The problem doesn't fit inside one computer anymore. To run a frontier model, you have to shard the algorithm, shard the pipeline, shard the data, shard the model itself across thousands of machines. The moment you do that, the network becomes the bottleneck, the CPU becomes the bottleneck, the cooling becomes the bottleneck. Specializing one component while ignoring the rest doesn't help.

今天的问题，再也塞不进一台机器里靠一颗 GPU 加速跑完。要训出一个前沿模型，你得把算法切片、把流水线切片、把数据切片、把模型本身切片到几千台机器上去。一旦这么干，瓶颈立刻换地方——网络成了瓶颈、CPU 成了瓶颈、散热成了瓶颈。光优化一颗芯片，没用。

Jensen's point: scale stops being additive. You need 10,000 computers to go a million times faster — and getting that compounding is a co-design problem, not a chip problem.

黄仁勋的核心论断：规模不再是简单加法。一万台机器要换出百万倍速度——这种「指数级复利」不是芯片问题，是协同设计问题。

Amdahl's Law: total speedup is capped by the slowest fraction of the workload. If computation is 50% of what your system does and you make the compute infinitely fast, you've doubled total throughput. That's the ceiling.

Amdahl 定律一句话：整体提速，被你最慢那块拖住。如果计算只占整体工作量 50%，就算你把计算搞到无限快，整体也只能翻一倍。这就是天花板。

Which is why making the GPU faster every year, alone, isn't the strategy. Memory has to keep up. Networking has to keep up. The software stack has to be co-designed with the hardware. Otherwise the rest of the system becomes the workload, and you're back to Moore's-Law-shaped progress.

所以光让 GPU 一年比一年快没用。内存得跟上，网络得跟上，软件栈得跟硬件一起设计。否则，剩下的部分就成了主要工作量——你又被踢回了「摩尔定律节奏」。

Jensen has 60 direct reports. He doesn't do one-on-ones. Every problem gets discussed in front of every relevant expert simultaneously — memory, GPU, optical, cooling, software — and whoever should have caught it gets called in. The company looks the way the product looks: extreme co-design, all the way up.

黄仁勋有 60 个直接下属。他不开 1-on-1。每个问题都摆在所有相关专家面前一起讨论——内存的、GPU 的、光通信的、散热的、软件的——谁该接住没接住，立刻被点名。公司的样子，就是产品的样子：极致协同设计，从下往上贯穿到顶。

Most companies use the same generic org chart whether they make hamburgers or cars. Jensen's argument is that the company should be shaped like what it produces.

绝大多数公司的组织架构图都长得一样，不管你做的是汉堡还是汽车。黄仁勋的反问：公司应该长得像它生产的东西。

Putting CUDA on every GeForce raised the cost of a consumer chip by ~50% on a 35% gross-margin business. NVIDIA's market cap fell to roughly $1.5 billion. Gamers paid for it without using it. The board signed off on a strategic loss with a payoff that was a decade away — populating an install base of CUDA-capable GPUs in every university lab, every gaming PC, every research cluster.

把 CUDA 塞进每一颗 GeForce，硬生生把消费级芯片成本拉高了 50%——而那是一家毛利只有 35% 的公司。NVIDIA 市值跌到大约 15 亿美元。游戏玩家替它付钱，却不用它。董事会签字接受了一笔战略性亏损，回报是十年后的事——把 CUDA-capable 的 GPU 灌进每个大学实验室、每台游戏 PC、每个研究集群里。

By the time deep learning happened, the install base was already there. Jensen calls NVIDIA "the house that GeForce built" — because GeForce is what got CUDA into the world.

等深度学习真的发生那天，装机量已经在那儿了。黄仁勋有句话：「NVIDIA 是 GeForce 盖起来的房子。」——因为是 GeForce 把 CUDA 送进了世界。

Pre-training scaled with data and parameters; people called it over when high-quality human data ran out. It wasn't — synthetic data took over. Post-training scaled with refinement. Then test-time scaling: making the model think longer at inference. Jensen's response to people who said inference would be a small commodity chip — thinking is harder than reading. Reasoning, search, planning, decomposition all consume compute.

预训练吃数据和参数；有人说高质量数据用完了 Scaling Law 就到头了。没有——合成数据接住了。后训练继续放大。然后是测试时（test-time）放大：让模型在推理那一刻多想一会儿。当年有人说推理简单、用个小芯片就行，黄仁勋当场反驳：思考比阅读难得多。推理、搜索、规划、问题分解——全是吃算力的活。

Then the fourth: agentic scaling. One agent spawns sub-agents the way Jensen hires employees. You can multiply intelligence by spinning off more workers. All four laws point at the same conclusion: intelligence scales by one thing, compute.

