, I am absolutely certain the world’s GDP is going to accelerate in growth. I am absolutely certain the percentage of that GDP used for computation will be 100 times more than the past, because it’s no longer a storage unit. It’s a product generation unit. Back into what NVIDIA does and how much of that new economics we address — I think we’re going to be a lot, lot bigger. Is it possible for NVIDIA to be a $3 trillion revenue company in the near future? Of course, yes. It’s not limited by any physical limits. There’s nothing that says $3 trillion is not possible. NVIDIA’s supply chain — the burden is shared by 200 companies. Do we have the energy? Surely we will. That number is just a number.

我还记得第一次跨过 10 亿美元，有位 CEO 告诉我无晶圆半导体公司理论上不可能超过 10 亿。然后有人说你永远超不过 250 亿，因为某家别的公司。那些都不是第一性原理。简单想法是：我们做什么，我们能创造的机会有多大。英伟达不在市场份额生意里。我刚才讲的几乎都不存在。如果英伟达是一家 100 亿美元公司想抢英伟达的份额，股东很容易看见拿 10% 会大多少。难的是想象我们能有多大，因为没有人可以让我去抢份额。挑战是对未来的想象力。但我有很多时间，会继续推理、继续讲，每次 GTC 会越来越真。某一天我们会到。我 100% 我们会到。

I still remember the first time we crossed a billion dollars, a CEO told me it’s theoretically impossible for a fabless semiconductor company to exceed a billion. Somebody told me you’ll never be more than $25 billion because of some other company. Those aren’t first-principled thinking. The simple way: what is it that we make, and how large is the opportunity we can create? NVIDIA is not in the market-share business. Almost everything I just talked about doesn’t exist. If NVIDIA was a $10 billion company trying to take NVIDIA’s share, it’s easy to see for shareholders that if they take 10% share they could be this much larger. It’s hard for people to imagine how large we could be because there’s nobody I could take share from. One of the challenges for the world is the imagination of the future. But I got plenty of time. I’ll keep reasoning about it, keep talking about it, and every single GTC will become more and more real. One of these days we’ll get there. I’m 100% we’ll get there.

Lex Fridman

token 工厂，每瓦每秒 token，每个 token 有价值，对不同人有不同种类、不同数量的价值。产品其实可以松散地想成 token。

Token factories, tokens per second per watt, every token having value — different kinds, different amounts, to different people. The actual product could be loosely thought of as the token.

Jensen Huang

Token 的 iPhone 到了。智能体。一般意义上的智能体。Token 的 iPhone 到了。历史上增长最快的应用，直线上升。毫无疑问 OpenClaw 是 token 的 iPhone。更可能的是你的 AI 一直在烦你，因为它做事太快，回报「做完了，下一步要我干什么」。大多数人还没意识到：跟你聊天、发短信最多的，会是你的爪子或龙虾。

The iPhone of tokens arrived. Agents. Agents in general. The iPhone of tokens arrived. It is the fastest-growing application in history. It went straight up. There’s no question OpenClaw is the iPhone of tokens. It’s more likely that your AI is bothering you all the time, because it’s getting stuff done so fast. It’s reporting back, “I got that done. What do you want me to do next?” The person who’s going to be chatting with them, texting them most, is their claws or lobster.

Lex Fridman

我读到你把很多成功归因于比谁都更能干活、比谁都更能扛苦。失败、成本、工程、人的问题、不确定、责任、疲惫、难堪、公司差点死掉，还有压力。现在各国围着这家公司做战略、做资金分配、做 AI 基础设施规划。你怎么扛这么大的压力？什么给你力量？

I read that you attribute a lot of your success to your ability to work harder than anyone and withstand more suffering than anyone. Failure, cost, engineering problems, human problems, uncertainty, responsibility, exhaustion, embarrassment, near-death company moments, and the pressure. As CEO of this company that economies and nations strategize around, how do you deal with this much pressure? What gives you strength?

