Jensen Huang: Will Nvidia's Moat Persist?
黄仁勋：英伟达的护城河，还能扛多久？
A 103-minute conversation with Dwarkesh Patel — on the electrons-to-tokens thesis, the five-layer AI stack, why Anthropic is a one-off (not a trend), the 50× Blackwell story, and the China debate Jensen refuses to concede.
黄仁勋与 Dwarkesh Patel 的 103 分钟深谈——从「电子换 token」的核心思路，到 AI 五层蛋糕；从 Anthropic 为什么是孤例、不是趋势，到 Blackwell 50 倍的秘密，再到那场关于中国、他半步不让的激辩。
The input is electrons, the output is tokens. In the middle is Nvidia.
输入是电子，输出是 token。中间那一层，是英伟达。
— Jensen Huang
— 黄仁勋
⚡ Electrons → Tokens ▶ 00:00
⚡ 电子换 token ▶ 00:00
Dwarkesh opens with the skeptic's case — Nvidia is "software that other people manufacture." Jensen's reframe runs the rest of the episode: something has to turn electrons into tokens, and making a token more valuable than the next one is a journey "far from deeply understood."
Dwarkesh 一开场就把质疑者的剧本摆上桌——"英伟达本质上是在写软件，别人在制造。"黄仁勋的反问贯穿了整场访谈：到最后，总得有人把电子变成 token，而让一个 token 比另一个更值钱，是一场"远远没走完、甚至还没被真正理解清楚"的旅程。
Nvidia sends a GDS2 file — someone else builds everything.
英伟达发出一个 GDS2 文件，其余都是别人造的。
Dwarkesh opens with the skeptic's case. Nvidia sends a GDS2 file to TSMC. TSMC builds the dies. SK Hynix, Micron, and Samsung make the memory. An ODM in Taiwan assembles the racks. "Nvidia is fundamentally making software that other people are manufacturing." If software gets commoditized, does Nvidia? Valuations across software are already crashing on that bet.
Dwarkesh 先把质疑者的逻辑摆上桌：英伟达把 GDS2 文件发给台积电，台积电做出芯片；SK 海力士、美光、三星提供 HBM；台湾的 ODM 厂把整机柜装好。"英伟达本质上是在写软件，而别人负责制造。"如果软件被 AI 大规模商品化，英伟达会不会也一起被商品化？整个软件板块的估值，已经在替这个猜想投票。
Making one token more valuable than another is "far from deeply understood."
让一个 token 比另一个更值钱，是一场"还没走完"的旅程。
Jensen's reframe: "In the end, something has to transform electrons to tokens. The transformation from electrons to tokens, and making those tokens more valuable over time, is hard to completely commoditize." Making one token more valuable than another, he argues, is an "incredible journey" of artistry, engineering, science, and invention that is "far from deeply understood."
黄仁勋的反问：到最后，总得有人把电子变成 token。"把电子变成 token、把这个 token 变得比那个 token 更有价值——这件事很难被彻底商品化。"让一个 token 比另一个更值钱，需要工艺、工程、科学、发明的叠加——他说，这场"旅程远远没走完，甚至还没被真正理解清楚。"
"As little as possible" means anything Nvidia doesn't do, someone in the ecosystem does.
"能不做就不做"——不碰的那些，都交给生态合作伙伴。
"What I mean by 'as little as possible' — whatever I don't need to do, I partner with somebody and make it part of my ecosystem." Nvidia runs the largest partner ecosystem in AI, upstream and downstream. The parts Jensen does keep in-house — NVLink, CUDA, extreme co-design — are the parts where "if we don't do it, nobody else would."
这是他反复强调的一句话。"不该碰的"——交给生态合作伙伴。英伟达的合作伙伴网络是 AI 行业最大的，从上游的台积电、HBM 三巨头，到下游的云厂商、应用开发者，五层全覆盖。"该做的"——NVLink、CUDA、极致协同设计这些硬骨头——"如果我们不做，没人会做"。
Agents won't kill Cadence and Synopsys — they'll cause Synopsys instances to "skyrocket."