第四条是 agentic scaling：一个 Agent 像黄仁勋招员工那样，把一堆 sub-agent 派出去。智能可以被「乘出来」。四条 scaling law 指向同一个结论：智能只会沿一个东西扩展——算力。

AI architectures change every six months. Hardware architectures change every three years. NVIDIA can't pivot a Vera Rubin rack on a week's notice — so they have to bet on what AI looks like in 2026, 2027. NVLink-72 was built because mixture-of-experts arrived. Vera Rubin's new "Rock" rack was built because agents bang on tools and storage.

AI 模型架构每六个月就换一茬。硬件架构三年才换一代。Vera Rubin 整机柜不可能在一周内调头——所以 NVIDIA 必须押宝 2026、2027 的 AI 长什么样。NVLink-72 是为 mixture-of-experts 准备的。Vera Rubin 上新的「Rock」机柜，是为「Agent 要疯狂调用工具和存储」准备的。

The way Jensen picks the bet: do basic research internally; talk to every AI company; and reason from first principles. If a humanoid robot enters your kitchen, it isn't going to grow a microwave from its hand. It will use the microwave. Then read the manual. The schematic for OpenClaw, Jensen claims, was on a GTC slide two years before it shipped.

黄仁勋怎么挑赌局？内部做基础研究、和每一家 AI 公司聊、再用第一性原理推一遍。如果一个人形机器人走进你家厨房，它不会从手里长出微波炉——它会用你家那台。然后读说明书。OpenClaw 那张原理图，他说自己两年前的 GTC 上就讲过。

If a team says it takes 74 days to do something, Jensen strips the estimate to zero and asks the physical-limits question. If you built it from scratch tomorrow, what could it be? Often the honest answer is six days. The other 68 days are real — cost, supply, dependencies, compromise — but at least they're now visible.

团队说「这件事得 74 天」。黄仁勋会把估算清零，问一个物理极限的问题：如果你今天从零做起，最少能多快？答案常常是六天。剩下的 68 天是真实的——成本、供应、依赖、妥协——但至少这 68 天现在看得见了。

Continuous improvement gets you 74 → 72. First-principles thinking gets you 74 → 6, then explains the 6 → 74 walk-back as a deliberate set of trades. Jensen runs every system this way: latency vs throughput, cost vs capacity, power, time, headcount, all benchmarked against the physical limit, not against last year.

「持续改进」让你从 74 走到 72。「第一性原理」让你从 74 走到 6，然后清楚地讲出 6 → 74 这条回路上每一笔交易。黄仁勋每件事都这么跑：延迟 vs 吞吐、成本 vs 容量、功耗、时间、人头——全跟物理极限对标，而不是跟去年对标。

A Vera Rubin pod is seven chip types, five purpose-built rack types, 40 racks, 1.2 quadrillion transistors, ~20,000 NVIDIA dies, over 1,100 Rubin GPUs, 60 exaflops, 10 petabytes per second of scale bandwidth. The NVL-72 rack alone packs ~1.3 million components and 1,300 chips into a 19-inch frame.

一套 Vera Rubin pod 是这样的：7 种芯片、5 种定制机柜、40 个机架、1.2 千万亿（quadrillion）个晶体管、约 2 万颗英伟达裸片、超过 1100 颗 Rubin GPU、60 exaflops 算力、10 petabytes/秒的扩展带宽。光是 NVL-72 一个机架，就把约 130 万个零件、1300 颗芯片塞进了 19 英寸的框子里。

NVIDIA's target throughput: roughly 200 of these pods per week. The supercomputer is no longer assembled at the data center; the supply chain ships supercomputers, two to three tons at a time per rack.

NVIDIA 的目标产能：每周大概要交付 200 套这种 pod。超级计算机不再在数据中心装配——供应链直接发货「整柜超算」，每柜两三吨。

Jensen rarely announces a strategic pivot cold. Years before NVIDIA's deep-learning declaration, he was laying bricks — at GTC, in board meetings, with suppliers, with customers. By the time the announcement came, employees said what took you so long.