Jensen Huang

我清楚英伟达的成功对美国很重要。我们贡献巨额税收，为国家建立技术领导力。技术领导力对国家安全重要，不只一个方面，是所有方面。国家更繁荣，才能把国内政策和社会福祉做得更好。我们在美国制造大量再工业化，创造大量工作，帮助把造东西搬回美国：工厂、芯片、计算机、这些 AI 工厂。我完全清楚。

I’m conscious that NVIDIA’s success is very important to the United States. We generate enormous amounts of tax revenues. We established technology leadership for our nation. Technology leadership is important for national security — not just one aspect, all aspects. When our country’s more prosperous, we can do a better job with domestic policies and social benefits. Because we’re generating so much re-industrialization in the United States, we’re creating mountains of jobs. We’re helping shift how we build things back to the United States: plants, chips, computers, these AI factories. I’m completely aware of that.

我还有一个真正的礼物：主流投资者，老师、警察，不知什么原因投了英伟达，或者因为看了 Jim Cramer 买了点股票，现在成了百万富翁。我也清楚英伟达处在一个很大的生态网络中心，上游下游。我处理的方式就是我刚才做的：推理我们在做什么、造成什么、对别人是正的还是负担，比如供应链。然后你要怎么办？几乎所有我感觉到的东西，我都拆开：情况是什么、什么变了、难在哪、我要怎么办。拆成我能做的可管理的事。然后：你做了吗？自己做了还是找人做了？你推理出该做、没做、也没找人做，那就别再哭了。

I have the benefit, a real gift, with mainstream investors — teachers, policemen who for whatever reason invested in NVIDIA, or because they watched Jim Cramer bought some stock and now are millionaires. I’m aware NVIDIA is central to a very large network of ecosystem partners behind us and downstream. The way I deal with that is exactly what I just did. I reason about what we’re doing, what it’s causing, the impact on other people — positively or even through great burden, for example to supply chain. Therefore, what are you going to do about it? In almost everything I feel, I break it down: what’s the circumstance, what has changed, what’s hard, and what am I going to do about it? The decomposition turns it into manageable things I can do. Did you do it? Did you get somebody else to do it? If you reasoned that you need to do it and you didn’t do it and you didn’t get anybody else to do it, then stop crying about it.

我对自己相当狠。但我也拆，所以我不恐慌。我能睡着，因为我列了该做的事，凡是可能让公司、伙伴、行业受害的，我都告诉了能做事的人。从胸口卸下来，或者我在做。然后，Lex，你还能怎样？

I’m fairly tough on myself. I also break things down so that I don’t panic. I can go to sleep because I’ve made the list of things that needed to be done, and I’ve made sure that everything that could put our company, my partners, our industry in harm’s way, I’ve told somebody who could do something about it. I’ve gotten it off my chest or I’m doing something about it. After that, Lex, what else can you do?

Lex Fridman

建英伟达这条路上那么多苦，你心理上撞过低谷吗？

Given all the intense suffering on the journey of building NVIDIA, have you hit low points psychologically?

Jensen Huang

哦有。当然。一直有。一部分是忘记。AI 学习最重要的属性之一，你知道，是系统性遗忘。你得知道什么时候该忘。不能什么都记、什么都带着。我很快拆问题、推理、把负担分出去。告诉所有人，本质上就是在分担。别只自己扛。别把他们吓疯。拆成小块，激励人去做。然后对自己狠一点：别哭了，起床。再然后你会被下一束光吸引，下一个未来、下一个机会：「那已经过去了，下一个是什么？」就像伟大运动员，只担心下一分。上一分已经过去，难堪、挫折已经过去。

Oh yeah. Sure. All the time. Part of it is forgetting. One of the most important attributes of AI learning is systematic forgetting. You need to know when to forget some things. You can’t memorize everything, keep everything, carry everything. I decompose the problem, reason about it, and share the load. When I say I tell everybody, I’m essentially sharing that burden. Don’t just keep it. Don’t freak them out. Decompose into smaller parts and inspire people to do something about it. Part of it is just forgetting. Be tough on yourself. Come on, stop crying about it. Let’s get going. You get out of bed. The other part is you’re attracted to the next shiny light, the next future, the next opportunity. “That’s behind us. What’s next?” You watch this with great athletes. They just worry about the next point. The last point is behind them. The embarrassment, the setback.

