BG2 Pod — 黄仁勋谈 OpenAI、算力未来与美国梦
主持人：Bill Gurley、Brad Gerstner、Clark Tang
发布时间：2025-09-26
时长：约 1 小时 30 分
主题：OpenAI Stargate 1000 亿美元大单 · Token 经济 · AI 工厂 · 资本支出周期 · 电力与基建瓶颈 · Extreme Co-Design · 中国与主权 AI
这一期录于 NVIDIA 刚刚宣布要向 OpenAI 投资 1000 亿美元、合作建设 10 GW "Stargate" 算力之后。Brad 和 Bill Gurley 本身就是顶级资金分配者，他们把华尔街的怀疑、周期性过剩担忧、ASIC 竞争、中国市场、主权 AI、美国梦这些最硬的问题一个个摆到 黄仁勋 面前。下面按主题把他的核心论述整理出来。
一、三条 Scaling Laws：一年前说 10 亿倍是低估
Brad 一开场就追溯到一年半前黄仁勋做的那个预测——"推理不是涨 100 倍或者 1000 倍，而是 10 亿倍"。黄仁勋笑着说：
"I underestimated. We now have three scaling laws. We have pre-training scaling. We have post-training—that's basically like AI practicing a skill until it gets it right. And then the third is inference. The old way of doing inference was one shot. The new way of doing inference is thinking. Think before you answer. The longer you think, the better the quality answer you get."
"我低估了。现在我们有三条 scaling law：预训练 scaling、post-training scaling（就是 AI 反复练习一门技能直到做对）、以及推理 scaling（旧的推理是 one-shot，新的推理是先思考再回答——你思考得越久，答案质量越高）。"
他对这件事今年的信心比去年更高，因为 agent 系统已经不再是单一模型，而是一整组模型并发运行、调用工具、做调研、生成多模态内容的复杂系统。
二、OpenAI Stargate：10 GW、4000 亿美金的"直签关系"
这是整期最重磅的一段。Brad 问："为什么投 1000 亿？" 黄仁勋先回答了最后一个问题：
"I think that OpenAI is likely going to be the next multi-trillion dollar hyperscale company. The opportunity to invest before they get there—this is some of the smartest investments we can possibly imagine. We don't have to invest. It's not required for us to invest. But they're giving us the opportunity to invest. Fantastic thing."
"我认为 OpenAI 很可能是下一个市值数万亿美元的超大规模云公司。能在他们到达那里之前投进去，这是我能想象到的最聪明的投资之一。我们并不必须投。他们给了我们这个机会，太棒了。"
然后他梳理了 NVIDIA 与 OpenAI 正在进行的四条并行 buildout：
1. Microsoft Azure 的持续建设，几年内几千亿美金的工程量；
2. OCI + SoftBank + OpenAI 的 5-7 GW；
3. CoreWeave 的全量；
4. 新的 Stargate — OpenAI 第一次自建 AI 基础设施。
"This new partnership is about helping OpenAI build their own self-build AI infrastructure for the first time. Us working directly with OpenAI at the chip level, software level, systems level, AI factory level. I mean, this is going to go on for some time."
他把 Stargate 的逻辑拆成两个指数的叠加：
"They're going through two exponentials. The first exponential is the number of customers is growing exponentially because the AI is getting better. The second exponential is the computational exponential of every use—instead of just one-shot inference, it's now thinking before it answers. These two exponentials compounding their compute requirements."
