11-19 更新 实时
首页/ 深度/ 正文

讣告

深度 来源:Stratechery 2026-11-16 19:02 0 阅读 机器翻译
讣告

收听本文:

亲爱的各位,我们今天聚集于此,是为了度过这所谓的人生。——Prince

2026 年 11 月 12 日(周四),黄仁勋辞去 Nvidia CEO 一职。自 1993 年 4 月 5 日创立该公司以来,他一直担任 CEO。Nvidia 最初是一家显卡公司;黄仁勋的离开——名义上是因为与一位可能是、也可能不是中国间谍的买家,以及至少两名公司员工的绯闻——归根结底是因为 Nvidia 被当权者视为国防承包商,而黄仁勋始终没有做出调整。

就此而言,黄仁勋是自己成功的最终受害者。Nvidia 最早的显卡只是把三角形转换成像素。2001 年的 GeForce 3 引入了可编程像素着色器,这意味着显卡的功能不再是被预先定义的。然而,那记杀手锏——也几乎要了公司命的一步——是 2006 年推出的 CUDA,它是一套用于为这些着色器编程的 API。此后 16 年的大部分时间里,黄仁勋一直在传布 CUDA 的福音,坚信它是加速计算未来的平台;然后,2022 年,ChatGPT 来了。

ChatGPT 让 GPU 最有价值的用途变得清晰:大语言模型,而 Nvidia 造出了最好的 GPU。不仅如此,黄仁勋还意识到,加速计算的未来不只是芯片,而是系统;这推动了 2019 年对 Mellanox 的收购,后者提供的技术让 Nvidia 得以从芯片向外扩展,从而走上成为那种让大语言模型成为可能的大规模训练的最佳选择之路。为此,Nvidia 收取了非常高的价格,毛利率超过 70%,因为他们有这个底气:他们拥有平台、技术,以及在过去几年里当其他人试图自研芯片时所拥有的供应链。

黄仁勋对亚洲供应链的影响尤其怎么强调都不为过。Byrne Hobart 在其著作《Boom: Bubbles and the End of Stagnation》中指出,泡沫的一个好处是它发挥了协调功能:对错失机会的恐惧,激励着多个彼此存在交叉依赖的主体同时投入建设;从很多方面看,黄仁勋本人就是一个"一人协调机制"。他每年发表多场主题演讲,他解释说这是他广泛与生态沟通的方式;他还以在亚洲各地一路吃吃喝喝而闻名,说服不情愿的供应商相信 AI 热潮是真实的,并为 NVIDIA 锁定他们产能的绝大部分,以及其对年度发布周期令人惊叹的坚持。

黄仁勋在相对默默无闻中耕耘了数十年,主要只有游戏玩家认识他,科技行业也对他知之甚少,如今他却成了摇滚明星,所到之处总被人群簇拥,永远穿着那件黑色皮夹克。而最终对他造成不利的是,他也成了一个政治人物。Nvidia 长期以来在华盛顿和北京都没有任何存在感,2022 年拜登政府实施芯片管制时被打了个措手不及。面对最初这些以互连速度为依据限制出口的管制,Nvidia 的即时反应是在中国推出 A800——即把 A100 的互连削减三分之一后的版本。这为我所认为的黄仁勋垮台真正原因的那场冲突定下了基调:黄仁勋强烈反对各种版本的芯片禁令,而 Nvidia 只肯做最低限度的事来遵守法律的字面规定,从不遵守其精神。

这种做法很可能也延伸到 NVIDIA 对新加坡实体的监管(或者说监管缺失)上,这些实体被指控是中国获取芯片的幌子。其中一些是合法的:中国企业被允许使用中国境外搭载领先 NVIDIA 芯片的云计算设施,尽管芯片管制倡导者正寻求堵上这一漏洞;有些则不合法,因为这些芯片最终被运往中国。Megaspeed 的高管 Alice Huang 是被指控与 Huang 有婚外情的第一人;该公司官方上从事的是前一类业务,但据称实际上从事的是后一类业务。有传言称,Megaspeed 的最终客户不是阿里巴巴,或者不只是阿里巴巴,还包括中国军方。需要明确的是,这些都只是传言;至少在公开层面,光是那些绯闻就足以导致 Huang 辞职。

