Lex Fridman #494 — 4 万亿美元的 NVIDIA 与 AGI 革命
主持人：Lex Fridman
发布时间：2026-03-23
时长：约 2 小时 25 分
主题：Extreme Co-Design · 机柜级工程 · AI scaling laws · 34 年 CEO 经验 · 供应链 · 电力瓶颈 · AGI 已实现 · 意识与死亡
这期访谈录于 NVIDIA 市值突破 4 万亿美元、黄仁勋成为"34 年 CEO"之后。他坐在 Lex 对面，把自己 34 年的方法论、正在发生的"AI 工业革命"、以及一些很私人的思考——关于死亡、意识、什么是"智能"——都摊开讲了。
一、Extreme Co-Design：跨芯片、系统、软件的整体优化
访谈一开始 Lex 就切入核心问题：现在的 NVIDIA 已经不再只设计 GPU，而是在整机柜级别（rack-scale）设计整套系统——GPU、CPU、内存、网络、存储、电源、冷却、软件、机柜、Pod 甚至整个数据中心。为什么必须这样？
黄仁勋的回答从第一性原理开始：
"The problem no longer fits inside one computer. You would like to go faster than the number of computers that you add. So you added 10,000 computers, but you would like it to go a million times faster. Then all of a sudden you have to take the algorithm, shard the pipeline, shard the data, shard the model. 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."
"问题已经不能装进一台计算机里了。你加上 1 万台计算机，却希望它跑得比 1 万倍还快 100 倍。于是你必须把算法拆开，把 pipeline、数据、模型都分片。在这种规模的分布式计算里，CPU 是问题，GPU 是问题，网络是问题，交换机是问题——每一样东西都挡在路上。"
他引出了 摩尔定律已死 这个他长年强调的判断：Dennard scaling 已经停下来了，摩尔定律带来的免费午餐吃完了。要继续让性能指数级增长，就必须跨层打通——架构、芯片、系统、系统软件、算法、应用全部一起优化。这就是 Extreme Co-Design。
他说 NVIDIA 直属员工超过 60 个（"More"），都是在内存、CPU、光学、GPU、架构、算法、设计等领域的世界级专家。他不开 不开 1-on-1 会议——不是因为牛，而是因为"做不到"：
"You can't have 60 people on your staff if you're gonna get work done. And no conversation is ever one person. We present a problem and all of us attack it. Because we're doing extreme co-design, whoever wants to tune out, tune out."
"如果你想干活，就不可能给 60 个直接下属一个个开 1-on-1。而且任何一场对话都不应该只有一个人——我们把一个问题端出来，所有人一起攻。这是 extreme co-design，谁想走神就走神，到时候我一定会喊他回来。"
二、CUDA：为什么把公司押在一个没有用户的平台上
Lex 问他：最接近"存亡级决策"的那一次是什么？黄仁勋说——是把 CUDA 平台 放进每一张 GeForce。
那是 NVIDIA 只有 70 亿美元市值左右的时候。决策的逻辑是：
"A computing platform is all about developers. And developers don't come to a computing platform just because it could perform something interesting. They come to a computing platform because the install base is large. Install base defines an architecture. Everything else is secondary."
"计算平台的本质是开发者。开发者之所以来到一个平台，不是因为它能做出什么酷炫的东西，而是因为装机量大。装机量定义一个架构，其它一切都是次要的。"
为了让 CUDA 平台 有装机量，他们决定把 CUDA 放进每一张 GeForce——即使玩家根本用不上。这让每一颗 GPU 的成本大涨，公司毛利直接崩盘，市值一度从 70 多亿美金跌到只有 15 亿美金。
"CUDA increased our cost so tremendously, it completely consumed all of the company's gross profit dollars. NVIDIA is the house that GeForce built, because it was GeForce that took CUDA out to everybody. Researchers, scientists, they discovered CUDA on GeForce because many of them were gamers."
"NVIDIA 是 GeForce 盖起来的这栋房子，因为是 GeForce 把 CUDA 带给了所有人——研究者、科学家都是通过 GeForce 发现 CUDA 的，因为他们很多人本身就是玩家。"
从决定到真正变现"花了十年"。
三、用"推理"而非"宣告"来领导 NVIDIA
Lex 追问：面对这种级别的存亡决策，你怎么做到让员工和董事会跟你一起下注？黄仁勋给了一段非常有启发的回答：
"Oftentimes in leadership, the leader stays quiet or they learn about something, and then they do some manifesto, a brand-new year, and somehow at the end of the year we have a brand-new plan. Big huge layoff, big huge organization change, new mission statement, brand new logos. I never do things that way."
