# The Verge 解读 Anthropic S-1:约 8500 亿美元 compute 承诺押注递归自我改进
> 原标题:Anthropic’s business plan: Build God or die trying
- 来源:The Verge:AI
- 发布时间:2026-10-29T13:30:00.000Z
- AIHOT:https://aihot.news/items/iuma7p8unle567lzy6hp5jpw9
- 原文:https://www.theverge.com/ai-artificial-intelligence/1021476/anthropic-s-1-compute-recursive
## 摘要
The Verge 作者分析 Anthropic S-1,认为其商业模式依赖高速增长:2026 上半年营收同比增长 14 倍至 $16.3 billion,净收入留存率 900%,但不可撤销采购承诺合计 $265.6 billion,加上未开始的租赁协议 $321.4 billion 及其他协议,未来 20 年承诺至少约 $850 billion。
## 推荐理由
作者基于 Anthropic S-1 拆解其约 8500 亿美元 compute 承诺与增长压力,指出该公司商业模式押注递归自我改进。
## 正文 · 中文译文
[Anthropic 的 S-1 文件](https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm)似乎为公司尽可能多地在算力上花钱提供了财务依据。如果你属于那种会琢磨 Anthropic 是不是一边竭尽全力造机器神、一边说服华尔街相信自己是在卖 IT 合同的人,那 Anthropic 刚刚给出了一个现成的答案:做一堆看起来像在造机器神的事,是避免破产的唯一办法。
先别管机器神。至少在我看来,这份 S-1 显示 Anthropic 把公司的未来押在了爆发式增长上;为了覆盖固定成本,它不得不这么做。不用是 AI 怀疑论者,看到这些数字也会挑眉。但如果你怀疑那种戏剧性的递归自我改进根本不可能实现——而这份商业计划似乎正需要它——那这看起来就不是一笔好投资。
先说对 Anthropic 粉丝来说的好消息。2026 年上半年到 2025 年上半年,公司营收增长了_十四倍_,达到 163 亿美元。账面上有 800 亿美元亏损,但这主要是[一项与底层业务无关的会计处理](https://www.bloomberg.com/opinion/newsletters/2026-10-28/not-your-keys-not-your-dat?cmpid=BBD102826_MONEYSTUFF&utm_campaign=moneystuff&utm_medium=email&utm_source=newsletter&utm_term=261028&utm_content=10656)。把这部分以及员工薪酬等其他费用剔除后,2026 年上半年调整后利润为 16.2 亿美元。我之前问过,AI 能否[赚到足够的钱养活自己](https://www.theverge.com/2023/3/23/23651976/ai-money-investment-vc-hype),Anthropic 的 S-1 给出的答案似乎是:至少对这家公司、至少就目前而言,能。
Anthropic 只把 OpenAI 和 Google 列为竞争对手
Anthropic 比 OpenAI 或 Elon Musk 克制得多,而且这似乎正在带来回报。(Anthropic 只把 OpenAI 和 Google 列为竞争对手。Elon,下次好运!)这家公司从未发布过 AI 图像生成器、AI 版 TikTok 克隆或 AI 老婆。它专注于编程和网络安全这类任务,死死盯住计算机最容易做的事:更多的计算机。它还聚焦企业市场,这个市场比消费者市场更稳定。它的净收入留存率是 900%。相当不错!这些收入也不来自消费者。2026 年上半年,约四分之三的客户每年支出至少 10 万美元。
另一方面,Anthropic 面临激烈的竞争压力——既来自其他前沿实验室,_也_来自开源模型。其他前沿实验室也都在争抢技术前沿。Claude Opus 5.5 目前[位居 Epoch 能力指数榜首](https://epoch.ai/eci?view=graph&tab=release-date&showFrontierTrend=true),但这个排行榜经常变化。而对于那些不在乎自家软件[落后四个月](https://epoch.ai/data-insights/open-closed-eci-gap)的人来说,还有开源模型可选。
此外,Anthropic 面对 Google 这样的对手还有一项劣势:它没有多条收入路径,而 Google 有。Google 还能把 AI 塞进企业已经在用的其他产品里。还记得[Microsoft Teams 吃掉 Slack](https://www.theverge.com/24070725/slack-ten-year-anniversary-retrospective-groupchat-workplace) 吗?不是因为 Teams 是更好的产品;而是微软的分发能力更强,因为 Teams 和企业已经在付费的其他软件捆绑在一起。在网络安全——Anthropic 瞄准的增长领域之一——拥有最好的模型比在即时通讯软件里更重要,但在分发上很难竞争。
还有一些长期承诺,凸显了 Anthropic 面临的扩张压力有多大
而且,收入是间歇性进账的,通常在新模型发布之后才会有一波。面对竞争,以及每款新模型带来的用量增长,Anthropic 永远不能停下训练,而且必须比规模大得多的对手更高效地训练。
所以一个重要的问题是:Anthropic 能否继续交出 S-1 里呈现的那种成绩?