黄仁勋很少冷不丁宣布一个战略转弯。在 NVIDIA 公开 All-in 深度学习之前的好几年，他一直在「砌砖」——GTC 上、董事会上、和供应商谈、和客户谈。等真正宣布那天到来，员工的反应是：你怎么才说？

The same playbook landed Mellanox, then HBM-as-mainstream-memory, then OpenClaw. Leadership in this mode isn't authority — it's belief-shaping. You convince the supply chain, the customers, the engineers, the board, your own future self. Then the future you've described becomes the easiest path forward.

同一套打法用在收购 Mellanox、把 HBM 推成数据中心主流内存、押注 OpenClaw 上。这种领导力不是「权威」，是「塑信念」。你说服供应链、说服客户、说服工程师、说服董事会、最后说服未来的自己。然后你描绘的那个未来，会变成所有路径里阻力最小的那条。

Asked how he carries the weight of a company that nations now build their AI plans around, Jensen describes a single operator: decompose the circumstance, reason about it, hand pieces of it to people who can act on them, then forget. The list is the relief.

有人问他：现在国家把 AI 计划都建在 NVIDIA 身上，这种重量你怎么扛？黄仁勋只用了一套动作：把这件事拆开、过一遍逻辑、把能动的部分交给能动的人，然后忘掉。「列单」本身就是减压。

If you've identified the problem and either acted on it or told someone who can, you've discharged your obligation to it. The rest, in Jensen's phrase, is just crying about it.

如果你已经识别了问题，要么自己动手，要么交给能动手的人——你对这件事的责任就履行完了。剩下的，用他原话说，就是「在那哭哭啼啼」。

Jensen's resilience stack has four parts. Curiosity — go into new experiences with the mind of a child, asking how hard could it be. Forgetting — let setbacks pass; the next point matters more than the last point. Belief — when your assumptions about the future haven't materially changed, the predicted output hasn't either, so keep going. Constant reevaluation — but don't confuse constant reevaluation with constant doubt.

黄仁勋的「韧性堆栈」有四层。好奇心——带着孩子的脑子进入新事，先问一句「这能有多难？」。会忘——让挫折翻篇，下一个点比上一个点重要。信念——只要你对未来的假设没有实质性变化，预测的输出也没变，那就继续走。持续重估——但别把「持续重估」搞成「持续怀疑」。

The combination is what gets a person through three decades of a company nearly dying repeatedly. Jensen's quip: if he'd known how hard NVIDIA would be, he wouldn't have done it. But almost everything worth doing fits that description.

这套组合拳，是一个人怎么撑过三十年里这家公司一次次「差点死掉」。黄仁勋有句俏皮话：当年要是知道做 NVIDIA 有多难，他就不会做了。但「值得做的事」差不多每一件都是这种描述。

Most CEOs explain conclusions. Jensen explains the reasoning steps that produced them. The benefit is structural: people in the room can disagree with a step, redirect his thinking, and move the company without confronting the boss. He calls it collective path-searching.

大多数 CEO 给员工讲「结论」。黄仁勋给员工讲「推到结论那几步」。结构上有个好处：屋里的人可以反对其中某一步、把他的思路掰回来、推动公司往前走——而不必正面顶撞老板。他自己叫这个「集体寻路」。

The cost is tolerance for embarrassment — when a step is wrong, everyone sees it. Jensen treats that as a feature. After 33 years of doing the job in public, you stop being protective of being right. You become protective of being correctable.

代价是「丢人的容忍度」——某一步错了，所有人都看得见。黄仁勋把这当成一项资产。在大众面前干了 33 年，你就不再护着「我对了」这件事——你开始护着「我能被纠正」这件事。

Watching xAI build the Memphis Colossus supercomputer in four months, Jensen highlights three traits: questioning every requirement to its physical floor; presence at the point of action — showing up to the rack and asking the engineer what plugging in cables actually looks like; and personal urgency that infects the whole supplier chain.

看 xAI 在四个月里把 Memphis 那台 Colossus 超级计算机搭起来，黄仁勋拎出三点：把每一项「需求」逼问到物理底线；亲身在「干活的现场」——直接走到机柜前问工程师「插线这件事到底怎么操作」；以及一种个人级别的「紧迫感」，能感染整条供应链。

None of it is mystical. All of it is hard. Jensen's parallel: extreme co-design IS systems engineering at the company level — the same posture, scaled to a different problem.