"很多 CEO 的做法是：自己闷头学、闷头想，然后突然来一个'宣言'——年初大变革、裁员、改架构、换使命语句、换新 logo。我从来不这样做。"
他的做法是——把自己的推理过程实时暴露给整个组织。每天、每场会议、每次跟客户和供应商聊天，他都在"铺砖头"，把还没确定的观察、新的工程突破、一点一点地注入到团队的共同信念里。
"When I come the day I say, 'Hey, let's buy Mellanox,' it's completely obvious to everybody that we absolutely should. I like to announce these things and imagine that the employees are kind of saying, 'Jensen, what took you so long?' "
"等我真的说'我们去收购 Mellanox 吧'的那一天，对所有人来说都是显而易见的——他们会说'Jensen 你怎么拖到现在才来宣布'。"
他把这叫做 "leading from behind"——真正的领导力看起来是从后面推的，因为当你宣布的时候，所有人已经 100% 同意了。GTC 的主题演讲是他"塑造整个行业信念系统"的工具：当 Vera Rubin 机架系统发布的时候，合作伙伴们早就已经准备好了，因为他早在两年前就开始铺垫。
四、四条 Scaling Laws 与"思考 = 很贵"
很多人把 Ilya Sutskever 的"预训练已死"理解成 scaling laws 结束了。黄仁勋的判断完全相反——他说现在有四条 scaling law：
- Pre-training scaling——数据越多，模型越聪明。合成数据会继续推进这条曲线。
- Post-training scaling——AI 反复练习一个技能直到学会，背后是强化学习和在线推理。
- Test-time scaling——思考时间越长，答案越好（test-time compute、思维链）。
- Agentic scaling——一个 agent 衍生出一堆 sub-agents，"多智能体"像"我扩招员工"一样把计算扩展出去。
他对"推理 = 简单 = 便宜"这个被广泛流传的观点非常不屑：
"Inference is thinking, and I think thinking is hard. Thinking is way harder than reading. Pre-training is just memorization and generalization, looking for patterns in relationships. You're reading. Versus thinking, reasoning, solving problems, breaking it down into solvable pieces. How could that possibly be compute light?"
"推理就是思考，而思考是困难的。思考比阅读难得多。预训练只是记忆和泛化，只是在找关系中的模式——那只是'读书'。而思考、推理、解决问题、把一个未见过的问题拆成可求解的部分……这怎么可能是轻量计算？"
更关键的是——这四条 scaling law 构成一个闭环：
"The agentic systems generate a lot more data and experiences. Some of it we say, 'Wow, this is really good.' That data set comes back to pre-training. We memorize and generalize it. We refine it back into post-training. We enhance it with test time. And then put it out to the industry. This loop is gonna go on and on. It kind of comes down to basically intelligence is gonna scale by one thing, and that's compute."
"Agent 系统会产生大量数据和经验——好的部分会回流到 pre-training，再精调到 post-training，再被 test-time 增强，然后释放到整个行业。这个循环会一直转下去。归根结底，智能会通过一样东西来扩展：计算。"
五、"我想让大语言模型变成一个数字员工"——Agent 架构的必然性
Lex 问："从 Grace Blackwell 到 Vera Rubin，机柜的形态都变了——你是从哪里知道要这么改的？" 黄仁勋说："其实比你想的简单，你只要 reason about it。"
"If we want the LLM to be a digital worker, what does it have to do? It has to access ground truth. That's our file system. It has to do research. It has to use tools. A lot of people say, 'AI is gonna destroy software, we don't need tools anymore.' That's ridiculous."
"如果我们想让大语言模型变成数字员工，它必须做什么？它必须访问 ground truth——也就是我们的文件系统。它必须做研究。它必须使用工具。很多人说'AI 来了我们就不需要软件和工具了'——这是荒谬的。"
他用了一个漂亮的思想实验：如果 10 年后我造出了最强的人形机器人，它走进我家，是更有可能用现成的微波炉，还是它的手会变成一把 10 磅的锤子或手术刀、手指会发射微波？ 当然是用微波炉。第一次它可能不会用，但它连上互联网读一遍说明书就变成专家了。
这就是 Vera Rubin rack 和新的 "Rock" 系统存在的原因——它不是为了跑纯 LLM，而是为了支持使用工具、调研资料、做决策、派生子 agent 的数字员工。他说他两年前的 GTC 主题演讲里就画过这张图。
六、供应链：提前 5 年说服 HBM 厂商做 AI 内存
黄仁勋花了一大段时间讲 NVIDIA 是怎么"说服整个世界上游和下游"的。
三年前，HBM 还只是超算里的小众内存。他挨个说服几家 DRAM 厂商的 CEO——"这会变成未来的数据中心主流内存"。第二个更怪的决定是把手机用的低功耗内存 LPDDR5 搬进超算。所有人第一反应都是"你在开玩笑吗？"但几位 CEO 选择了相信他，去投资扩产。
"At first it sounded ridiculous, but several of the CEOs believed me and decided to invest in building HBM memories. The volumes are so incredible. All three of them had record years in history, and these are 45-year companies."