"他们正在穿越两个指数——第一个是用户数的指数增长，因为 AI 越来越好用；第二个是每次调用的计算量的指数增长，因为推理从 one-shot 变成了'先思考再回答'。两个指数在互相叠加。"
他把 OpenAI 和 Meta/Google/Azure 放到同一类 hyperscaler——"他们想要一种像 Elon Musk 和 X 那样的直签关系，也像 Zuck 和 Meta、Sundar 和 Google、Satya 和 Azure 那样的直签关系"，每一家都超大规模、都自建。
三、为什么华尔街是错的：三层论证
Brad 抛出一个很有趣的数据点——华尔街 25 位卖方分析师的共识预期认为 NVIDIA 从 2027 年开始增长会降到 8%。黄仁勋淡定地说："我们对这个预期很舒服，我们有能力稳定地跑赢。"
但他用三个层次论证为什么这个预期从根本上错了：
第一层——物理定律。 通用计算已死，未来是加速计算和 AI 计算。
"General purpose computing is over. The future is accelerated computing and AI computing. How many trillions of dollars of computing infrastructure in the world has to be refreshed?"
第二层——超大规模云的刷新。 这是他最关键的判断：超大规模云自己已经在从 CPU 迁到 GPU。搜索、推荐、TikTok、YouTube Shorts、Instagram——所有这些不是 AI 的"新用例"，而是本来就存在的百亿级流量被迁到了 AI 上。
"You can't do TikTok without AI. You can't do YouTube Short without AI. Meta's late getting to GPUs, search with GPUs, for sure, brand spanking new. Hundreds of billions of dollars just feeding the Metas, the Googles, the ByteDances, the Amazons—taking their classical hyperscaling and moving it to AI."
第三层——AI 作为"智能放大器"。 他做了一个非常清晰的"员工账本"：
"Suppose I were to hire a $100,000 employee, and I augmented that employee with a $10,000 AI, and that AI made the employee twice or three times more productive. Would I do it? Heartbeat I'm doing it across every single person in our company. Every software engineer, every chip designer already has AIs working with them. 100% coverage."
"假设我雇一个 10 万美元的员工，我再花 1 万美元给他配一个 AI，让他效率提高 2-3 倍。我会不会这么做？当然会。我们公司现在每一个人都已经配了 AI 同事——100% 覆盖。"
然后他把 NVIDIA 的故事套到世界 GDP 上：人类智能贡献了世界 GDP 的 55-65%，大约 50 万亿美元。如果其中有 10 万亿美元被 AI 增强，按 50% 毛利算，需要 5 万亿美元的年度基础设施投资——每年 5 万亿美元的 capex 是合理的。
"If you told me that on an annual basis the cap ex of the world was about $5 trillion, I would say the math seems to make sense. And that's kind of the future."
四、"我们已经站在 1 万亿美金 AI 收入门口了"
Brad 追问："2026 年 AI 收入我们估 1000 亿，2030 年要到 1 万亿——你觉得能做到吗？"
"Yes. And I would also say we're already there. Because the hyperscalers, they went from CPUs to AI. Their entire revenue base is all now AI driven. You can't do TikTok without AI. Correct. It's all of that stuff used to be humans creating four choices that are then selected by a recommender engine. And now it's infinite number of choices generated by an AI."
"可以。而且我还要说一句——我们已经到了。因为超大规模云厂商已经从 CPU 切到 AI 了。他们的整个收入基盘现在都是 AI 驱动的。以前是人创造 4 个候选、推荐引擎从中选；现在是 AI 创造无限个候选。"
Brad 问他未来会不会出现供给过剩（glut）。他说：
"Until we fully convert all general purpose computing to accelerated computing and AI, I think the chances are extremely low. That'll take a few years. And then until all recommender engines are AI based, until all content generation is AI based."
"直到所有通用计算都转成加速计算和 AI，出现过剩的概率都非常低。这要花几年。然后还要等所有推荐引擎和内容生成都变成 AI 的。"
关于订单节奏：NVIDIA 处在供应链末端，响应需求。过去几年他们深度打通供应链（wafer start、CoWoS、HBM），随时可以翻倍。客户每次给的预测都是错的——但是往小了错，"比去年显著增长，但还是不够"。
五、关于"循环收入"与"泡沫"的反驳
CNBC 上每天有人在喊"glut"、"bubble"、"round-tripping"（循环收入），Brad 把这个问题直接丢给他。黄仁勋的回答简单有力：
"Ten gigawatts is like $400 billion. And that $400 billion will have to be largely funded by their offtake—their revenues, which is growing exponentially. It has to be funded by their capital, the money they've raised through equity and whatever debt they can raise. Those are the three vehicles."