Lex Fridman

我是游戏迷。得谢谢英伟达这么多年了不起的图形。

I’m a big gaming fan. I have to say thank you to NVIDIA for many years of incredible graphics.

Jensen Huang

顺便说，GeForce 到今天仍是我们第一营销策略。人在青少年时期认识英伟达，上大学就知道我们是谁。一开始是玩 Call of Duty、Fortnite，后来用 CUDA，再用 Blender、Dassault、Autodesk。

GeForce is still, to this day, our number one marketing strategy. People learn about NVIDIA while they’re in their teenage years. Then they go to college and they know who NVIDIA is. In the beginning it’s playing Call of Duty, Fortnite. Later they’re using CUDA, then Blender and Dassault and Autodesk.

Lex Fridman

我跟朋友说要跟你聊，他说「哦，他们做很棒的游戏 GPU」。DLSS 5 有些争议。网上玩家担心游戏会看起来像 AI 垃圾。你怎么看这场戏？

I mentioned to a friend that I’m talking with you. He said, “oh, they make great gaming GPUs.” There was some controversy around DLSS 5. Gamers online were concerned that it makes games look like AI slop. What do you think of this drama?

Jensen Huang

他们的视角说得通，我能理解从哪来的，因为我自己也不喜欢 AI 垃圾。AI 生成内容越来越像，都好看。我同情他们在想什么。但那不是 DLSS 5 要做的。DLSS 5 是 3D 条件、3D 引导，被 ground truth 结构数据引导。几何是艺术家定的。我们对每一帧的几何完全忠实。它被纹理、被艺术家的艺术性条件化。每一帧它增强，但不改任何东西。

Their perspective makes sense. I can see where they’re coming from, because I don’t love AI slop myself. All of the AI-generated content increasingly looks similar and they’re all beautiful. I’m empathetic towards what they’re thinking. That’s just not what DLSS 5 is trying to do. DLSS 5 is 3D-conditioned, 3D-guided. It’s ground-truth structure data guided. The artist determined the geometry. We are completely truthful to the geometry maintained in every single frame. It’s conditioned by the textures, the artistry of the artist. Every single frame, it enhances but it doesn’t change anything.

增强这件事：系统是开放的，你可以训练自己的模型，未来甚至可以 prompt：我要 cartoon shader，给个例子，它会按那种风格生成，同时跟艺术家的艺术性、风格、意图一致。这是给艺术家的。我觉得他们印象是游戏按出厂那样出，然后我们做后处理。那不是 DLSS 的意图。DLSS 跟艺术家集成，是给艺术家生成式 AI 这个工具。他们可以不用。

The question is about enhancing. Because the system is open, you could train your own models, and you could even in the future prompt it. I want it to be a toon shader. Give it an example. It would generate in the style of that, all consistent with the artistry, the style, the intent of the artist. All of that is done for the artist so they can create something more beautiful but still in the style they want. I think they got the impression the games are going to ship the way they do and then we’re going to post-process it. That’s not what DLSS is intended to do. DLSS is integrated with the artist. It’s about giving the artist the tool of generative AI. They could decide not to use it.

Lex Fridman

得值超过十亿美元。你知道把那些组件做齐有多难。一个 Open-Claude 去做创新、找客户、销售、管理、建团队——有些智能体、有些人。这离我们还有五年、十年、十五年、二十年？

It has to be worth more than a billion dollars. You know how hard it is to do all those components. How far are we from an Open-Claude that does all the incredibly complex stuff: innovate, find customers, sell to them, manage, build a team of some agents, some humans? Is this five, 10, 15, 20 years away?