"这些都是 45 年的老公司，三家都创下了历史最高业绩。"
他说他自己的工作是"上游和下游同时 shape、inform、inspire"——从 ASML、TSMC、SK Hynix、一路到 GE、Caterpillar（因为电力和重型设备）。每个机架 130 到 150 万个零件，200 家供应商。
Lex 问："你怎么能睡得着？" 他说：
"I can go to sleep because I checked it off. Because I told them what I needed. They understood what I need. They told me what they're gonna go do, and I believe them. After that, Lex, what else can you do?"
"我能睡着，是因为我把它从清单上划掉了。我告诉了他们我需要什么，他们理解了，他们告诉了我他们要做什么，我相信他们。之后，Lex，我还能做什么？"
他讲了一个"减压 = 分解 + 分享"的方法论：
"I decompose the problem. Whatever worries me, I tell somebody else. Don't just keep it. Don't freak them out. Decompose into smaller parts and inspire people to go do something about it. Everything that I feel could put anybody in harm's way, I've told someone. So I've gotten it off my chest. After that, what else can you do?"
"我把问题拆开。任何让我担心的事，我都告诉另一个人。不要自己扛着，也不要吓到他们，而是把它拆成小块，让别人被激发去做点什么。任何可能伤害到人的事，我都告诉了能处理它的人。之后还能做什么？"
七、电网的"闲置电力"——最低垂的果实
Lex 问："AI 的最大瓶颈是不是电？" 黄仁勋说了一个大多数人根本没意识到的事实：
"Our power grid is designed for the worst case condition. 99% of the time we're nowhere near the worst case condition. Most of the time we're probably running around 60% of peak. 99% of the time, our power grid has excess power, and they're just sitting idle."
"电网是按最坏情况设计的——夏天和冬天最极端的那几天。99% 的时间我们远远没到这种极限，平均大概只用到 60%。99% 的时间电网都有闲置电力，就放在那里。"
他的提议是：让数据中心和电网签一份"可降级服务"的合同——在真正用电高峰（医院、机场需要），数据中心自动降速、把关键任务转到别的数据中心，或短暂使用备用发电机。
"We could engineer data centers that gracefully degrade. We're just gonna move our workload around. We can reduce the computing rate and use less energy. The quality of service degrades a little bit."
"我们可以设计出'优雅降级'的数据中心——工作负载会自动迁移，计算速率降低，服务质量略微下降。"
目前卡住这件事的是三方：客户要求 6 个 9 的可用性、云服务提供商照搬给电力公司、电力公司为了兜底被迫扩容。他说："我要去和所有 CEO 聊这件事——他们的数据中心团队签的合同，他们自己可能都不知道。"
八、Elon 和 Colossus：为什么他造数据中心这么快
Lex 问他怎么看 Elon Musk 在 Memphis 用四个月建起 Colossus 超算（20 万卡，后来扩到 50 万卡）的事。黄仁勋的观察很细：
"Elon is deep in so many different topics. He's able to think through multiple disciplines, and he questions everything. Number one, is it necessary? Number two, does it have to be done this way? Number three, does it have to take this long? He has the ability to question everything to the point where everything is down to its minimal amount that's necessary."
"Elon 在很多领域都钻得很深，他能跨学科思考，他会质疑一切：第一，这有必要吗？第二，必须这么做吗？第三，必须花这么久吗？他会一直问一直问，问到一切都只剩下最小必要的东西。"
最关键的是他人在现场：
"He'll just go there. 'Show me the problem.' When you act personally with so much urgency, it causes everybody else to act with urgency. Every supplier has a lot of customers, a lot of projects going on. He makes it his business that he's the top priority of everybody else's projects. And he does that by demonstrating it."