Agent 不会干掉 Synopsys，它会让 Synopsys 的实例数爆表。
The market thinks agents will kill tool vendors like Cadence and Synopsys. Jensen sees the opposite. "I think the number of agents is going to grow exponentially, and the number of tool users is going to grow exponentially. The number of instances of Synopsys Design Compiler is going to skyrocket." Agents don't replace tools — they multiply them. "The reason it hasn't happened yet is the agents aren't good enough at using tools yet."
市场的判断是：AI Agent 会干掉 Cadence、Synopsys 这些工具厂。黄仁勋的判断正好相反。"Agent 的数量会指数级增长，工具使用者的数量也会指数级增长。Synopsys Design Compiler 的实例数，会爆表。"Agent 不是替代工具，而是让每个工具被调用的次数暴涨。"现在还没发生，是因为 Agent 还不够会用工具。等它们会用了——全面开花。"
Our job is to do as much as necessary and as little as possible.
我们要做的，是该做的都做到，不该碰的一概不碰。
— Jensen Huang
— 黄仁勋
🏗️ The Moat Is the Supply Chain ▶ 02:50
🏗️ 真正的护城河，是供应链 ▶ 02:50
Jensen's mental model of AI is a five-layer cake — energy, chips, systems, models, applications. Every layer has to succeed. The policy implication runs throughout the conversation: conceding any one layer threatens the whole stack.
黄仁勋眼中的 AI 栈，是一块五层蛋糕——能源、芯片、系统、模型、应用，每一层都必须赢。整场访谈的政策观点，都从这个结论往回推：让出哪一层，整座大厦都会跟着晃。
The five-layer AI cake: Energy → Chips → Systems → Models → Applications.
AI 五层蛋糕：能源 → 芯片 → 系统 → 模型 → 应用。
Jensen's mental model of AI: five layers stacked from the ground up. Energy → Chips → Systems → Models → Applications. Every layer has to succeed. "We want the United States to win every single layer, including the chip layer." The policy implication runs throughout the conversation: conceding any one layer threatens the whole stack.
黄仁勋眼中的 AI 栈，从下往上五层：能源 → 芯片 → 系统 → 模型 → 应用。每一层都必须赢。"我们希望美国在每一层都赢，芯片层也不例外。"这句话不是客套——整场访谈的政策观点，都从这个结论往回推。
$250B in supply-chain obligations — and that's before partners' implicit investments.
供应链义务约 2500 亿美元——还没算上游伙伴的隐性投入。
Nvidia's latest filings show close to $100 billion in purchase commitments with foundries, memory makers, and packagers. SemiAnalysis estimates the total obligation closer to $250 billion. That's before partners' implicit upstream investments — HBM capacity, CoWoS packaging, EUV machines, silicon photonics — built because Jensen personally informed, inspired, and aligned supplier CEOs with his forecast. It's not just contracts. It's a decade of "let me reason through this with you."
最新财报显示，英伟达与晶圆厂、存储、封装的采购承诺已接近 1000 亿美元；SemiAnalysis 估算加总后的义务规模大约 2500 亿美元。这还没算上游伙伴的隐性投入——HBM 产能扩张、CoWoS 封装、EUV 机台、硅光器件——这些投资之所以敢下，是因为黄仁勋这些年挨个跑去和上游 CEO 谈："让我告诉你这个行业会有多大，让我和你一起推导一遍。"不是合同，是十年的信任账户。
No bottleneck lasts more than 2–3 years — except plumbers and electricians.
没有瓶颈能撑过两到三年——除了水管工和电工。
"None of the bottlenecks last longer than a couple of years — two, three years, none of them." CoWoS packaging was the crisis of 2023. Nvidia "swarmed the living daylights out of it" and it became mainstream. HBM was specialty. Now it's standard. Silicon photonics was a gap — Nvidia invested in Lumentum and Coherent and built a supply chain with TSMC. The only bottleneck that persists, Jensen says, is plumbers and electricians. Energy. Labor. Everything else gets solved.