没什么神秘的。每一条都很难。黄仁勋的对照：极致协同设计，本质上就是把同一种姿势——系统工程的姿势——放到公司这一层来用。

Three decades of business. Hundreds of billions of dollars. No contract. Jensen says the deepest misunderstanding of TSMC is that their moat is transistors. Their actual moat is orchestrating dynamic demand from hundreds of customers as the world shape-shifts — wafer starts, wafer stops, emergency reroutes — while still hitting yield, cost, and delivery promises. Plus a culture that is simultaneously bleeding-edge technology AND bleeding-edge customer service.

合作三十年，几千亿美元的生意，没有合同。黄仁勋说，外界对台积电最深的误解，是以为它的护城河是晶体管。其实护城河是它能在「世界一直在变形」的同时——加单、撤单、紧急换线、客户来回切换——把上百家客户的动态需求调度好，同时还把良率、成本、交期全打满。再加上一个罕见组合：技术做到极致前沿，客户服务也做到极致前沿。

Plus the intangible Jensen names last: trust. "I trust them to put my company on top of them. That's a very big deal."

最后还有一样，黄仁勋放在最后才说：信任。「我敢把整个公司架在他们身上。这件事很大。」

In 2013, Morris Chang offered Jensen the role of TSMC CEO. He was, in his words, "deeply honored" — Chang remains one of his closest friends and TSMC one of the most consequential companies in history. He declined.

2013 年，张忠谋亲口请黄仁勋去当台积电 CEO。黄仁勋的话：「深感荣幸」——张忠谋至今是他最亲近的朋友之一，台积电是历史上最重要的公司之一。他婉拒了。

"The work I'm doing here is really important. It's my responsibility to make this happen." Then a beat: "So now I can help both companies."

「这边的工作太重要了。把它做成，是我的责任。」停了一拍，他补一句：「不过现在我两家都能帮上。」

For sixty years, a computer was a retrieval machine. You pre-recorded a thing — a document, an image, a transaction — stored it, and the computer fetched it back when you asked. Warehouses, in Jensen's metaphor. Warehouses don't make money; they hold inventory.

过去六十年，计算机是一台「检索机」。你预录一个东西——一份文档、一张图、一笔交易——存进去，需要的时候让它取回来。黄仁勋的比喻：仓库。仓库不赚钱，仓库只是放货的地方。

AI computers don't retrieve. They generate, in real time, contextually, from a question you've never asked before. That makes them factories. Factories generate revenue. The unit of computing went from chip → server → cluster → AI factory — and the economic relationship between the machine and the company that owns it inverted.

AI 计算机不是检索。它是在你提出一个从没人问过的问题时，实时、上下文相关地生成回答。这是工厂。工厂直接产生收入。计算单元从芯片 → 服务器 → 集群 → AI 工厂——机器和拥有它的公司之间的经济关系，从此倒过来了。

What the factory produces is tokens. And tokens are starting to segment the way phones segmented in 2010 — free tokens, mid-tier tokens, premium tokens. Specialized intelligence will command specialized prices. Jensen names it directly: someone willing to pay $1,000 per million tokens isn't a hypothetical, it's an inevitability.

工厂产出的东西叫 Token。Token 正在像 2010 年的手机那样分档——免费 token、中端 token、高端 token。专门化的智能会卖出专门化的价格。黄仁勋直接讲：愿意付每百万 token 1000 美元的人，不是「假设」，是「迟早」。

Once you accept that intelligence is a commodity with a price curve and that price curve has a high end, the question isn't whether the world needs more compute. It's how many factories the world will eventually need — and what fraction of GDP routes through them.

一旦你接受「智能是一种商品、有价格曲线、曲线还有高端」——那问题就不再是「世界还需不需要更多算力」，而是「世界要建多少座工厂」、以及「GDP 里有多少最终会从这些工厂里走过去」。

Power is the headline blocker. But Jensen's structural fix isn't building generation faster — it's using what's already there. The grid is sized for worst-case demand. Most days it runs at ~60% of peak. The other 40% sits idle waiting for a few extreme-weather days a year.

功耗是公开的瓶颈。但黄仁勋的结构性解法不是更快地建发电厂——是先把已经在那儿的用起来。电网是按最坏情况设计的。绝大多数日子里它跑在峰值的 60% 左右。剩下 40% 闲在那等着每年那几个极端天气日。

His pitch: data centers should sign contracts that gracefully degrade — slow down, shift workloads, drop to backup — when the grid genuinely needs the headroom. Customers who insist on six-nines availability force CSPs to demand six-nines from utilities, and the result is a grid burdened by promises it shouldn't have made. Re-engineer the contract, and the grid releases capacity nobody had to build.

