AI is slowing down

人工智能正在放缓

If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 5,000 to 18,000 words, including vast, detailed analyses of NVIDIA, Anthropic and OpenAI’s finances, and the AI bubble writ large (updated to version 3.0 last week). My Hater’s Guides To the SaaSpocalypse, Private Credit and Private Equity are essential to understanding our current financial system, and my guide to how OpenAI Kills Oracle pairs nicely with my Hater’s Guide To Oracle.

如果你喜欢这篇文章，欢迎订阅我的付费通讯。年费为 70 美元，或每月 7 美元。作为回报，你每周将收到一份 5,000 到 18,000 字不等的通讯，其中包括对 NVIDIA、Anthropic 和 OpenAI 财务状况的详尽分析，以及对整个 AI 泡沫的宏观解读（上周已更新至 3.0 版本）。我的《SaaSpocalypse（SaaS 启示录）、私募信贷和私募股权仇恨者指南》对于理解当前的金融体系至关重要，而我撰写的《OpenAI 如何扼杀甲骨文》与我的《甲骨文仇恨者指南》堪称绝配。

Over a three week period in May, I published an exhaustive three-part guide to how the AI bubble might collapse, the events that might trigger it, and the consequences. For something lighter, check out last week’s premium, where I re-introduce you to the antagonists of the AI bubble (and their fatal weaknesses) in caustic, slightly sweary terms. Subscribing to premium is both great value and makes it possible to write these large, deeply-researched free pieces every week.

在 5 月份的三个星期里，我发布了一份详尽的三部分指南，探讨了 AI 泡沫可能如何破裂、可能引发破裂的事件及其后果。如果想看点轻松的内容，可以看看上周的付费文章，我用尖刻且略带粗口的语言重新介绍了 AI 泡沫中的反派角色（以及他们的致命弱点）。订阅付费内容不仅物超所值，也让我能够每周撰写这些篇幅宏大、深度研究的免费文章。

Last week I went on Bloomberg and discussed the state of the AI bubble with a clarity that rattled even the sweatiest boosters, mostly because I spoke with clarity about an investment frenzy whipped up through hype, deceit and mythology. Some were equal parts frustrated and angry that I don’t have money in the market, or, as they’d put it, “skin in the game.”

上周我参加了彭博社的节目，以一种让最狂热的鼓吹者都感到不安的清晰度讨论了 AI 泡沫的现状，主要是因为我直言不讳地揭露了这场由炒作、欺骗和神话煽动起来的投资狂潮。有些人既沮丧又愤怒，因为我没有在市场上投入资金，或者用他们的话说，没有“利益攸关”。

I get it! When your entire worldview is dictated by what a series of venture capitalists and psuedo-journalists on Twitter want you to believe, it must be difficult to imagine someone having “morals” or “beliefs” or that one might hold a position that wasn’t entirely based on greed or tribalism. It must be confusing — upsetting, even! — to hear that somebody is willing to accurately and vociferously tear into a tech industry largely controlled by people with no regard for their users or workers, who are willing to bathe their products in mediocrity all because it’s the thing that everybody else is doing.

我理解！当你的整个世界观都被推特上的一系列风险投资家和伪记者所灌输的内容所左右时，你一定很难想象有人会有“道德”或“信仰”，或者有人持有的立场并非完全基于贪婪或部落主义。听到有人愿意准确且大声地抨击一个由完全不顾用户或员工利益的人所控制的科技行业，这一定让你感到困惑，甚至不安！这些人仅仅因为随大流，就愿意让他们的产品平庸至极。

This is a hysterical era perpetuated by liars, cowards, imbeciles, craven boosters and the easily-fooled. Those excited about generative AI are either the victim or the perpetrator of a con centered around a technology to ingratiate at the highest cost possible.

这是一个由骗子、懦夫、白痴、卑劣的鼓吹者和易受骗者所延续的歇斯底里的时代。那些对生成式 AI 感到兴奋的人，要么是这场骗局的受害者，要么是其参与者，这场骗局围绕着一种以最高成本去讨好市场的技术展开。

AI Cannot Afford To Slow Down — It Needs $3 Trillion Or More In Revenue By End Of 2030 To Sustain Its Existence. I also think that everybody is a little flippant about what has to happen for me to be wrong.

AI 承受不起放缓的代价——到 2030 年底，它需要 3 万亿美元或更多的收入才能维持其生存。我还认为，对于“我必须在什么情况下才会出错”这个问题，每个人都表现得过于轻率。

If we take Sightline Climate’s data from February at face value, there are 190GW of data centers planned. If we take NVIDIA CEO Jensen Huang’s statement that data centers will cost $80 billion to $100 billion a gigawatt at face value, this means that said data centers will cost anywhere from $9.5 trillion to $15 trillion. Bloomberg incorrectly states that this is a “$3 trillion” buildout.

如果我们按字面意思采纳 Sightline Climate 2 月份的数据，目前规划中的数据中心容量为 190GW。如果我们按字面意思采纳 NVIDIA 首席执行官黄仁勋关于数据中心每吉瓦（GW）成本为 800 亿至 1000 亿美元的说法，这意味着这些数据中心的建设成本将在 9.5 万亿至 15 万亿美元之间。彭博社错误地将其称为“3 万亿美元”的建设规模。

This money will have to come from somewhere. The Financial Times reported in May that banks are concerned they might “choke” on data center debt when I estimate there’s barely $250 billion a year being issued. They will, to actually make these data centers happen, have to start issuing anywhere from $500 billion to a trillion a year.