"没有哪个瓶颈能撑过两到三年。"2023 年的 CoWoS 封装危机？英伟达"拿全部火力扑上去"，如今已是主流。HBM 曾是小众，现在标配。硅光缺口？他们投资 Lumentum、Coherent，和台积电一起做 COUPE。真正两三年解决不了的瓶颈只有一个——水管工和电工。能源、劳动力、电网。硬件问题都能解。
Three loops — install base, ecosystem richness, versatility — each feeding the others.
三条轨道——装机量、生态丰度、无处不在——互相咬合越转越快。
Three reinforcing loops drive the flywheel. Install base: hundreds of millions of CUDA GPUs deployed across every cloud, every edge, every robot. Ecosystem richness: every framework runs on CUDA first — Triton, vLLM, SGLang, verl, NeMo RL; even custom frameworks are built with Nvidia engineers. Versatility: Nvidia runs on Google Cloud, AWS, Azure, OCI, on-prem — anywhere a customer wants compute. Each loop feeds the others.
飞轮的三条轨道互相咬合，越转越快：装机量——几亿张 CUDA GPU 已经部署，每一朵云、每一台边缘设备、每一台机器人；生态丰度——所有框架首发 CUDA：Triton、vLLM、SGLang、verl、NeMo RL，连定制框架也是英伟达工程师"亲自下场"帮着调出来的；无处不在——Google Cloud、AWS、Azure、OCI、自建，客户在哪，英伟达就在哪。
AI is a five-layer cake. We have ecosystems across the entire five layers.
AI 是一块五层蛋糕。每一层都有我们的生态。
— Jensen Huang
— 黄仁勋
🏎️ GPUs vs ASICs — The Race Isn't What You Think ▶ 15:00
🏎️ GPU 对 ASIC——这场竞赛的形状，和你想的不一样 ▶ 15:00
Dwarkesh presses on TPUs and custom silicon. Jensen reframes the premise — Nvidia builds accelerated computing, not a tensor processor — and argues that programmability, not matrix multiplies, is what makes AI advance.
Dwarkesh 在 TPU 和定制芯片上步步紧逼。黄仁勋索性把前提推翻——英伟达做的是"加速计算"，不是张量处理器——他的论点是：让 AI 一路突飞猛进的，不是矩阵乘法，而是可编程性。
Nvidia built accelerated computing, not a tensor processor — market reach is far wider.
英伟达做的是加速计算，不是张量处理器——市场半径大得多。
When Dwarkesh points out that Claude and Gemini were trained on Google TPU, Jensen pushes back on the premise. "What Nvidia built is accelerated computing, not a tensor processing unit." CUDA accelerates molecular dynamics, quantum chromodynamics, fluid dynamics, particle physics, data processing, AND AI. "Our market reach is far greater than any TPU or ASIC can possibly have." TPUs optimize for one workload. GPUs absorb every workload that emerges.
Dwarkesh 指出 Claude 和 Gemini 都是在 Google TPU 上训练的。黄仁勋不接这个前提。"英伟达做的是加速计算，不是张量处理器。"CUDA 能加速的东西远不止 AI——分子动力学、量子色动力学、流体力学、粒子物理、结构化/非结构化数据处理。"我们的市场半径，比任何 TPU 或 ASIC 都大得多。"TPU 为一种工作负载优化；GPU 吞下每一种新冒出来的工作负载。
Programmability lets tomorrow's algorithms emerge — ASICs bake in today's.
可编程性让明天的算法还能长出来——ASIC 把今天的焊死在硅里。
Matrix multiplies aren't the whole story. When researchers invent a new attention mechanism, a hybrid SSM, a fused diffusion-autoregressive model — they need a programmable architecture. "The ability to invent new algorithms is really what makes AI advance so quickly." ASICs bake in today's workload. GPUs let tomorrow's workload emerge.