他的提法：数据中心应该签那种「优雅降级」的合同——电网真有压力时，数据中心愿意降速、迁移负载、回到备用电源。客户死磕「6 个 9 可用性」，CSP 就只能逼着电网也给「6 个 9」，电网为此被绑死在不该签的承诺上。把合同重写一遍，电网立刻释放出谁都不必新建的那部分容量。

What captured public attention wasn't a model release. It was OpenClaw — agents that read files, hit databases, use tools, and report back. Jensen calls it the iPhone of tokens: the fastest-growing application in history.

真正抓住公众注意力的，不是哪一个新模型——是 OpenClaw，那种能读文件、查数据库、用工具、再回报给你的 Agent。黄仁勋的说法：这是 token 的 iPhone——历史上增长最快的应用，一路飙直线。

The reason it landed harder than Claude Code or Codex isn't model capability. It's that consumers could reach it. Agents — not chatbots — are the consumer-shaped form factor of intelligence. NVIDIA's read: every model becomes an agent, agents need storage and tool-call infrastructure, and the rack architecture has to ship before the agents do.

它为什么比 Claude Code、Codex 落得更重？不是模型能力——是普通消费者够得着。Agent 不是聊天机器人，而是「智能」的消费级形态。NVIDIA 的判断：每个模型最终都会变成 Agent；Agent 需要存储和工具调用基础设施；机柜架构必须先于 Agent 上线。

In 2019-2020, computer vision became superhuman. The prediction: radiologists were finished. The actual outcome: the number of radiologists grew, and there's now a shortage. The mistake was conflating the task (reading scans) with the purpose (diagnosing disease, helping patients). Faster scan-reading meant more scans, more diagnoses, more patients seen, more demand.

2019–2020 年，计算机视觉变得「超人」。当时的预言：放射科医生要被淘汰了。实际结果：放射科医生不仅没减少，反而紧缺了。这次预言失误的根，是把「一个具体动作（读片）」和「一份职业的真正目的（诊断疾病、帮病人）」混为一谈。读片更快，意味着扫描更多、诊断更精、看的病人更多、需求更大。

Jensen's translation for the present: software engineers won't shrink, they'll grow. The number of coders — people who can specify what software should do — goes from ~30 million today to roughly a billion. Every carpenter becomes an architect. Every plumber becomes a service designer. The task automates. The purpose elevates.

黄仁勋把这一课直接搬到今天：软件工程师不会萎缩，会膨胀。「码农」——也就是「能跟计算机说清楚要做什么」的人——会从今天的约 3000 万人变成大约 10 亿人。每个木匠都会变成建筑师，每个水管工都会变成服务设计师。动作被自动化，目的被抬高。

Jensen's framing on AGI is functional, not mystical. He says he's surrounded by 60 people more intelligent than him in their respective specialties. They went to better schools. They're deeper in their fields. They're, in his word, superhuman. And yet a dishwasher — his self-description — sits in the middle and orchestrates them.

黄仁勋对 AGI 的定义很「功能化」，不带神秘色彩。他说自己被 60 个比他更聪明的人围着——每一位在自己专业里都比他强。他们读的学校更好，钻得更深，用他自己的话讲，是「超人级」的。然而一个「洗碗工」——他对自己的形容——坐在中间，把他们 60 个人协调起来。

The point isn't humility. It's that intelligence is a commodity; orchestration, judgment, character, tolerance for embarrassment — those are the scarce resources. The word society has elevated is intelligence. The word it should elevate is humanity.

重点不是谦虚，是：智能是商品；调度、判断、品格、丢得起脸的能力——才是稀缺资源。社会把「智能」这个词拔到了最高位。黄仁勋说，应该被拔上去的，是「人性」。

The conversation arcs from the rack to the radiologist, from the speed of light to the soul. Jensen's bet — and it is a bet, even if he calls it inevitable — is that compute will keep scaling, that intelligence will keep getting cheaper, and that the people who win the next decade will be the ones who treat humanity, not intelligence, as the rare thing. NVIDIA is the engine. The rest is the question of what we choose to use it for.

这场对话从机柜聊到放射科、从光速聊到灵魂。黄仁勋的赌注——是的它就是一个赌注，哪怕他自己叫它「必然」——是：算力会一直 scale，智能会一路便宜，下一个十年的赢家，是那些把「人性」而不是「智能」当稀缺资源对待的人。NVIDIA 是那台引擎。剩下的问题是：我们准备拿它做什么。