NVIDIA 的持久性——至少其利润率的持久性——本就是一个悬而未决的问题。先从 CUDA 说起:当市场卖的是为 GPU 编写的软件时,一个平台——不仅包括 API,还包括面向各种可能的加速计算市场的大量库——几乎是不可逾越的护城河。这就是像 GPT 这样的大语言模型是用 CUDA 编写的原因。

不过,LLM 既给 NVIDIA 带来了惊人的成功,也埋下了它最大的风险。第一,绝大多数开发者的精力都向上层迁移,把 LLM 的 API 当作平台来使用;对大多数公司而言,CUDA 并不重要。第二,Anthropic 从 Amazon 和 Google 获得了融资——这两家原本只是 Microsoft-OpenAI 合作关系的陪跑者——因此不得不学习如何使用 Amazon 的 Trainium 芯片和 Google 的 TPU;这很困难,但最终意味着这家领先的 LLM 公司完全没有任何 NVIDIA 依赖(Anthropic 仍运行在 NVIDIA 芯片上,因为他们学会了在任何芯片上运行)。第三点最具毁灭性:LLM 在编写软件方面如此高效,以至于任何基于软件的护城河都必然会干涸。

NVIDIA 的第二道护城河——其网络与系统级方案——更为持久,尽管尤其受到 Google(进而也包括 Anthropic)的挑战。NVIDIA 的系统级方案,对于希望自建硬件的企业和政府,以及希望建设独立于超大规模云厂商的数据中心的新兴云厂商(neocloud)而言,也是最省事的答案。后一市场向超大规模云厂商、前沿实验室以及个别公司出租芯片,但随着显而易见的电力来源和政治许可消失,这一市场正面临挑战。他们还需要资金才买得起 NVIDIA 的系统,而这些资金越来越多地来自 NVIDIA 本身。

Nvidia 的第三道护城河是它对供应链的掌控,而这道护城河如今突然显得脆弱得多。黄仁勋曾夸耀,在与 TSMC 合作三十年间,他从未签过一份合同;这段关系建立在握手与互信之上,且这种互信是专门针对黄仁勋个人的。如上所述,这是一种非常亚洲式的经商方式,而黄仁勋正是此道高手。当握手不再存在时,这种信任与承诺意愿能否延续,就很成疑问了。

然而,Nvidia 最重要的护城河在于它在多大程度上不像一家正常公司。黄仁勋把 Nvidia 组织成围绕一个人——他自己——的外骨骼。Nvidia 的管理层极少——黄仁勋有 60 名直接下属是出了名的——这与没有继任计划出于同样的原因:Nvidia 之所以能如此迅速、如此激进地行动,是因为这家公司的行为方式仍像一家初创公司。确实,许多成熟企业——包括那些仍由创始人领导的企业——都声称自己仍像初创公司,但真正按这种方式组织起来的只有 Nvidia 一家。这让它那些成功的押注得到了回报;也造就了一个对黄仁勋毫无制衡的环境。而如今他已不在,这套外骨骼崩塌的可能性远比你预想的要大。

AI 建设的第一阶段由自由现金流提供资金,尤其是来自超大规模云厂商的自由现金流。随后,在去年下半年,几乎所有人都开始转向债务。现在——从刚过去的这个夏天 Google 开始——主要融资方式变成了股权。从长远看,这应该会绕回原点,未来的建设将再次由自由现金流提供资金,而这些现金流将来自眼下正在建设的基础设施。