矩阵乘法不是 AI 的全部。当研究者想发明新的 attention 机制、混合 SSM、diffusion+自回归的融合架构——他们需要的是一个通用可编程的底座。"算法创新的能力，才是 AI 进化这么快的真正原因。"ASIC 把今天的工作负载焊死在硅里；GPU 让明天的工作负载还能长出来。
GPU vs ASIC: you're not saving what you think you're saving.
GPU 对 ASIC：你以为自己省了很多钱，其实没省多少。
Jensen's point: you're not saving what you think you're saving. A GPU handles every workload, runs every framework, and can be operated by anyone with a cluster. An ASIC wins on one workload, locks you into vendor tooling, and forces a respin when algorithms shift. Margins tell the story — GPU roughly 70%, ASIC roughly 65%. The delta disappears; the flexibility cost does not.
黄仁勋的潜台词：你以为省了很多，其实没省多少。GPU 什么活都接、所有框架都跑、任何有集群的人都能用；ASIC 只赢在一种工作负载，绑死一家工具链，算法一变就要重新流片。毛利率的差距——GPU 约 70%，ASIC 约 65%——会被灵活性成本吃掉，而灵活性本身却不会消失。
Hopper → Blackwell: 50× energy efficiency — transistors alone were only 75% better.
Hopper → Blackwell：能效 50 倍——晶体管本身只贡献了 75%。
Hopper to Blackwell: 50× more energy-efficient. Jensen originally announced 35× — Dylan Patel later wrote Nvidia had sandbagged the number. The transistors alone? Only ~75% better. "Moore's Law is dead." The other 30× came from architecture: MoE parallelization, disaggregation, offloading compute into the NVLink fabric and the Spectrum-X network, and CUDA extensible enough to rewrite kernels for the new shape of the workload.
Hopper 到 Blackwell，能效提升 50 倍。黄仁勋最初官宣 35 倍，后来 Dylan Patel 撰文指出英伟达"藏了实力"——其实是 50 倍。那 50 倍怎么来的？晶体管本身只贡献了约 75%——"摩尔定律已死"。剩下那 30 多倍，全是架构红利：MoE 并行化、拆分、把计算下放到 NVLink 织物和 Spectrum-X 网络，再加上 CUDA 灵活到能给新工作负载重写 kernel。
Anthropic is a one-off — 100% of TPU and Trainium growth traces to one lab.
Anthropic 是特例——TPU 和 Trainium 的增长，100% 来自这一家实验室。
"Anthropic is a unique instance, not a trend. Without Anthropic, why would there be any TPU growth at all? It's 100% Anthropic." Why did Anthropic go to TPUs and Trainium? Because Google and AWS could each put $5–10 billion into the lab in exchange for compute. "We just weren't in a position to do that at the time. That was my miss." Nvidia has now invested roughly $30B in OpenAI and $10B in Anthropic. "I'm not going to make that same mistake again."
"Anthropic 是特例，不是趋势。没有 Anthropic，TPU 的增长从哪来？100% 来自 Anthropic。"Anthropic 为什么去了 TPU 和 Trainium？因为 Google 和 AWS 各拿得出 50 亿到 100 亿美元注资，换它们用自家算力。"当时我们还没到能这么砸钱的阶段。那是我的失算。"现在，英伟达向 OpenAI 投了约 300 亿美元，向 Anthropic 投了约 100 亿美元。"这种错误，我不会再犯第二次。"
GPU is an F1 car — anyone gets to 100 mph, but 3× speed-ups need Nvidia engineers on-site.