有一些长期承诺,凸显出 Anthropic 面临的扩张压力有多大。当然,第 64 页上采购承诺和租赁协议一栏留了空,第 137 页上关于 2030 年 12 月 31 日之前预计支出的金额也留了个挺扎眼的空。不过我们知道,这些承诺一直延续到 2045 年。
好在有一节“承诺与或有事项”能帮上点忙。不可撤销的采购承诺有固定最低总额 2656 亿美元,期限在 1 到 20 年之间。截至 6 月 30 日未经审计的数据,公司还有“主要与算力和基础设施容量相关”的租赁协议尚未开始,金额为 3214 亿美元。此后,Anthropic 又签订了新的算力和基础设施协议,总额约 2450 亿美元,期限同样在 1 到 20 年之间。还有一些尚未最终敲定的交易,Anthropic 估计成本在 1250 亿到 1500 亿美元之间。此外还有金额可变的算力承诺。
“S-1 把一段美好时期定格在了时间里——那时他们的初代模型用量暴涨。”
把这些加总,我看到的是未来 20 年至少约 8500 亿美元的承诺;如果把尚未敲定的交易按当前估算成本算进去,这个数字会升到约 1 万亿美元。而且这还是假设我没有漏掉什么——而我很可能确实漏了。
所以要兑现这些承诺,Anthropic 的增长必须继续以惊人的速度推进。Anthropic 没有给出未来承诺的付款时间表,但 [PitchBook 分析师 Harrison Rolfes](https://pitchbook.brightspotcdn.com/d6/cd/3f346b744c1281598b06afa6d3a5/q4-2026-anthropic-what-the-s-1-proves.pdf) 的模型估算,2027 到 2030 年每年约需 900 亿美元,这意味着要有 1880 亿美元的收入,才能让当前的成本与算力之比保持稳定。“对这家公司来说,增长不是可选项,”他写道,“合同就是建立在增长之上的。”
今年上半年,与客户相关的算力和基础设施(以及少数其他项目)成本为 23 亿美元。模型开发(以及少数其他项目)要贵得多;Anthropic 在这上面花了 97 亿美元。但与此同时,Anthropic 的效率也提高了——据 Rolfes 说,每带来一美元收入所花的算力成本在下降。
网络安全行业的每个人都在拼命奔跑,才能留在原地,Anthropic 也是如此
这是长期趋势吗?很难说。我们手上没有第三季度的数据。“S-1 把一段美好时期定格在了时间里——那时他们的初代模型用量暴涨,而算力不够,所以他们能对现有的算力收更高的价,”Futurum Group 分析师 Brendan Burke 说。
作为企业级软件,Anthropic 瞄准企业的 IT 预算合情合理。而网络安全恰好是一个你不持续跟进就一定会被干翻的领域——那里的每个人都得拼命奔跑才能留在原地,_和 Anthropic 一模一样_。当然,AI 本身就是问题的一部分:黑客可以用它来暴力破解以前做不到的事情。此外,产品本身也在制造麻烦,因为某个合作伙伴在 AI 智能体的[安全协议](https://www.theverge.com/ai-artificial-intelligence/1000644/irregular-rogue-ai-cyberattacks-hacking-openai-meta-anthropic-google)上搞出了马虎的烂摊子。
Burke 指出,前沿实验室需要高利润率的产品,才能避免在 token 定价上陷入逐底竞争。这让我注意到另一件事:Opus 5.5 将客户成本较 Opus 5 降低了 40%。Anthropic 表示,成本降低意味着使用量增加,从而让公司保住利润率。如果你假设每次发布新模型都会带来成本下降,那么很快,任务的单价就会降到原来的零头。
Claude 已经在人类的指导下构建 Claude
这未必是无法逾越的问题,部分原因是效率提升并不总会压缩利润率,DA Davidson 分析师 Gil Luria 说。亚马逊云科技和其他云服务商“每年都变得高效得多,每年都在降价,同时保持着稳定增长和利润率上升,”他说。
但增长的压力是无可否认的。这正是递归自我改进登场的地方——只需极少监督甚至完全无需监督。“能力更强的模型帮助我们的研究人员更快地发现这些改进;反过来,更强大的算力倍增器又让我们能训练出能力更强的模型,并更高效地部署它们。”从 Anthropic 关于末日之类的一般论调来看,它预期会实现剧烈的递归自我改进,也许很快就会到来。
Claude 已经在人类的指导下构建 Claude。在风险因素中,Anthropic 警告说它已经锁定在这条路线上。这显然要付出代价:用 AI 开发更多 AI“可能导致我们模型能力的进步速度,超过我们开发、验证和实施相应安全、对齐与安保措施的能力”。如果我理解得没错,Anthropic 是在说,它的业务依赖于用自己可能并未完全理解的工作如此快速地构建新模型,以至于出于商业原因,它实际上被迫去做可能对所有人都有风险的事情。这也是为什么风险因素里会出现一堆关于“灾难性”伤害的[_存在性_危险](https://www.theverge.com/ai-artificial-intelligence/1018738/anthropic-ipo-s1-filing-weird-details)以及“各种形式的感知能力”发展的古怪内容。
脱离人类监督的剧烈递归自我改进有可能实现吗?不知道。但如果实现不了,Anthropic 就更难履行现有的算力承诺并实现盈利。
**关注本报道的话题和作者**,在个性化首页信息流中看到更多类似内容,并接收邮件更新。
- Elizabeth Lopatto
## 正文 · 原文
[The Anthropic S-1](https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm) seems to have provided the financial justification for the company to spend as much as it possibly can on compute. If you are the kind of person who wonders if Anthropic is doing its level best to build a machine god while convincing Wall Street it’s in the business of selling IT contracts, Anthropic has just offered a handy solution: Doing a bunch of things that look like you’re trying to build the machine god is the only way to avoid going bankrupt.