"他就出现在问题现场，说'给我看看这个问题'。当你自己都带着如此强烈的紧迫感行动时，它会感染所有人。每家供应商都有很多客户、很多项目——Elon 的做法是让自己成为所有人的 top priority，而他是通过亲自示范做到的。"
九、"Speed of Light"：对抗"持续改进"的思维方式
Lex 问他有没有一个自己的工程哲学，黄仁勋讲了他 30 年前就开始用的一个方法——Speed of Light。
"Speed of light is my shorthand for what's the limit of what physics can do. Every single thing that we do is compared against the speed of light. Memory speed, math speed, power, cost, time, effort, number of people, manufacturing cycle time."
"Speed of light 是我的缩写——指物理能做到的极限。我们做的每一件事都要对照光速：内存速度、数学速度、功耗、成本、时间、人力、制造周期时间。"
他最讨厌的一种工程思维是"continuous improvement"（持续改进）：
"Somebody says, 'Hey, it takes 74 days to do this today. We can do it for you in 72 days.' I'd rather strip it all back to zero and say, 'Explain to me why 74 days in the first place. And if I were to build it from scratch, how long would it take?' Oftentimes you'd be surprised. It might come to six days. Now the rest, 74 to six, is very well-reasoned compromises and cost reductions. But at least you know what they are."
"有人说'这个事今天要 74 天做完，我们能帮你做到 72 天'——不，我更愿意从零开始问：首先为什么要 74 天？如果我从头造，需要多久？往往你会惊讶——可能是 6 天。剩下的 74 天减 6 天，都是有充分理由的妥协和成本削减，但你至少知道它们是什么。"
十、NVIDIA 的护城河：装机量、速度、信任
Lex 问："NVIDIA 最大的护城河是什么？" 他说：
"Our single most important property as a company is the install base of our computing platform. It's the install base of CUDA. If somebody came up with a GUDA or TUDA, it wouldn't make any difference. Because it's never been just about the technology. It was 43,000 people that made CUDA successful."
"我们最重要的资产是计算平台的装机量——也就是 CUDA 平台 的装机量。如果今天有人推出一个 GUDA 或者 TUDA，根本不会有任何意义。因为这从来不只是技术的问题——是 43,000 个人让 CUDA 成为 CUDA。"
他站在一个开发者的角度推导：
"If I support CUDA, tomorrow it'll be 10 times better. I just have to wait six months on average. If I develop on CUDA, I reach a few hundred million computers. I'm in every cloud, every computer company, every industry, every country. And I trust 100% that NVIDIA is going to keep CUDA around and maintain it for as long as they shall live."
他说"计算单位"的定义也在变：早年是 GPU，后来是一台计算机，再后来是集群，现在是整个 AI 工厂。他的"mental model"早就不是那颗芯片了，而是"一个连着电网的千兆瓦级怪物，装它就要上千人"。他说他希望下一次 click 是"行星级"——planet scale。
十一、AGI 已实现：我们已经到了
Lex 抛出一个经典问题："AGI 什么时候到？——我给它的定义是：一个 AI 能够启动并运营一家市值超过 10 亿美金的科技公司。" 黄仁勋的回答令全场哑然：
"I think it's now. I think we've achieved AGI."
"我觉得现在就是。我觉得我们已经实现 AGI 了。"
他的理由很有意思——你只说"10 亿美金"，没说"永远维持 10 亿美金"。一个 agent 完全可以造出一个 cute 的 web app、爆红、被几十亿人用几个月、再消失，就像互联网早期那些昙花一现的公司。
但他紧接着提醒："10 万个 agent 组队造出一家 NVIDIA 的概率是 0。"
然后他讲了一个重要的故事——radiologist（放射科医生）。2019-2020 年计算机视觉就已经超越了人类，所有人都预测放射科医生会消失。结果呢？
"The number of radiologists grew. And we now have a shortage of radiologists. The alarmist warning went too far and it scared people from doing this profession that is so important to society."
"放射科医生的数量增加了。现在全世界缺放射科医生。这种危言耸听的警告走得太远，吓得大家不敢再从事这个对社会如此重要的职业——它是在做伤害。"
原因是什么？因为放射科医生的目的（诊断疾病）和任务（看片子）是两回事。当看片子变快，医生就能看更多片子、诊断更多病人、帮更多医院赚更多钱、于是需要更多放射科医生。
"The purpose of your job and the tasks and tools that you use to do your job are related, not the same."
"你工作的目的，和你做这份工作所用的任务与工具，是相关的，但不是同一件事。"
同样地，他预测编程者的数量会从 3000 万涨到 10 亿：
"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."