NVIDIA 面临的问题是,最有能力驾驭这一演进进程的正是超大规模云厂商,它们拥有挑战 NVIDIA 在芯片市场主导地位的资源。为此,NVIDIA 一直非常急于确保自己不仅向超大规模云厂商销售,也向其他潜在竞争对手销售,比如新型云厂商(neoclouds);很可能在某个时点,人们会就 NVIDIA 究竟是如何为追求这一目标而分配算力的,提出严肃质疑。

然而,在融资方面,新型云厂商处于明显的劣势。首先,它们的债务成本高得多,因为它们唯一的业务就是出租 GPU(而不像那些经营着全球最有价值广告业务、或全球最大电商平台、或主导企业软件的公司)。其次,鉴于这些信贷所附带的重大风险,它们的股权正变得越来越不值钱。如果一家新型云厂商付不出钱,最先被清零的就是股权持有人。

原文较长,译文为节选;译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Obituaries)

Listen to this post:

Dearly beloved, we are gathered here today to get through this thing called life.
― Prince

Jensen Huang, CEO of Nvidia, 1993–2026

Jensen Huang resigned as CEO of Nvidia on Thursday, November 12, 2026. He had been CEO of the company ever since he founded it on April 5, 1993. Nvidia started out as a graphics card company; Huang’s exit — nominally because of affairs with a buyer who may or may not have been a Chinese spy, as well as at least two company employees — was ultimately because Nvidia was, by the powers that be, considered a defense contractor, and Huang never adjusted.

In this, Huang was the ultimate victim of his own success. Nvidia’s first graphics cards simply translated triangles into pixels. The 2001 GeForce 3 introduced a programmable pixel shader, which meant a video card’s functionality was no longer pre-defined. The killer move, however — which almost killed the company — was the 2006 introduction of CUDA, which was an API for programming those shaders. For much of the next 16 years Huang sold the gospel of CUDA, convinced it was the platform for a future of accelerated computing; then, in 2022, came ChatGPT.

ChatGPT made clear what GPUs were most useful for: large language models, and Nvidia made the best GPUs. More than that, Huang had realized that the future of accelerated computing was not just chips, but rather systems; that drove the purchase of Mellanox in 2019, which provided the technology for Nvidia to scale out from chips, setting it down the road to being the best option for the sort of large scale training that made large language models possible. For this Nvidia charged very high prices, with margins exceeding 70%, because they could: they had the platform, the technology, and, over the past few years as others sought to make their own chips, the supply chain.

Huang’s impact on the Asian supply chain in particular cannot be overstated. Byrne Hobart, in his book Boom: Bubbles and the End of Stagnation, notes that one of the benefits of bubbles is that they provide a coordination function, as the fear of missing out incentivizes multiple entities with interlocking dependencies to all commit to building at the same time; Huang was, in many respects, a one-person coordination function. Huang gave multiple keynotes a year, which he explained as how he communicated with the ecosystem broadly, and was famous for eating and drinking his way across Asia in particular, convincing reluctant suppliers that the AI boom was real, and locking up the lion’s share of their output for Nvidia and its stunning commitment to a yearly release cycle.

Huang, who had toiled for decades in relative obscurity, known primarily to gamers, and only barely to the tech industry, became a rock star, swarmed wherever he went, always in his black leather jacket. He also, to his ultimate detriment, became a political figure. Nvidia had long had no presence in either Washington D.C. or Beijing, and was blindsided by the Biden administration’s imposition of chip controls in 2022. Nvidia’s immediate reaction to those initial controls, which restricted exports based on interconnect speed, was to immediately release the A800 in China, which was an A100 with a third of its interconnects disabled. That set the tone for the conflict that I believe was the true cause of Huang’s downfall: Huang strenuously disagreed with the various iterations of the chip ban, and Nvidia only bothered to do the bare minimum to abide by the letter of the law, never the spirit.