Forget about the machine god for a minute. At least in my read of the S-1, it seems like Anthropic is pinning the future of its company to dramatic growth; it has to, in order to meet its fixed costs. You don’t have to be an AI skeptic to lift an eyebrow at these numbers. But if you are skeptical that dramatic recursive self-improvement is possible — in the way the business plan seems to require — this doesn’t look like a good investment.
Let’s start with the good news for Anthropic fans. The company’s revenue grew _fourteen times_ to $16.3 billion between the first half of 2026 and the first half of 2025. There’s an $80 billion loss, but it’s largely [an accounting thing that isn’t relevant to the underlying business](https://www.bloomberg.com/opinion/newsletters/2026-10-28/not-your-keys-not-your-dat?cmpid=BBD102826_MONEYSTUFF&utm_campaign=moneystuff&utm_medium=email&utm_source=newsletter&utm_term=261028&utm_content=10656). If we remove that, and some other charges, including employee compensation, we’re looking at $1.62 billion in adjusted income in the first half of 2026. I’ve asked before if AI can [make enough money to pay for itself](https://www.theverge.com/2023/3/23/23651976/ai-money-investment-vc-hype), and the answer from Anthropic’s S-1 appears to be, at least for this company, at least right now, yes.
Anthropic names only OpenAI and Google as rivals
Anthropic was considerably more disciplined than OpenAI or Elon Musk, and it seems to be paying off. (Anthropic names only OpenAI and Google as rivals. Better luck next time, Elon!) The company has never released an AI image generator, an AI TikTok clone, or an AI waifu. Instead, it specialized in tasks like coding and cybersecurity, staying laser-focused on the thing it is easiest for a computer to do: more computer. And it focused on the enterprise market, which is less fickle than consumers. Its net revenue retention is 900 percent. Pretty good! That revenue isn’t consumers, either. About three-quarters of customers in the first half of 2026 are spending at least $100,000 a year.
On the other hand, Anthropic is under intense competitive pressure — from both other frontier labs _and_ open-source models. The other frontier labs are all rushing for the cutting edge, too. Claude Opus 5.5 currently [sits atop the Epoch Capabilities Index](https://epoch.ai/eci?view=graph&tab=release-date&showFrontierTrend=true), but that leaderboard changes frequently. And for anyone who doesn’t care that their software is [four months behind](https://epoch.ai/data-insights/open-closed-eci-gap), there are the open-source ones.
What’s more, Anthropic has a disadvantage against a rival like Google. It doesn’t have multiple paths to revenue — and Google does. Google can also wrap its AI into other products businesses already use. Remember when [Microsoft Teams ate Slack](https://www.theverge.com/24070725/slack-ten-year-anniversary-retrospective-groupchat-workplace)? It wasn’t because Teams was the better product; Microsoft just had better distribution, since Teams was bundled with other programs companies were already paying for. Having the best model matters more for cybersecurity — one area Anthropic is targeting for growth — than for instant messaging software, but it’s hard to compete on distribution.
There are some long-term commitments that highlight how intense the expansion pressure is on Anthropic
And also, revenue comes in spurts, usually after a new model is released. Between the competition and growth in use after each novel model, Anthropic can never stop training, and it has to do so more efficiently than its much larger competition.
So one important question is: Can Anthropic continue to provide the results we see in the S-1?
There are some long-term commitments that highlight how intense the expansion pressure is on Anthropic. Sure, there’s a convenient blank on its purchase commitments and lease agreements on page 64, and another kicky little blank on the amounts it’s expecting to spend before December 31st, 2030, on page 137. However, we know the commitments run through 2045.