"未来每个木匠都会是程序员——只不过带着 AI 的木匠同时也是建筑师。他们给客户的价值凭空翻了一倍。"
十二、关于意识、智能和"人性"
这是访谈里最安静也最深的一段。Lex 问：人类意识里有没有什么东西是计算无法复制的？黄仁勋慢慢地说：
"I don't know if the chip will ever get nervous. I believe AI will be able to recognize and understand those emotions. I don't think my chips will feel those. How that anxiety, how that feeling manifests in human performance—extremely amazing human performance, athletic performance, average, lesser than average—that entire spectrum comes out of exactly the same circumstances for different people, manifesting a different outcome."
然后他走向一个让我想反复读几遍的论断：
"Intelligence is a commodity. I'm surrounded by intelligent people more intelligent than I am in each one of the spaces they're in. 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?"
"智能是一种商品。我身边都是比我聪明的人，在他们各自领域都远比我博学。我有 60 个这样的人，他们对我来说都是超人——但我这个'洗盘工'却坐在 60 个超人中间指挥他们。你必须问自己：到底是什么东西，让一个洗盘工能坐在一群超人中间？"
"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."
"智能是一个功能性概念。而'人性'不是功能性的，它是一个大得多的词。我们的生活经验、我们对痛的忍耐、我们的决心——这些词和'智能'不是一件事。"
他的结论非常温柔也非常坚定：
"The word we should really elevate is humanity. Character, humanity. Compassion, generosity—I believe those are superhuman powers. Intelligence is gonna be commoditized. Don't let this commoditization of intelligence cause you anxiety. You should be inspired by that."
"我们真正应该被抬升的词是'人性'。品格、人性、慈悲、慷慨——这些才是真正的超能力。智能会被商品化。不要让智能的商品化让你焦虑——你应该被它激励。"
十三、关于死亡：我相信继任规划就是"每天传递知识"
Lex 问："你怕死吗？" 他说：
"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—this is a once in a humanity experience, what I'm going through."
但他不相信传统意义上的"继任规划"（succession planning）：
"If you're worried about succession planning, what should you do about it? 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. Which is the reason why I continuously reason about everything in front of my team. Every single meeting is a reasoning meeting. Nothing I learn ever sits on my desk longer than a fraction of a second. I'm passing that information, that knowledge— 'Oh my gosh this is cool, get on this'—before I even finish learning all of it myself."
"如果你担心继任，那你今天就应该尽可能频繁地、持续地传递知识、信息、洞察、技能、经验。这就是为什么我每场会议都是一场推理会议。我学到的任何东西在桌上停留的时间都不会超过一刹那——'哇这个太酷了，你快去研究一下'——我甚至在自己还没完全学完的时候就已经在传给别人了。"
然后他说了一句很柔软的话：
"I hope that I die on the job. And hopefully I die on the job instantaneously, no long periods of suffering."
"我希望自己死在工作岗位上——而且最好是瞬间的，不要有长时间的痛苦。"
十四、尾声：对人类的乐观
Lex 最后问："当你望向未来，什么给你希望？"
"I've always had 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 the assumption that people want to do good. And vastly, I am proven right."
他描绘了一个未来：
"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. And understanding the biological machine is right around the corner. It's five years, probably."
"我们可以合理地期待疾病的终结。我们可以合理地期待污染大幅减少。我们可以合理地期待光速旅行在我们未来是可能的。而理解生物这台机器就在眼前——可能五年。"
访谈的最后他半开玩笑地说——他会先把一个人形机器人送上宇宙飞船，让它一边飞一边进化，等到时候到了，他会把自己的意识（"我已经把人生很多东西上传到互联网了——邮箱里的所有东西、说过的所有话"）用光速发过去，追上那个机器人。
金句集
- "Every problem no longer fits inside one computer to be accelerated by one GPU."
- "I never do things that way." — 关于"年初大变革"式的领导力
- "Install base defines an architecture. Everything else is secondary."
- "Inference is thinking, and I think thinking is hard."
- "Intelligence is gonna scale by one thing, and that's compute."
- "I think it's now. I think we've achieved AGI."
- "The purpose of your job and the tasks you use to do your job are related, not the same."
- "Intelligence is a commodity. The word we should really elevate is humanity."
- "I hope I die on the job. Hopefully instantaneously."
原文出处：Lex Fridman Podcast #494 — Jensen Huang: NVIDIA The $4 Trillion Company & the AI Revolution