Jensen Huang

我觉得就是现在。我觉得我们已经实现 AGI。

I think it’s now. I think we’ve achieved AGI.

Lex Fridman

你觉得能有一家公司由这样的 AI 系统来运营吗？

Do you think you could have a company run by an AI system like this?

Jensen Huang

可能。因为你说了十亿，你没说永远。并非不可能：一个 Claude 做出一个网络服务、一个有趣的小应用，忽然几十亿人用，每人 50 美分，然后很快又倒了。互联网时代我们见过一堆那样的公司，那些网站大多并不比今天 Open-Claude 能生成的更复杂。

Possible, and the reason is this. You said a billion, and you didn’t say forever. It is not out of the question that a Claude was able to create a web service, some interesting little app that all of a sudden a few billion people used for 50 cents, and then it went out of business again shortly after. We saw a whole bunch of those type of companies during the internet era, and most of those websites were not anything more sophisticated than what Open-Claude could generate today.

Lex Fridman

这句话会让很多人兴奋。你的意思是我可以启动一个智能体就赚很多钱？

You’re gonna get a lot of people excited with that statement. What do you mean, I can just launch an agent and make a lot of money?

Jensen Huang

顺便说，这正在发生。你去中国会看到一堆人在教他们的 Claude 去找工作、做事、赚钱。我不会惊讶如果某种社交的东西发生，有人做出一个超级可爱的数字网红，或某种养电子宠物的社交应用，突然成功，很多人用几个月然后淡掉。十万个那样的智能体建成英伟达的概率是零。

By the way, it’s happening right now. When you go to China you’re going to see a whole bunch of people teaching their Claudes to go look for jobs, do work, make money. I wouldn’t be surprised if some social thing happened, or somebody created a digital influencer, super cute, or some social application that feeds your little Tamagotchi, and it became out of the blue an instant success. A lot of people use it for a couple of months and it kind of dies away. The odds of 100,000 of those agents building NVIDIA is zero percent.

我不想做、也希望我们都做的一件事，是承认人真的在担心工作。我只想提醒：工作的目的，和你用来做工作的任务、工具，相关，但不是同一回事。我做这工作 33 年了。我是世界上在任最长的科技 CEO，34 年。这 34 年里我用来做工作的工具一直在变，有时两三年就变得很剧烈。计算机科学家、AI 研究者说会消失的第一份工作是放射科。因为计算机视觉会达到超人类水平，它确实达到了。CV 在 2019、2020 年就已经超人类。预测是放射科医生会消失，看片子是过去的事，AI 会做。他们完全说对了：计算机视觉完全超人类，今天每个放射平台和套件都由 AI 驱动，但放射科医生人数增长了。现在全世界还短缺放射科医生。

The one part I won’t do, and I want to make sure we all do, is recognize that people are really worried about their jobs. I just want to remind them that the purpose of your job and the tasks and tools you use to do your job are related, not the same. I’ve been doing my job for 33 years. I’m the longest-running tech CEO in the world, 34 years. The tools I’ve used have changed continuously, sometimes quite dramatically over two, three years. The first job computer scientists, AI researchers said was going to go away was radiology, because computer vision was going to achieve superhuman levels, and it did. CV was superhuman in 2019, 20, maybe 2020. The prediction was radiologists would go away. Studying radiology scans was a thing of the past. AI will do that. They were absolutely right. Computer vision is completely superhuman. Every radiology platform and package today is driven by AI, and yet the number of radiologists grew. We now have a shortage of radiologists in the world.