That approach, in all likelihood, extended to Nvidia’s oversight (or lack thereof) of Singaporean entities that were alleged to be fronts for chips for China. Some of these were legal: Chinese companies are allowed to access cloud computing facilities outside of China that house leading edge Nvidia chips, although chip control advocates are seeking to close that loophole; some were not legal, in that the chips were ultimately ferried to China. Megaspeed, whose executive, Alice Huang, was the first person Huang was alleged to have an affair with, was officially in the former business, and alleged to actually be in the latter. There are whispers that Megaspeed’s ultimate customer was not, or not just, Alibaba, but also the Chinese military. These are whispers, to be clear; it was the affairs that were enough, at least publicly, to lead to Huang’s resignation.

Nvidia, Most Valuable Company in the World, 2024–2026

Nvidia’s durability — at least that of their margins — was already an open question. Start with CUDA: when the market was software written for GPUs, a platform — including not just an API but a huge set of libraries targeting every conceivable market for accelerated computing — was a nearly impenetrable moat. That is why LLMs like GPT were written using CUDA.

LLMs, though, both brought Nvidia its incredible success, and also seeded its greatest risks. The first is that the vast majority of developer energy moved up the stack, using the LLM’s API as a platform; for most companies, CUDA didn’t matter. The second is that Anthropic, which found funding from Amazon and Google, originally also-rans to the Microsoft-OpenAI partnership, had to learn how to use Trainium chips from Amazon and TPUs from Google; this was difficult, but ultimately meant that the leading LLM did not have any Nvidia dependency at all (Anthropic runs on Nvidia chips, because they learned how to run on anything). The third is the most devastating: LLMs are so effective at writing software that any software-based moat will inevitably dry up.

Nvidia’s second moat, its networking and systems approach, is more durable, although challenged by Google (and by extension, Anthropic) in particular. Nvidia’s systems approach is also the easy answer for enterprises and governments seeking to own their own hardware, and neoclouds seeking to build data centers that are independent of the hyperscalers. The latter market rents chips to both hyperscalers and frontier labs, as well as individual companies, but is challenged as obvious sources of power and political permission disappear. They also need money to afford Nvidia’s systems, which has increasingly come from Nvidia itself.

Nvidia’s third moat is its mastery of the supply chain, and this is the moat that suddenly looks much more tenuous. Huang has bragged about the fact that he has never signed a contract with TSMC over their three decades of partnership; it’s a relationship built on a handshake and mutual trust with Huang specifically. This is, per the point above, a very Asian way of doing business, and Huang was the master of that. It’s questionable whether or not that trust and willingness to commit will persist when the handshake is no longer there.

Nvidia’s most important moat, however, was the degree to which it was not a normal company. Huang organized Nvidia as an exoskeleton around one man: himself. There is very little management at Nvidia — Huang famously has 60 direct reports — for the same reason that there was no succession plan: Nvidia could move so quickly and so aggressively because the company still behaved as if it were a startup. Indeed, lots of established companies — including those still led by founders — claim to still be a startup, but only one, Nvidia, really was organized that way. That paid off with all of those successful bets; it also created an environment where there were no checks on Huang. And, now that he is gone, the exoskeleton is much more poised to collapse than you might expect.

The AI Bubble, 2025–2027?

The first part of the AI buildout was funded by free cash flow, particularly from the hyperscalers. Then, during the second half of last year, nearly everyone started turning to debt. Now — starting with Google this past summer — the primary funding mechanism is equity. In the long run, this should come full circle, and future build-outs will once again be funded by free cash flow, generated from the infrastructure that is being built now.

The problem for Nvidia is that the companies most able to navigate this progression are the hyperscalers, which have the resources to challenge Nvidia’s hold on the chip market. To that end, Nvidia has been very eager to ensure that it is selling not only to the hyperscalers, but to other would-be competitors, like the neoclouds; there will likely, at some point, be serious questions raised about exactly how Nvidia went about allocating compute in pursuit of this goal.