GPU 是 F1——任何人都能开到 100 迈，但 3 倍加速需要英伟达工程师驻场。
A CPU is a Cadillac — smooth cruiser, everyone drives it, never goes too fast. Nvidia's GPU is an F1 car. Any hyperscaler can get it to a hundred miles an hour. But it takes Nvidia's own engineers, embedded in the labs, to push it to the limit. "It's not unusual that by the time we're done optimizing their stack, their model sped up by 3×, 2×, 50%." On a fleet of Hoppers and Blackwells, that's billions in additional revenue extracted from the same silicon.
CPU 是凯迪拉克——平顺、人人会开、永远不飙车。英伟达的 GPU 是 F1——任何大厂都能把它开到 100 迈。但要榨到极限，必须靠英伟达自己的工程师驻场。"我们帮客户调完栈，他们的模型常常快 2 倍、3 倍、50%。"放到 Hopper 和 Blackwell 的整个装机量上——这就是同一块硅，多挖出几十亿美元营收。
Nvidia's GPUs are like F1 racers. Everybody can drive it at a hundred, but it takes expertise to push it to the limit.
英伟达的 GPU 是 F1 赛车，CPU 是凯迪拉克——人人能开到 100 迈，但只有行家能把它推到极限。
— Jensen Huang
— 黄仁勋
🕰️ The Long Arc ▶ 45:00
🕰️ 长期主义的时钟 ▶ 45:00
The roadmap clock, the FIFO allocation rule, and the handshake with TSMC that has nothing written down. Jensen's read: the entire AI supply chain holds together because two multi-decade institutions just show up every year.
路线图这只钟、先到先得的分配规则，以及和台积电那只没有合同的握手。黄仁勋的判断是：整条 AI 供应链之所以能扛住这轮狂奔，就是因为两家跨了几十年的机构——年年都按时出现。
No highest-bidder auctions — allocation is FIFO and prices don't move with demand.
不玩价高者得——先下单先拿货，需求爆表也不涨价。
Jensen rejects the idea that Nvidia plays favorites or auctions scarce supply. "We never do that. If I quoted you a price, we quoted you a price. If demand goes through the roof, so be it." Allocation is first-in-first-out. Adjustments happen only when a customer's data center isn't ready to receive racks. "I prefer to be dependable. To be the foundation of the industry. You don't need to second-guess."
黄仁勋否认英伟达会给稀缺供给开"高价者优先"的口子。"我们从来不这么干。我给你报的价就是价，不会因为需求爆表就涨。""先下单先拿货"是唯一规则。唯一的例外是客户数据中心还没准备好——机柜发过去也没地方放，就先服务下一个。"我宁可做一家可以被依赖的公司。你不需要揣摩我是什么意思。"
Larry and Elon never "begged" — they just had to place an order.
拉里和马斯克从没"求"过 GPU——他们只是下了订单。
An article once claimed Larry Ellison and Elon Musk "begged" Jensen for GPUs over dinner. Jensen on the record: "We absolutely had dinner. It was a wonderful dinner. At no time did they beg for GPUs. They just had to place an order. Once they place an order, we do our best to get the capacity to them. We're not complicated."
曾有报道称拉里·埃里森和马斯克在晚宴上"跪求" GPU。黄仁勋上了节目直接澄清："我们确实一起吃了那顿饭，吃得挺开心。但他们从头到尾没求过任何 GPU。他们只需要下订单。单子一下，我们尽最大努力把产能送过去。我们不搞那些复杂的。"
Vera Rubin, Ultra, Feynman — every year a new architecture, every year a 10× token-cost drop.
Vera Rubin、Ultra、Feynman——每年一代架构，每年 token 成本掉一个数量级。
Vera Rubin this year. Vera Rubin Ultra next year. Feynman the year after. "The year after that, I haven't introduced the name yet." Every single year, a major architecture step — and roughly a 10× drop in token cost. "You're going to have to go find another ASIC team where you can bet the farm that they will be here for you every single year." No one else can make that claim. Not even TSMC, except for Nvidia.