"10 GW 大约 4000 亿美元。这笔钱主要要靠他们自己指数级增长的收入来买单，加上他们自己融到的股权和债。这是三个来源。"
他的投资和他们的采购是完全分开的——"投资那一侧和任何事都没绑定，只是一个让我们能以早期价格进入未来多万亿美元公司的机会"。
"The equity and debt they could raise has something to do with the confidence of the revenues that they could sustain. Smart investors and smart lenders will consider all of these factors. Fundamentally, that's their company. It's not my business. My only regret is that they invited us to invest early on—we were so poor, we didn't invest enough. I should have given all my money."
Brad 顺势补了一句很关键的事实检查：每个月全球都有 15 亿人为 ChatGPT 付钱——"这不是把收入在两家公司之间倒腾，是真实的付费用户"。
六、Token 经济：Perf-per-Watt = 收入
这是整期最漂亮的一段推理。Brad 问："你说过就算竞争对手的 ASIC 免费送，客户也应该买 NVIDIA 的系统——这怎么可能？"
黄仁勋的回答是：所有人都是电力受限的。
"Everybody's power limited. Let's say you were able to secure two more gigawatts of power. You would want that to translate to revenues. So if your performance per watt was twice as high as somebody else's, you can produce twice as much revenues from your data center. Who doesn't want twice as much revenues?"
"所有人都是电力受限的。假如你多拿到 2 GW 电力，你当然想把它变成收入。如果我的 perf-per-watt 是别人的两倍，我的客户就能从同一个数据中心生出两倍的收入——谁不想要两倍的收入？"
更狠的一步：
"Blackwell is 30 times Hopper. Let's pretend somebody else's ASIC is Hopper. So you've got to give up 30 times revenues in that 1 gigawatt. It's too much to give up. So even if they gave it to you for free, your opportunity cost is so insanely high, you would always choose the best performance per watt."
"Blackwell 是 Hopper 的 30 倍。假如别人的 ASIC 相当于 Hopper，你就要在那 1 GW 里放弃 30 倍的收入——这代价太大了。就算他们免费送，你的机会成本也高到你必须选 perf-per-watt 最高的那家。"
这就是为什么"土地、电、外壳本身就要 150 亿美金"的情况下，芯片价格几乎不是主要变量。这是整个行业过去两年才真正内化的账。
七、什么是 Extreme Co-Design？一年一代的暴力美学
Brad 问："对普通人怎么解释 extreme co-design？" 黄仁勋说：
"Extreme co-design means that you have to optimize the model algorithm, system and chip at the same time. You have to innovate outside the box. Because Moore's law said you just have to keep making the CPU faster. Everything got faster. You were innovating within the box. Well, if that chip doesn't go any faster, then what are you going to do? Innovate outside the box."
"Extreme co-design 就是你得同时优化模型算法、系统和芯片——你必须在框外创新。因为摩尔定律告诉你只要让 CPU 更快就行，所有东西都会跟着快——那是框内创新。但如果那颗芯片不能更快了，你怎么办？框外创新。"
他给了一组震撼的数字：Blackwell 对 Hopper 是 30 倍——不是因为晶体管密度，而是因为他们同时重做了 CPU、GPU、网络芯片、NVLink scale-up、Spectrum-X scale-out、冷却、电源、机架——全部重新设计。他说摩尔定律连十分之一都给不了。
"Blackwell to Hopper is 30 times. No Moore's law could possibly achieve that. That's because NVIDIA got into networking and switching and scale up and scale out and scale across, building CPUs and building GPUs and building NICs. That's the reason why NVIDIA is so rich in software and people."