警报走太远，吓得人不敢做对社会那么重要的职业，造成了伤害。错在哪？放射科医生的目的是诊断疾病、帮病人和医生诊断。片子看得快那么多，你可以看更多、诊断更好、住院更快、见更多人。医院赚更多，病人更多，你需要更多放射科医生。这么明显会发生。英伟达的软件工程师人数会增长，不会下降。软件工程师的目的和写代码这个任务相关，不是同一回事。我要软件工程师解决问题，我不在乎他们写了多少行。解决问题、团队协作、诊断问题、评估结果、找新问题、创新、连点，这些都不会消失。

The alarmist warning went too far and it scared people from doing this profession that is so important to society. It did harm. Why was it wrong? The purpose of a radiologist is to diagnose disease and help patients and doctors diagnose disease. Because we’re able to study scans so much faster, you could study more scans, diagnose better, in-patient faster, see people more. Hospitals are making more money. You have more patients. You need more radiologists. It’s so obvious this was going to happen. The number of software engineers at NVIDIA is going to grow, not decline. The purpose of a software engineer and the task of coding are related, not the same. I wanted my software engineers to solve problems. I didn’t care how many lines of code they wrote. Solving problems, working as a team, diagnosing problems, evaluating the result, looking for new problems, innovation, connecting dots. None of that is going to go away.

Lex Fridman

就拿编程来说，你觉得世界上程序员的数量可能增加而不是减少吗？

Let’s even take coding. Do you think the number of programmers in the world might increase, not decrease?

Jensen Huang

会。今天编码的定义就是规格说明，你想更指令化一点，还可以给它软件架构。有多少人能描述规格、告诉计算机去建什么？我觉得我们刚从 3000 万到大概 10 亿。未来每个木匠都会是编码者，只是带 AI 的木匠也是建筑师。他们能交付给客户的价值刚刚升高，艺术性大幅抬升。每个会计师也是你的财务分析师、财务顾问。这些职业刚刚被抬高了。如果我是木匠，看见 AI 我会完全疯掉。水管工也是。

Yes. What is the definition of coding? As of today it is simply specifying, specification, and maybe if you want to be rather directive you could even give it an architecture. How many people could describe a specification, telling the computer what to go build? I think we just went from 30 million to probably 1 billion. Every carpenter in the future will be a coder, except a carpenter with AI is also an architect. They’ve just increased the value they could deliver to the customer. Their artistry just elevated tremendously. Every accountant is also your financial analyst, also your financial advisor. All of these professions have just been elevated. If I were a carpenter, I see AI, I would just completely go berserk. If I were a plumber, completely go berserk.

Lex Fridman

你觉得人性、人类意识里有没有本质上不可计算的东西？也许无论芯片多强都复制不了？

Do you think there’s some things about human nature, about human consciousness, that is fundamentally non-computational? Maybe something a chip, no matter how powerful, can never replicate?

Jensen Huang

我不知道芯片会不会紧张。造成焦虑、紧张或任何情绪的那些条件——我相信 AI 能识别、能理解。我不认为我的芯片会感觉到那些。那种感觉如何在人类表现里显现：极致的运动表现，或平均、低于平均。完全相同的情境，不同的人，不同的结果、不同的表现。没有任何我们在建的东西表明：两台不同的计算机被给予完全相同的上下文，会因为感觉不同而表现不同。当然统计上会产出不同结果，但不是因为它感觉不同。

I don’t know if the chip will ever get nervous. The conditions that cause anxiety or nervousness or whatever emotion — I believe AI will be able to recognize those and understand those. I don’t think my chips will feel those. How that anxiety, that feeling, that excitement, all of those feelings manifest in human performance: extremely amazing athletic performance, average or lesser than average. That entire spectrum from exactly the same circumstances for different people, manifesting a different outcome, a different performance. I don’t think there’s anything about anything we’re building that would suggest two different computers presented with exactly the same context would perform differently because they felt different. Of course it would produce statistically different outcomes, but it’s not because it felt different.

Lex Fridman

主观体验有真正特别的地方。我跟你聊之前挺紧张。希望、恐惧、焦虑，生命本身的丰盛，我们爱得有多深、心碎有多深、多怕死、亲人离去有多痛。很难想象计算设备能那样。但这整件事还有很多谜，我愿意被惊喜。缩放在智能空间里能创造奇迹。

There’s something truly special about the subjective experience we humans feel. I was pretty nervous talking to you. The hope, the fear, the anxiety, life itself, the richness of life. How deeply we fall in love, how deeply our hearts get broken, how afraid we are of death and how much pain we feel when our loved ones pass away. It’s very hard to think a computational device being able to do that. But there are so many mysteries we’re yet to uncover that I am open to be surprised. Scaling can create some incredible miracles in the space of intelligence.