The neoclouds, however, are at a meaningful disadvantage when it comes to funding. First, their debt is significantly more expensive, given that their only business is renting GPUs (as opposed to, say, running the most valuable advertising businesses in the world, or the world’s largest e-commerce platform, or dominating enterprise software). Second, their equity is increasingly worthless, given the significant risks attached to that credit. If a neocloud can’t pay, it’s the equity holders that get wiped out first.

This led to what I termed Nvidia’s Risky Business: Nvidia increasingly used its cash flow and credit rating to backstop and fund the companies who were buying its chips. Bears called this circular financing; bulls called it a necessary bridge to a future where AI revenue paid down these companies’ debt and drove massive upside. What is undoubtedly true is that this was the sort of risky business that is uniquely possible for a multi-trillion dollar company that is completely controlled by one man, who has made an entire career of pushing the limits and making risky bets — and seeing them pay off.

It’s also a series of bets that are unlikely to continue. The neoclouds should be very nervous about losing their patron, along with their creditors and equity holders. The other problem is that the revenue necessary to make all of this build-out work is something on the order of a trillion dollars industry-wide, and it needs to materialize sooner rather than later; if it falls short, it’s the hyperscalers who will prioritize their own compute first, and they will likely be seen as a safe haven for everyone committing to spending. In that world, Nvidia’s neocloud hedge — absent Huang’s willingness to shore the ecosystem up — is liable to turn into a millstone, with their assets flooding the market even as the supply chain ramp-up brings ever more supply to market.

AI as Normal Technology and the U.S. Semiconductor Industry, 1947~2047

Perhaps because he had been preaching the gospel of accelerated computing for so long, Huang was the leading figure pushing back against the doomer narrative surrounding AI. Yes, AI was a massive business opportunity, but it was still normal technology; the safety issues presented were engineering challenges to be solved. What required consideration was not just the next five years, but the next fifty. It’s a view that I am, in full disclosure, generally aligned with.

Many others are not. On one hand there are the aforementioned doomers, certain that fast takeoff is imminent — or has already started — as AI increasingly builds better AI in a recursive loop. The end state is an AI that is massively more capable than humans, and uncontrollable by them; the only hope is to ensure that AI is aligned with humanity, by which adherents to this viewpoint mean aligned with them and their view of the world. This contingent is, broadly speaking, represented by Anthropic. And, in this view, actively working to enable and accelerate uncontrollable open weights models is a threat to humanity.

On the other hand there are the China hawks, who view AI as a national security issue, and the key to U.S. military dominance in a world where China owns the means of production of most militarily relevant hardware. In this world giving China access to chips is effectively treason against the United States.

For both of these groups, Huang was public enemy number one, because of his determination to do what was best for Nvidia’s business, which was to sell chips to China. This made sense for Nvidia for multiple reasons:

  • The first and most obvious reason is that it is good for a company to be able to sell to everyone, including the second largest market in the world (China has traditionally made up about a quarter of Nvidia’s sales, long before AI was even a thing).
  • The second is that China is a large enough market, with sufficiently motivated and capable developers, that it can actually create a meaningful CUDA alternative; furthermore, such an alternative would, in the long run, not be constrained to China, but rather compete with Nvidia all over the world.
  • The third is that China’s AI ecosystem, unconstrained as it is by such niceties as terms of service, distills frontier lab models and releases them as open weights models; those models, meanwhile, are tuned to run on Nvidia chips. These models give Nvidia leverage versus the frontier labs all over the world, including in the West.
  • The fourth is that China is uniquely positioned to create a fundamentally different approach to systems design using optical interconnects to scale out training clusters. This is much more inefficient in terms of power usage than Nvidia’s more elegant designs, but China is also much more capable of manufacturing excess amounts of power to make it work.
  • The fifth is that, in the very long run, China can create an entire alternative supply chain for semiconductors broadly, rendering Nvidia’s mastery of the Western-aligned supply chain much less important.