今年 Vera Rubin，明年 Vera Rubin Ultra，后年 Feynman。"再后面那一代，名字我还没公布。"每年一代架构，token 成本差不多掉一个数量级。"你去别处找一家 ASIC 团队，能让你把整个家当押上去、相信它每年都如约而至的——几乎没有。"能做这种承诺的，除了英伟达，就是台积电。
No legal contract with TSMC in 30 years — just rough justice and two institutions that show up.
和台积电合作近三十年，没有法律合同——靠的是粗糙的公道和两家年年都到场的机构。
Nvidia and TSMC have been in business for nearly 30 years. "Nvidia and TSMC don't have a legal contract." Everything runs on rough justice. Sometimes Jensen gets a better deal; sometimes a worse one. "But overall, the relationship is incredible. I can completely trust them. I can completely depend on them." Jensen argues this trust is why the entire AI supply chain holds together: two multi-decade institutions who just show up every year.
英伟达和台积电合作快三十年了。"我们之间没有法律合同。"全靠"粗糙的公道"——黄仁勋有时候拿到更好的条件，有时候稍亏。"但总的来说，这段关系无比坚固。我可以百分百信任他们。"他的论点是：整条 AI 供应链之所以能扛住这轮狂奔，就是因为两家跨了几十年的机构——年年都按时出现。
Every single year you can count on us. Your token cost will decrease by an order of magnitude every single year.
我们每年都如约而至。你的 token 成本，一年降一个数量级——准得像钟。
— Jensen Huang
— 黄仁勋
🌏 The China Debate ▶ 57:00
🌏 关于中国的那场辩论 ▶ 57:00
The hardest section of the interview. Jensen rejects the "loser premise" — 7nm is Hopper, architecture beats lithography, and the telecom industry is the historical warning he keeps pointing at. He's not asking to send everything to China. He's asking for the messy middle.
这是全场最针锋相对的一段。黄仁勋拒绝"输家前提"——7nm 就是 Hopper、架构打败光刻，而他反复举例的历史教训，是美国电信业。他并不主张"什么都往中国送"，他要的是那条中间路径。
50% of AI researchers are Chinese, 40% of the global tech industry is in China.
50% 的全球 AI 研究者是华人，全球科技产业 40% 在中国。
50% of the world's AI researchers are Chinese, per Jensen. 60%+ of mainstream chips are manufactured in China. China is the second largest computing market in the world — about 40% of the global tech industry. "They have plenty of energy. They have plenty of chips. They have most of the AI researchers. The question is: what is the best way to create a safe world?"
50% 的全球 AI 研究者是华人，60% 以上的主流芯片在中国生产。中国是全球第二大计算市场，大约占全球科技产业的 40%。"他们能源充足，芯片充足，研究者也充足。那问题就变成了——要让世界变得更安全，最聪明的打法是什么？"
7nm is Hopper — the dangerous-model threshold was already crossed inside China.
7nm 就是 Hopper——能训出"危险模型"的算力门槛，中国早已跨过。
Dwarkesh argues SMIC is stuck at 7nm while Nvidia moves to 3nm and 1.6nm. Jensen: irrelevant. "Today's models are largely trained on Hopper, Hopper generation. 7nm chips are plenty good." China has abundant energy — enough that they can simply gang more chips together to compensate for density. "If your amount of watts is completely abundant, what do you care about performance per watt for?" The Hopper-class threshold for dangerous models, Jensen says, has already been crossed inside China.
Dwarkesh 强调中芯国际卡在 7nm，而英伟达要去 3nm、1.6nm。黄仁勋：这不重要。"今天世面上的模型，大多数是在 Hopper 这一代训练出来的。7nm 对他们够用了。"中国的能源极其充沛——芯片密度不够，就多插几张、把墙插满。"如果你的电是免费的、无限的，性能每瓦还有那么重要吗？"他的结论是：能训练出"危险模型"的 Hopper 级算力门槛，中国早就跨过去了。
"The day DeepSeek comes out on Huawei first — that is a horrible outcome for our nation."