Jensen Huang

真正重要的是拆开智能是什么。这个词我们一直用，它不是神秘词，有含义：感知、理解、推理、规划的能力。那个循环从根本上就是智能。智能不是一个词，不等于人性。我们有两个词。我不过度幻想、不过度浪漫化智能。智能是商品。我周围都是聪明人，在各自领域都比我聪明。可我在那个圈子里有角色。他们比我受教育好、学校更好、每个领域更深。我有 60 个，对我来说都是超人。不知怎的我坐在中间编排这 60 个。你得问自己：一个洗碗工凭什么坐在超人们中间？智能是功能性的东西。人性不是按功能规定的，是大得多的词。生命经验、对痛的耐受力、决心，那些是和智能不同的词。我们把智能这个词抬得很高。

It’s really important to break down what intelligence is. That word we use all the time is not a mysterious word. Intelligence has a meaning. It’s something we do that includes perception and understanding and reasoning and the ability to plan. That loop is fundamentally what intelligence is. Intelligence is not one word that is exactly equal to humanity. We have two words for that. I don’t over-fantasize, I don’t over-romanticize about intelligence. Intelligence is a commodity. I’m surrounded by intelligent people more intelligent than I am in each of the spaces they’re in. And yet I have a role in that circle. They’re more educated than I am. They went to better schools. They’re deeper in any of the fields they’re in. All of them. I have 60 of them. They’re all superhuman to me. And somehow I’m sitting in the middle orchestrating all 60 of them. You gotta ask yourself: what is it about a dishwasher that allows that dishwasher to sit in the middle of superhumans? Intelligence is a functional thing. Humanity is not specified functionally. It’s a much, much bigger word. Our life experience, our tolerance for pain, our determination — those are different words than intelligence. Intelligence is a word we’ve elevated to a very high form over time.

Lex Fridman

我们真正该抬高的词是人性。

The word we should really elevate is humanity.

Jensen Huang

品格，人性。同情、慷慨，你刚才说的那些。我相信那些是超人的力量。现在智能会被商品化。

Character, humanity. Compassion, generosity, all of the things you said just now. I believe those are superhuman powers. And now intelligence is going to be commoditized.

Lex Fridman

英伟达的成功和我提到的几百万人的生活依赖你。但你只是一个人，像我们所有人一样会死。你想过自己的死亡吗？你怕死吗？

So much of the success of NVIDIA and the lives of millions of people depend on you. But you’re just one human, a mortal like all of us. Do you think about your mortality? Are you afraid of death?

Jensen Huang

我真的不想死。我有很好的生活、很好的家庭、非常重要的工作。这不是「一生一次」那种说法——一生一次意味着很多人经历过，只是不是同一个人。这是人类一次的体验，我正在经历的。英伟达是历史上最重要的科技公司之一。我们做非常重要的工作。我非常认真。当然有些实际的事，比如怎么想接班计划。我出名的一点是：我不相信接班计划。

I really don’t wanna die. I have a great life. I have a great family. I have really important work. This is not a once-in-a-lifetime experience — that suggests it has been experienced by many people, just not one person. This is a once-in-a-humanity experience, what I’m going through. NVIDIA is one of the most consequential technology companies in history. We’re doing very important work. I take it very seriously. Some of the things that of course are practical, like how do we think about succession planning? I’m famous in saying that I don’t believe in succession planning.