这笔钱必须有来源。英国《金融时报》5 月报道称，银行担心数据中心债务会让他们“窒息”，而据我估计，目前每年的发行量仅为 2500 亿美元左右。为了真正实现这些数据中心的建设，他们必须开始每年发行 5000 亿至 1 万亿美元的债务。

Jensen Huang has also said that NVIDIA projects a trillion dollars worth of revenue through the end of 2027. 54% of NVIDIA’s revenue comes from three clients, which means that NVIDIA’s future largely depends on three unnamed companies — likely Taiwanese ODMs building servers for Microsoft, Google and Meta — and their counterparties’ ability to raise debt on a near-perpetual basis, as the number of firms that can afford to buy thousands of $7.8 million racks of Vera Rubin GPUs is dwindling.

黄仁勋还表示，NVIDIA 预计到 2027 年底将实现 1 万亿美元的收入。NVIDIA 54% 的收入来自三个客户，这意味着 NVIDIA 的未来在很大程度上取决于这三家未具名的公司（很可能是为微软、谷歌和 Meta 构建服务器的台湾 ODM 厂商）及其交易对手能否持续不断地筹集债务，因为买得起数千个价值 780 万美元的 Vera Rubin GPU 机架的公司数量正在减少。

Even then, every part of this puzzle requires more and more debt or at the market dumps like Google’s $85 billion equity sale or Meta’s planned multi-billion dollar dump. The fact that hyperscalers are doing equity sales is, as economist Paul Kedrosky raised in our conversation on my show last week, a sign that debt is becoming harder to acquire.

即便如此，这个拼图的每一部分都需要越来越多的债务，或者像谷歌 850 亿美元的股票抛售或 Meta 计划的数十亿美元抛售那样的市场倾销。正如经济学家 Paul Kedrosky 上周在我的节目中提到的，超大规模云厂商正在进行股票抛售，这表明债务正变得越来越难获得。

Anthropic has made $330 billion in compute and chip commitments between Google, Amazon, and Microsoft, another $30 billion with CoreWeave and another $15 billion with SpaceX. To pay for this compute, Anthropic must meet its projected revenue of $174 billion a year by 2029.

Anthropic 在谷歌、亚马逊和微软之间达成了 3300 亿美元的计算和芯片承诺，与 CoreWeave 达成了 300 亿美元的承诺，与 SpaceX 达成了 150 亿美元的承诺。为了支付这些计算费用，Anthropic 必须在 2029 年前实现每年 1740 亿美元的预期收入。

Anthropic has raised $95 billion across rounds in February, April (from Google and Amazon), and May. These funds will be insufficient to cover Anthropic’s costs, as will Anthropic’s cash flow, meaning that it will have to raise at least another $200 billion in the next year.

Anthropic 在 2 月、4 月（来自谷歌和亚马逊）和 5 月的几轮融资中筹集了 950 亿美元。这些资金将不足以覆盖 Anthropic 的成本，其现金流也一样，这意味着它在明年至少还需要再筹集 2000 亿美元。

OpenAI has projected to burn at least $852 billion through the end of 2030, and has made over $770 billion in compute commitments across Microsoft, Amazon, CoreWeave, Cerebras, and Oracle. The $122 billion funding round from March will be insufficient to cover these costs, and it will require, at the very least, another $250 billion in funding by the end of the year.

OpenAI 预计到 2030 年底至少将烧掉 8520 亿美元，并已在微软、亚马逊、CoreWeave、Cerebras 和甲骨文之间做出了超过 7700 亿美元的计算承诺。3 月份的 1220 亿美元融资将不足以覆盖这些成本，到今年年底，它至少还需要再筹集 2500 亿美元。

Whatever obtuse fantasies you have about the current state of generative AI are irrelevant to a much larger problem: that the infrastructure being built and compute commitments being made are being done so at a level that demands that generative AI and AI compute generate over $2 trillion in annual revenue by 2030. When I say that, I mean it absolutely has to do that otherwise none of the data center capex makes sense, and neither Anthropic nor OpenAI can pay their commitments.

无论你对生成式 AI 的现状有什么愚钝的幻想，都与一个更大的问题无关：目前正在建设的基础设施和做出的计算承诺，其规模要求生成式 AI 和 AI 计算在 2030 年前产生超过 2 万亿美元的年收入。当我这么说时，我的意思是它绝对必须做到这一点，否则所有的数据中心资本支出都毫无意义，Anthropic 和 OpenAI 也都无法履行他们的承诺。

OpenAI expects to spend $50 billion on compute in 2026, and I wouldn’t be surprised if Anthropic spends anywhere from $30 billion to $50 billion. Between them, Anthropic and OpenAI represent the vast majority of all AI compute demand — at a minimum 70%, if not 80% to 90%. Put another way, there’s barely a few billion dollars of demand outside of two companies that lose billions — or tens of billions — of dollars a year.

OpenAI 预计 2026 年在计算上支出 500 亿美元，如果 Anthropic 支出 300 亿到 500 亿美元，我也不会感到惊讶。Anthropic 和 OpenAI 两者合计占据了所有 AI 计算需求的绝大部分——至少 70%，甚至可能达到 80% 到 90%。换句话说，除了这两家每年亏损数十亿甚至数百亿美元的公司之外，几乎没有其他几十亿美元的需求。

Let’s break down these numbers a little further: 190GW of data center capacity assuming a PUE of 1.35 suggests a critical IT load of a…

让我们进一步拆解这些数字：190GW 的数据中心容量，假设 PUE（电源使用效率）为 1.35，这意味着关键 IT 负载为……