This last point is worth dwelling on. The chief objection that China hawks in particular make to the idea that it is a good idea to sell China chips in order to prevent them from surpassing the U.S. is that China is already heavily invested in building their own ecosystem. This is certainly true, but I think it underplays the complexity of chips and what is necessary to actually compete with the West.

Specifically, the Chinese government can only control the supply side, by spending money and making top-down imperatives; building an ecosystem, however, whether that be for AI specifically, or semiconductors broadly, requires demand as well. Progress in models and systems and chips depends on a feedback loop between the user and the developer, or the customer and the manufacturer. The demand side, however, is much more short-term oriented than the Chinese government might prefer. The model companies’ concern is the other model companies: that means if they are distilling, we must distill too. If they are using Nvidia chips, then we must too. This matters up and down the stack, just as it has always mattered in technology, which really is different in this regard from almost every other industry.

A useful analogy is to consider the Internet. Why is it that country after country can profess a desire to have an indigenous alternative to the American tech giants, yet miserably fail in every attempt to generate it? It’s because the tech giants are only a click away! In this regard, China’s Great Firewall was one of the most brilliant geostrategic moves ever: by virtue of cutting off access to the American tech companies, China created the demand that pulled China’s own tech giants into existence. You need demand, not just desire.

This is why China itself is currently banning the import of Nvidia’s two-generations-behind H20 chips, after the Trump administration foolishly banned their import, only to reverse themselves while bragging about the reason to do so — i.e. keep China addicted to Nvidia. China is replaying the Great Firewall playbook: force its internal demand to use its much worse alternatives, because that is the only way those alternatives become better. And, in the long run, those alternatives, and the entire ecosystem that supports them, will flood the world.

Huang was, to be perfectly clear, talking his book when it came to his insistence on selling chips to China, or, in the case of the current controversy, allegedly turning a blind eye towards nominally legal buyers who were intent on violating the chip ban; that doesn’t mean, however, that he was wrong about the long-term effects, at least if AI is a normal technology (and I haven’t gotten to the risks this imposes on the Taiwan Strait). What is clear now, however, is that the doomers and the hawks have won: AI, at least in terms of policy, will be treated as anything but normal, and in the fullness of time, it is the U.S. chip ecosystem that will pay the ultimate price.

Tech Company Independence, 1968–2026

The birth date for the last obituary was the invention of the transistor; the birth date for this one is the founding of Intel. Robert Noyce and co-founders had spent the last decade at Fairchild Semiconductor, but were constrained by their parent company’s fundamental conservatism and chafed at their lack of upside. Arthur Rock effectively created the venture capital industry in Silicon Valley through his financing of Intel; Noyce created the concept of the commercial-market-led technology company that made it up in volume. At that time the largest single purchaser of chips was the U.S. government; Noyce realized that the fastest way to progress was to make chips for everyone, including the government, and on his terms, not theirs.

From that point on, tech as an industry operated with the assumption that the world was its customer, and that their own executives defined the route to progress. Government was free to buy their products, but not to define them, and if they wanted the best products, then they best not restrict who they could sell them to either. That’s how important volume was, and, over time, it became clear how powerful and essential ecosystems were.

Other industries operate differently; I thought this article in the Financial Times was striking, and not just because of the headline: Bessent says Nvidia needs ‘shake-up’ with external CEO:

Bessent said that it is important “to examine whether Nvidia had a culture of compliance around the US export restrictions” and whether its top executives knew they were being skirted. “Was it a low-level fail? Was it in the middle, or was it up in the C-suite? And who knew? These shell companies were obvious, they were being investigated, and did Nvidia use them as a workaround to sell chips to the Chinese?” he said. “Were they pressing these seemingly illicit sales to goose their revenues and their earnings?” The Treasury secretary added that Nvidia needs to provide answers within a matter of days: “I think we need answers in the next 10 days from the board and from the outside counsel”. He added that “we will see if Nvidia intentionally ignored the warning signs of these shell companies, many of them in Singapore”.