原因不是我不死。原因是：如果你为接班计划焦虑，那你今天该做什么？拆到底。如果你在乎你走了以后公司的未来，今天最重要的事是尽可能频繁、持续地把知识、信息、洞察、技能、经验传出去。所以我在团队面前不停推理。每一次会都是推理会。我在公司里外花的每一刻都是尽快把知识传给别人。我学到的东西在桌上停留不超过一瞬间。还没学完我就指给别人：「盯这个，这太酷了，你会想学。」我不停传知识、赋能、抬高周围每个人的能力。我寻求的结果是：我在岗位上死去。希望是瞬间死去，没有很长的受苦。

The reason isn’t because I’m immortal. If you’re worried about succession planning, all that anxiety, then what should you do about it? Break it all the way back down. The most important thing you should do today, if you care about the future of your company post you, is to pass on knowledge, information, insight, skills, experience as often and continuously as you can. That’s why I continuously reason about everything in front of my team. Every single meeting is a reasoning meeting. Every moment I spend inside a company, outside a company, is about passing on knowledge as fast as I can. Nothing I learn ever sits on my desk longer than a fraction of a second. Before I even finish learning all of it myself, I’m already pointing it to somebody else. “Get on this. This is so cool. You’re gonna wanna learn this.” I’m constantly passing knowledge, empowering people, elevating the capability of everybody around me, so that the outcome I seek, that I hope for, is that I die on the job. Hopefully I die on the job instantaneously. No long periods of suffering.

Lex Fridman

从粉丝角度看，你对文明有极大的正面影响，我当然希望你继续。看英伟达在做什么也很有趣，创新的速率，大量工程。这是对人性、对伟大建造者、对伟大工程的庆祝。展望 10、20、50、100 年，什么让你对人类的未来有希望？

From a fan perspective, given your extremely enormous positive impact on civilization, of course I hope you keep going. It’s just fun to watch what NVIDIA is doing, the rate of innovation. I’m a huge fan of engineering. It represents something special. When you look out 10, 20, 50, 100 years from now, what gives you hope about humanity?

Jensen Huang

我一直对善良、慷慨、同情、人的能力有很大信心。有时超过我该有的，我会被占便宜，但从不因此停下来。我总是从人想做好事、想帮别人开始。绝大多数时候我被证明是对的，常常超出预期。所以我对人的能力有完全的信心。让我极有希望的，是我现在看见的可能，以及按我们在做的事外推、很可能发生的事。我们想解决那么多问题，想建那么多东西，想做那么多好事，现在到了我们够得着的地方，到了我有生之年够得着的地方。你不可能不对此浪漫。疾病的终结是可以合理期待的。污染大幅减少是可以合理期待的。短距离以光速旅行实际上在我们的未来，是可以合理期待的。很快我会把一个人形机器人送上飞船，它会在飞行中不断改进。时机到了，我那么多生命已经上传到互联网：收件箱、做过的、说过的，会变成我的 AI。到时候我们把它以光速发出去，追上我的机器人。

I’ve always had a great confidence in the kindness, the generosity, the compassion, the human capacity. Sometimes more so than I should. I get taken advantage of, but it doesn’t ever cause me not to. I start with always that people want to do good, people want to help others. Vastly I am proven right, constantly, and often it exceeds my expectations. I have complete confidence in the human capacity. The things that give me incredible hope are what I see now as possible, and as I extrapolate based on the things we’re doing, what will very likely happen. There are so many things we want to solve, so many things we want to build, so many good things we want to do that are now within our reach, and within the reach of my lifetime. You just can’t possibly not be romantic about that. It’s a reasonable thing to expect the end of disease. It’s a reasonable thing to expect that pollution will be drastically reduced. It’s a reasonable thing to expect that traveling at the speed of light is actually in our future — not for long distances, but short distances. Very soon I’m going to put a humanoid on a spaceship, my humanoid, send it out as soon as possible, and it’s going to keep improving and enhancing along the flight. When the time comes, so much of my life has been uploaded in the internet. Take all my inbox, everything I’ve done, everything I’ve said. It’s been collected and becoming my AI. When the time comes, we’ll just send that at the speed of light, catch up with my robot.