一年一代的节奏本身就是护城河——每年 6-7 颗芯片同步进化，整个系统每年推倒一次。他反问："谁会给一个首次流片的新架构下 500 亿美金的 PO？NVIDIA 可以，因为架构已经验证过，客户规模极大，供应链规模极大。"
八、ASIC vs GPU：三种芯片、两种生意
Clark 问到 Google TPU 和 Amazon Trainium 这些 ASIC 的威胁。黄仁勋的回答是三类芯片论：
"There are three categories of chips. First, architectural chips—x86 CPUs, ARM CPUs, NVIDIA GPUs. They have an ecosystem above and rich IP. Second, ASICs. I worked for the original company that invented ASICs—LSI Logic. LSI Logic is not here anymore. ASICs is fantastic when the market size is not very large. Third, COT—customer owned tooling. That's what Apple does for the iPhone chip. So where will TPUs go when it becomes a large business? Customer owned tooling. No question about it."
"芯片有三类：第一类是架构型（x86、ARM、NVIDIA GPU）——有生态、有丰富 IP；第二类是 ASIC——我当年在发明 ASIC 这个概念的 LSI Logic 工作过，那家公司今天已经不存在了。ASIC 在市场不大的时候很好用；第三类是 COT（客户自持工艺），Apple 的手机芯片就是这样。TPU 等到规模够大的时候会去哪里？毫无疑问——COT。"
然后他讲了 ASIC 的根本困境：你三五年前立项的时候，整个行业看起来只是"一颗 GPU"；今天它是整个 AI 工厂——需要 CPX（context processing）、长短期记忆、KV cache 处理器、视频转码……而 transformer 架构本身每六个月就在变。
"If not for the fact that CUDA is easy to operate on and iterate on, how do they try all of their vast number of experiments? CUDA helps you do all that because it's so programmable."
最尖锐的一句：
"That industry three, four, five years ago was super adorable and simple. Now it's giant and complex and in another two years it's going to be completely massive. The battle of getting into a very large market as a nascent player is just hard."
"三四年前那个行业非常可爱也非常简单。今天它已经庞大复杂，再过两年会巨大到让人窒息。作为一个小玩家想进入已经巨大的市场，是非常非常难的一件事。"
九、Elon Musk 与 Colossus：人脑作为系统集成者
Brad 问：为什么 Elon 能用四个月建起 Colossus，而大部分公司不行？黄仁勋说：
"These AI supercomputers are complicated things. The technology is complicated, procuring it is complicated because of financing issues, securing the land, power and shell, powering it, building it all, bringing it all up. This is unquestionably the most complex systems problem humanity has ever endeavored."
"这些 AI 超算是非常复杂的东西——技术复杂、融资复杂、拿地拿电拿外壳复杂、通电复杂、系统集成复杂。这毫无疑问是人类尝试过最复杂的系统问题。"
然后他说了一句让我反复回味的话：
"Elon has a great advantage that in his head, all of these systems are interoperating and the interdependencies reside in one head, including the financing."
"Elon 有一个巨大优势——所有这些系统、所有这些相互依赖，都在他一个大脑里同时跑，连融资也在。"
Brad 补了一句："他自己就是一台 GPT。" 黄仁勋说："He's the ultimate GPU."
十、主权 AI：每个国家都需要 AI 基础设施
Brad 把话题拉到地缘政治——"30 年前你很难想象今天你会和酋长、国王、白宫聊这些"。黄仁勋说：
"Nobody needs atomic bombs. Everybody needs AI. AI is modern software. We've reinvented computing. There's not a new species on Earth. We just reinvented computing. Everybody needs computing. There's nobody in the world that says, 'I used to use computers yesterday. I'm pretty good with clubs and fire tomorrow.' "
"没有人需要原子弹，但所有人都需要 AI。AI 就是现代版的软件。我们重新发明了计算——这地球上没有新物种，我们只是重新发明了计算。所有人都需要计算，世界上没有人会说'我昨天还在用计算机，明天我要回去用棍棒和火把'。"
他对"主权 AI"的立场是务实的——每个国家都应该用 OpenAI、Anthropic、Gemini、Grok，但也要有自己的基础设施去发展工业模型、制造模型、国家安全模型：
"In order to participate in AI, you have to encode within AI your history, your culture, your values. Every country needs to have some sovereign capability. They need to build it not just for language models, but for industrial models, manufacturing models, national security models."