"DeepSeek 第一次发布在华为上的那天——对我们国家来说是个极糟的结局。"
Dwarkesh's worst case: Chinese labs find zero-days first using models trained on Chinese chips. Jensen's worst case: DeepSeek — or the next DeepSeek — ships optimized for Huawei instead of Nvidia. "The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation." Today, the world's open-source models run on CUDA. Tomorrow, if they're born on Huawei, the American tech stack becomes the second-class citizen — in Africa, the Middle East, Southeast Asia, the global South.
Dwarkesh 最担心的是：中国实验室靠国产算力先一步找到零日漏洞。黄仁勋担心的版本不一样——"DeepSeek 第一次发布时，是在华为上跑，而不是在英伟达上。这对我们国家是个极糟的结局。"今天全球开源模型首发 CUDA；明天如果它们首发在华为上——美国技术栈就变成了非洲、中东、东南亚、全球南方的二等公民。
Architecture beats lithography — CS delivers 10× per year, Moore delivers 25%.
架构打败光刻——计算机科学每年 10 倍，摩尔定律每年 25%。
"Blackwell is 50× Hopper. Is the lithography 50× more advanced? Not even close — 75%." Most of the gain came from architecture, networking, numerics, system design. Most of AI's gains over the last decade came from algorithms, not transistors. Moore's Law delivers 25% per year. Computer science delivers 10× per year. "China's army of AI researchers is their fundamental advantage. We see it. DeepSeek is not an inconsequential advance."
"Blackwell 是 Hopper 的 50 倍。光刻先进了 50 倍吗？差远了——只提升了 75%。"其余全靠架构、网络、数值、系统设计。过去十年 AI 的进步，绝大部分来自算法而非晶体管。摩尔定律一年 25%；计算机科学一年能挤出 10 倍。"中国那一大群 AI 研究者，才是他们的根本优势。我们看见了。DeepSeek 不是小事。"
Telecom got "policed out of the world" — the chip layer could be next.
电信业被"赶出了全世界"——芯片层可能是下一个。
"The policies you're advocating resulted in the American telecommunications industry being policed out of basically the world. We don't control our own telecommunications anymore." Jensen's read: export restrictions designed to slow China ended up starving US firms of scale, gifting the global market to Huawei-class competitors, and leaving America dependent on foreign infrastructure. His warning is that the chip layer is next if the same logic is applied. "We are not a car. Ecosystems are hard to replace. x86 is sticky. ARM is sticky."
"你提倡的这套政策，结果是把美国电信产业赶出了大半个世界。今天我们连自己的通信基础设施都不能完全掌控。"黄仁勋的解读：所谓"掐住中国"的出口限制，反而把美国企业的规模饿瘪，把全球市场拱手送给华为级的对手，让美国在基础设施上反被卡脖子。下一个被送走的层，会不会就是芯片？"电脑不像车。生态不是说换就换——x86 粘，ARM 粘。"
Best tech stays home first — and Nvidia competes globally. Both can happen simultaneously.
最好的技术留在美国且优先——同时在全球赢。这两件事可以同时做。
Not "send everything to China at all times." Nobody argues that. Jensen's position is the messy middle: "We should always have the best technology here. We should always have the most technology here, and the first. But we should also try to compete and win around the world. Both of those things can simultaneously happen." The alternative he rejects is absolutism. "Life is not absolutes."
不是"什么都往中国送"——没人主张这个。他坚守的是那条中间路径："最好的技术必须留在美国，并且先给美国。但我们同时要在全球市场竞争、赢下来。这两件事可以同时做。"他拒绝的是绝对主义——非此即彼、零和对决。"生活不是非 0 即 1。"
You're not talking to somebody who woke up a loser. That loser attitude, that loser premise, makes no sense to me.