Bessent has pointed the spotlight on Howard Lutnick’s commerce department for failing to ban exports of Nvidia’s chips to Megaspeed, the Singaporean company linked to the alleged Chinese spy who had an affair with Huang. US authorities have been investigating whether Megaspeed, which builds AI data centres in Asia, has been involved in smuggling high-end Nvidia chips, which are powering the AI boom, into China. Bessent has blamed the commerce department more than once this week for failing to ban exports of Nvidia’s chips to Megaspeed. Commerce delayed efforts to add Megaspeed to Washington’s blacklist of companies the US considers a risk to national security. “Many other departments were pushing for this, and commerce had the pen,” Bessent repeated on Sunday. “We have all the rules we need. We just need them enforced.”

Lutnick is under mounting pressure from within the administration and from Congress, which has demanded that he explain how he allowed chips to be exported to a company linked to an alleged Chinese spy. US lawmakers are also accelerating efforts to crack down on advanced AI chips reaching China. If Nvidia “were a bank and they hadn’t complied like this, the [Treasury’s Office of the Comptroller of the Currency] would be in there right now, and a lot of the executives would be removed”, Bessent said.

Scott Bessent is the Treasury Secretary, which is to say that he operates an apparatus that is arguably just as, if not more, powerful than the U.S. military. Treasury controls the flow of money around the world, and its soldiers are the banks; if Treasury has sanctioned an entity, then it is up to banks to ensure those sanctions are enforced. They need to know their customers, and they need to analyze every dollar flowing through their systems; if they do not, they are breaking the law. Ignorance is not a defense.

Ignorance, on the other hand, has been tech’s defense basically forever. Yes, there are bad people in the world, which means that if the whole world is using your products, they will be used to do bad things. That was considered an acceptable trade-off, however, because most people are good and the benefits that come from scale — and an avoidance of the friction that comes from trying to track the activities of billions of people — were worth the risk.

This was certainly what Huang believed, and, to his detriment, he aggressively belittled anyone who dared disagree (the extent to which Huang is despised in Washington D.C., as evidenced by the giddiness and bloodlust online last week, is extraordinary). In Huang’s view, as long as the Commerce Department said that it was ok to sell to Megaspeed, then it was ok to sell to Megaspeed — and to sleep with one of its executives.

Bessent views the matter entirely differently: he thinks that Nvidia should operate like a bank. It’s not just their job to abide by the letter of the law, but the spirit of it. It’s not the government’s job to enforce the chip ban, but the company’s. An attestation from a Singaporean customer that chips aren’t going to China isn’t worth anything; actually putting boots on the ground and ensuring compliance is the only way Nvidia cannot be in violation of the U.S. government’s dictates.

Again, for tech, this is extraordinary, even if it is normal in banking. Extraordinary enough, in fact, that Huang is being labeled by many — not just in Washington, but also in the finance world — as a traitor. Huang, in my estimation, is just a businessman who cares about nothing else other than Nvidia; that worked out great when he made GPUs for gaming that no one in government cared about or understood. That’s just a normal technology; when AI was deemed to be anything but normal, however, business as usual wasn’t just over: it was a crime, and it cost Huang his job — and maybe more.

It’s worth pointing out that while this episode was shocking, I should not, in retrospect, have been surprised. Earlier this year, when Anthropic faced off against the U.S. government, I wrote in Anthropic and Alignment:

At the same time, what is the standard by which it should be decided what is allowed and not allowed if not laws, which are passed by an elected Congress? Anthropic’s position is that Amodei — who I am using as a stand-in for Anthropic’s management and its board — ought to decide what its models are used for, despite the fact that Amodei is not elected and not accountable to the public.