他打了一个非常好的类比——AI 基础设施就像每个国家的能源基础设施、通信基础设施，是国家级必须品。
十一、中国：被迫放弃 95% 市场份额是"单方面缴械"
这是整期最锋利的一段。Brad 提到华为正在用"中国市场的垄断利润"发展、三年计划超过 NVIDIA。黄仁勋承认：
"95% market share. And the facts are just wrong. They are nanoseconds behind us. Not two years, three years. They're nanoseconds behind us."
"我们曾有 95% 的中国市场份额。关于中国的那些'事实'大多是错的——他们不是落后我们两三年，他们落后我们的是纳秒级。"
他给出了一个非常干练的判断：
"50% of the world's AI researchers are Chinese. They are a formidable, innovative, hungry, fast moving, under-regulated. People don't realize this. They are lightly regulated, less regulated ironically than we are in the capitalist system. Some of the things I heard—they could never build AI chips—that just sounded insane. They can't manufacture? If there's one thing they could do, it's manufacture."
"全世界 50% 的 AI 研究者是中国人。他们是一群可怕的、有创新力的、饥饿的、快速行动的、被轻监管的人——很多人没有意识到这一点，讽刺的是，他们比我们资本主义体系下的监管还要轻。那些'他们永远造不出 AI 芯片'的说法太疯狂了。'他们不会制造'？如果这个世界上有一件事他们会做，那就是制造。"
他的中国战略很清晰——竞争，而不是脱钩：
"The single best industry we have is our technology industry. Why would we not allow this industry to go compete for its survival, proliferate American technology around the world, maximize economic success, maximize geopolitical influence? I've never heard President Trump say the word 'decouple.' You can't decouple the two most important relationships of the next century. Decoupling is exactly the wrong concept."
"我们最强的产业就是科技产业。我们为什么不让这个产业去竞争、去求生存、去把美国技术扩散到全世界、去最大化我们的经济成功和地缘影响？我从来没有听 Trump 总统说过'脱钩'这个词——你怎么可能和未来 100 年里最重要的两对关系脱钩？脱钩就是错的概念。"
被问到"你只是想多卖芯片吧"，他直接回应：
"Just because I want America, ecosystem and economy to grow, doesn't make me wrong. Everything that has been said so far about China has proven to be wrong. The facts are just wrong."
十二、美国梦 与 Invest America
访谈的后半段转向一个很温暖的话题——Brad 推动的 "Invest America" 计划。这个想法是：每个在 2026 年之后出生的美国孩子，国家出 1000 美元为他开一个投资账户，投资于最好的美国公司。黄仁勋 已经决定 NVIDIA 不仅为员工的孩子追加资金，还会扩展到更多孩子。
他谈到自己就是美国梦的象征：
"My parents didn't have any money, sent us over here. We started from nothing. You guys know I bus tables, wash dishes, clean toilets, and here I am. This is the American dream. You go to Wikipedia, you can look up American dream. My picture."
"我父母没钱，把我们送到这里来。我们从一无所有开始。你们都知道我洗过盘子、收过桌子、擦过厕所——然后我站到了这里。这就是美国梦。你去 Wikipedia 查 American Dream，会看到我的照片。"
对于 H1B 签证 10 万美元收费的争议，他说"这是一个好的开始"，但明确：
"America has a singular brand reputation that no country in the world has. No country in the world is in the position to be able to say, 'Come to America and realize the American dream.' What country has the word 'dream' behind it? We are utterly singular."