你面前站的，不是一个醒来就认输的人。那种输家心态、输家前提——我没法接受。
— Jensen Huang
— 黄仁勋
🔭 Beyond AI ▶ 1:35:00
🔭 不止 AI ▶ 1:35:00
The closing stretch: why Nvidia doesn't run parallel wafer-scale bets, why "premium tokens" just became a market, and the counterfactual Jensen answers without a pause — accelerated computing, the same thing we've been doing all along.
节目尾声：为什么英伟达不同时押几条晶圆级路线、"溢价 token"为什么突然成了一个市场，以及那个反事实——黄仁勋几乎没停顿就给出了答案：加速计算，他们一直做的那件事。
"We could diversify — it's just that we don't have a better idea. We simulate it all — provably worse."
"我们可以多线并进——问题是没有更好的想法。都仿真过了，结果更差。"
Dwarkesh: with your cash, why not also run a Cerebras-style wafer-scale project, a Dojo-style mega-package, a non-CUDA architecture? Diversify. Jensen: "We could. It's just that we don't have a better idea. We simulate it all — provably worse. We're working on exactly the projects we want to work on." The one exception: when the workload changes shape — not the algorithm, the market for the algorithm — they adapt. Recent example: Nvidia folded Groq into the CUDA ecosystem for ultra-low-latency inference.
Dwarkesh 问：你们现金那么多，为什么不同时干一条 Cerebras 式的晶圆级路线？一条 Dojo 式的大封装？一条绕开 CUDA 的架构？分散一下风险。黄仁勋："我们可以做——问题是没有更好的想法。都在仿真里跑过，结果更差。我们做的，就是我们确实想做的那些。"唯一会偏离主线的情况，是工作负载本身换了形状——不是算法变，是市场要的东西变。最近就有一例：他们把 Groq 融进了 CUDA 生态，专做超低延迟推理。
Premium tokens just became a market — lower throughput, higher ASP, Groq's segment.
溢价 token 突然成了一个市场——低吞吐、高 ASP，这是 Groq 的地盘。
Until recently, tokens were nearly free. Now software engineers, agents, and other high-value users justify paying more for faster tokens. "If I can give them much more responsive tokens so they're even more productive, I would pay for it." That opens a second inference tier: lower throughput, higher ASP (average selling price). Groq slots into that segment. The inference market is disaggregating — not one market, but a spectrum from commodity to premium.
从前 token 几乎免费。现在软件工程师、Agent、那些高价值用户开始愿意为更快的 token 多付钱。"如果我能给他们响应更快的 token，让他们生产力再高一截，我会为此埋单。"这就开辟出推理的第二档：吞吐低一些，但 ASP（平均售价）高得多。Groq 就是冲这个段来的。推理市场正在分层——不再是一个市场，而是从低端到顶级的一整条光谱。
Without deep learning, Nvidia would still be very large — general-purpose computing ran its course.
没有深度学习，英伟达仍会很大——通用计算已经走到头了。
Final question: if deep learning hadn't ignited in 2012, where would Nvidia be? Jensen, without hesitation: accelerated computing — the same thing we've been doing all along. Molecular dynamics. Seismic processing for energy. Computational lithography. Quantum chemistry. Particle physics. Data processing. "If AI didn't exist, Nvidia would be very, very large. The reason is fundamental: general-purpose computing has largely run its course. The way forward is domain-specific acceleration." AI isn't the mission. AI is what accelerated computing unlocked.
最后一个问题：如果 2012 年深度学习根本没点着，英伟达会在哪里？黄仁勋几乎没停顿：加速计算——我们一直做的那件事。分子动力学。地震波反演。计算光刻。量子化学。粒子物理。数据处理。"即便没有 AI，英伟达也会很大。原因很根本：通用计算已经走到头了，往前的出路只有一条——领域专用加速。"AI 不是公司的使命，AI 是加速计算点燃的那把火。
If deep learning didn't happen, Nvidia would still be very, very large.
就算深度学习没发生，英伟达也会很大、很大。
— Jensen Huang
— 黄仁勋