And, on the second point, who decides when and in what way American military capabilities are used? That is the responsibility of the Department of War, which ultimately answers to the President, who also is elected. Once again, however, Anthropic’s position is that an unaccountable Amodei can unilaterally restrict what its models are used for.

My point in that Article was that if AI is as exceptional as claimed, then the government will demand a say in how it is used, whether or not private company executives agree; moreover, that was a good thing, or at least a least-bad thing. The same analysis applies to Nvidia and Huang: Huang may have fundamentally disagreed with the chip ban, but at the end of the day, it was his duty to accept that it was the law of the land. To the extent he did not do so is ultimately his responsibility.

American Exceptionalism, 1776–2026

Perhaps the most tragic aspect of the Huang affair is the extent to which he was the personification of The American Dream. Huang was sent to America as a nine year old who didn’t speak a lick of English, and somehow ended up at a boarding school in rural Kentucky for misbehaving youth, where he would trade help on homework for protection from bullies. From there he went to a state school in Oregon, became an engineer, before famously ending up at a Denny’s diner where he and his co-founders decided to start Nvidia. Then, over the following 33 years, he took risk after risk, created innovation after innovation, until he became a hero not just in the United States, but across the world.

And then, you might say, he flew too close to the sun. Certainly he did so in his private life: it’s not just the affairs, which may be par for the course for rich and powerful men; rather, it’s notable just how reckless the affairs were. The ones that were revealed last week were with a customer and potential Chinese spy, and two of his own employees; both are fireable offenses. Meanwhile, the way in which Huang treated his opponents — including actively seeking to get people fired who he felt stood in the way of getting his way — bore the marks of a man who has constructed a life, and a company, where no one can or will stand up to him. That is a recipe for disaster, and here we are.

At the same time, it really does bum me out that what ultimately doomed Huang was a set of rules that really are a big departure from the status quo — and, in my opinion, misguided. Huang’s actions — or lack of actions, in terms of enforcing compliance with the chip ban — are treated as if they are an obvious crime, when in fact this is a dramatic shift in expectations from a short time ago (although there is certainly a callback to previous controversies around selling advanced computing to adversaries). Huang should have adjusted and accepted the new reality, make no mistake; I’m sympathetic to the fact he couldn’t, and recognize the failure in not having structures in place that could have warned him about the dangerous road he was embarking on, even as I realize Huang’s willingness to navigate dangerous roads is precisely why he was an American success story.

Beyond Huang, however, there is the question of how we compete with China, and I am disappointed, although not surprised, that we are pursuing the route of bans and forced compliance, instead of innovation and openness. Make no mistake, I am no China dove, and I have been very early to the argument that we erred greatly in willingly exporting entire industries to the communist state. That, however, makes it even more imperative that we sprint in the opposite direction of China, towards more freedom in a bet on inventing new industries, not more control and a focus on gatekeeping what we have. I get the logic, particularly in the case of AI, which is extraordinary; I lament the implications of what it says about our belief in ourselves and our ability to stay ahead.


Presented by

Stripe

As AI models improve, you might expect the market will consolidate around a handful of providers. But user data tells a different story.

Among OpenRouter customers spending at least $100 a month, the median number of AI model families in use has doubled over the past year, from three to six.

Even as AI models become more capable, businesses are using more models, not fewer. This makes sense when you consider that intelligence isn’t one-dimensional. A model that’s exceptional at writing code may not be the fastest at answering customer questions, or the most economical at processing millions of documents.

The question is shifting from “which AI model should we build on?” to “which model should handle each task?” The answer, increasingly, is ‘it depends.’ Which is great for flexibility, less great for simplicity.

OpenRouter, which recently joined Stripe, allows businesses to access and govern hundreds of models with the flexibility to optimize for cost, performance, and reliability as the market evolves.

Learn more at openrouter.ai.


正文保留英文原文,标题与摘要由机器翻译。

本文由采集器自动同步

原文发布于 Stratechery(Obituaries),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