十三、"China Hawk" 是耻辱，不是荣耀
这一段是我第一次听到他用这么重的词。Brad 提到一位在美国顶级实验室的华人 AI 研究员说：三年前 90% 的中国顶尖毕业生会来美国，今天只有 10-15%。黄仁勋说：
"This is precisely a concern that we have. This is a source of existential crisis. Definitely the early indicators of a future problem. Smart people's desire to come to America, smart students' desire to stay—those are KPIs."
然后他直接对"China Hawk"这个标签开火：
"There's a phrase, 'China Hawks.' If you're a China Hawk, you get to wear that label with pride. It's almost like a badge of honor. It's a badge of shame. Although they want what's in the best interest of our country, destroying that pipeline of the American dream is not patriotic. Not even a little bit."
"有个词叫 'China Hawks'。如果你是 China Hawk，你可以骄傲地戴着这个标签——但它其实是耻辱的勋章。尽管他们也想为这个国家好，但摧毁美国梦的输送管道不是爱国行为。一点都不是。"
"We need to have the confidence of a great country. And to somebody who wants to compete with us, have the attitude 'bring it on.' Bring it on. Because I believe in our people. I believe in our culture. I believe in our country. I believe in our system. Bring it on."
十四、加速进步：上车，别想着在未来某个路口拦截它
访谈尾声 Brad 引了 Ray Kurzweil 的话："21 世纪不是 100 年的进步，而是 20,000 年的进步。" 黄仁勋给了一个特别好的实操建议：
"If you have a train that's about to get faster and faster and go exponential, the only thing you really need to do is get on it. Once you get on it, you'll figure everything else out along the way. To predict where that train's going to be and try to shoot a bullet at it, or predict where that train's going to be and it's going exponentially faster every second and figure out what intersection to wait for it—that's impossible."
"如果一列火车即将越开越快、进入指数级加速——你真正要做的只有一件事：上车。一旦你上了车，剩下的你会在路上想明白。想预判这列车的位置再开一枪、或者预判在哪个路口拦截它——这是不可能的，它每一秒都在指数级变快。"
他预测未来 5 年会发生的一件"很酷的事"：AI 和机电一体化的融合。
"We're going to have AIs wandering around us. We all know that we're going to grow up with our own R2-D2—our own companion that remembers everything about us and coaches us along the way. Every human will have their own GPUs associated with them in the cloud. 8 billion people, 8 billion GPUs—that's a viable outcome."
"我们每个人都会陪着自己的 R2-D2 长大——一个记得我们一切、一路指导我们的陪伴者。每个人在云端都会有自己专属的 GPU。80 亿人，80 亿 GPU——这是完全可能的结局。"
结尾他对 AI 与工作的看法依然是那个乐观的乘数逻辑：
"Intelligence is not a zero-sum game. The more intelligent people I'm surrounded by, the more geniuses I'm surrounded by, surprisingly, the more ideas I have, the more problems I imagine that we can go solve. The more work we create, the more jobs we create."
"智能不是零和游戏。我身边的聪明人和天才越多，奇怪的是我自己的想法反而越多，想解决的问题越多，创造的工作越多，雇的人越多。"
然后他对 Brad 说："Don't scare them, bring them along."（别吓到他们，带他们一起走。）
金句集
- "I underestimated it. Inference is going up by a billion times." — 关于三条 scaling law
- "OpenAI is likely going to be the next multi-trillion dollar hyperscale company."
- "$5 trillion of annual global capex—the math seems to make sense."
- "We're already there." — 关于 1 万亿美金 AI 收入
- "Even if somebody's ASIC is free, opportunity cost is too high. You'd always choose the best perf per watt."
- "Blackwell is 30 times Hopper. No Moore's law could possibly achieve that."
- "Nobody needs atomic bombs. Everybody needs AI."
- "China is nanoseconds behind us."
- "China Hawk is a badge of shame, not honor."
- "Get on the train. You'll figure everything else out along the way."
- "Intelligence is not a zero-sum game."
原文出处：BG2 Pod — NVIDIA CEO Jensen Huang on OpenAI, Future of Compute, and the American Dream (2025-09-26)