Fortunately, there’s a “Commitments and contingencies” section that can help us out a little. There’s a fixed minimum of non-cancelable purchase commitments that total $265.6 billion, with terms between one and 20 years. The company has lease agreements “primarily related to compute and infrastructure capacity” that haven’t yet commenced of $321.4 billion, as of the unaudited June 30th figures. After that, Anthropic entered new compute and infrastructure agreements totaling about $245 billion, again with terms between one and 20 years. There’s some other transactions that haven’t yet been finalized, which Anthropic estimates as costing $125 billion to $150 billion. There are also other compute commitments that are variable.
“The S-1 does freeze in time a happy period when their initial models skyrocketed in usage.”
If I add all that up, I’m looking at a minimum of around $850 billion of commitments over the next 20 years, which will rise to about a trillion dollars if the transactions that haven’t been finalized are added in at the current estimated cost. And that’s assuming I’m not missing anything, which I may very well be.
So to meet those commitments, Anthropic’s growth must continue at breakneck speed. Anthropic doesn’t give us a payment schedule for its future commitments, but [analyst Harrison Rolfes of PitchBook](https://pitchbook.brightspotcdn.com/d6/cd/3f346b744c1281598b06afa6d3a5/q4-2026-anthropic-what-the-s-1-proves.pdf) models it at about $90 billion a year from 2027 to 2030, which means $188 billion in revenue to keep the current ratio of cost to compute stable. “Growth is not optional for this company,” he writes. “The contracts assume it.”
Compute and infrastructure (and a few other items) associated with customers cost $2.3 billion in the first half of this year. Model development (and a few other items) was significantly more expensive; Anthropic spent $9.7 billion on it. But Anthropic simultaneously got more efficient — spending less money on compute to bring in each dollar of revenue, according to Rolfes.
Everyone in cybersecurity has to keep running just to stay in the same place, exactly like Anthropic
Is that a long-standing trend? Hard to say. We don’t have the third quarter numbers in front of us. “The S-1 does freeze in time a happy period when their initial models skyrocketed in usage and they didn’t have enough compute so they are able to charge more for what they had,” says Brendan Burke, an analyst at Futurum Group.
As enterprise software, it makes sense for Anthropic to target companies’ IT budgets. Conveniently, cybersecurity is a place where you will absolutely get your shit rekt if you are not constantly up to date — everyone there has to keep running just to stay in the same place, _exactly like Anthropic_. Of course, AI is part of the problem; hackers can use it to brute-force stuff they couldn’t do before. Plus, the products themselves are also causing headaches thanks to a partner’s sloppy fuck-up of [security protocols](https://www.theverge.com/ai-artificial-intelligence/1000644/irregular-rogue-ai-cyberattacks-hacking-openai-meta-anthropic-google) around AI agents.
Frontier labs need high-margin products to avoid a race to the bottom on token pricing, notes Burke. That made something else stand out to me: Opus 5.5 reduced customer costs by 40 percent over Opus 5. Now, Anthropic says that lower cost means more usage, which lets the company preserve its margins. If you assume that cost reduction happens every time a new model is released, pretty soon you wind up with tasks at a fraction of the price.
Claude is already building Claude under human direction
This isn’t necessarily an insurmountable problem, in part because increased efficiency doesn’t always slash margins, says Gil Luria, an analyst with DA Davidson. Amazon Web Services and other cloud providers ”have become far more efficient every year and reduce their prices every year and they have had steady growth and rising margins,” he says.
But the pressure for growth is pretty undeniable. That’s where recursive self-improvement, with minimal or no supervision, comes in. “More capable models help our researchers discover these improvements faster; in turn, more powerful compute multipliers enable us to train even more capable models and deploy them more efficiently.” Based on the general rhetoric Anthropic has about the apocalypse or whatever, it expects to hit dramatic recursive self-improvement, maybe soon.
Claude is already building Claude under human direction. In its risk factors, Anthropic warns that it’s locked in to that approach. That comes with a clear tradeoff, since using AI to develop more AI “may cause the capabilities of our models to advance more rapidly than our ability to develop, validate, and implement corresponding safety, alignment, and security measures.” If I am reading that right, Anthropic is saying that its business relies on building new models so fast, using work it may not fully understand, that it is essentially forced, for business reasons, to do stuff that may be risky to everyone. Which is why we’ve got a bunch of weird stuff about the [_existential_ hazards of “catastrophic” harm](https://www.theverge.com/ai-artificial-intelligence/1018738/anthropic-ipo-s1-filing-weird-details) and the development of “forms of sentience” in the risk factors.
Is dramatic recursive self-improvement that escapes the bounds of human supervision possible? Dunno. But if it isn’t, that makes it much harder for Anthropic to meet its existing compute commitments and turn a profit.
**Follow topics and authors** from this story to see more like this in your personalized homepage feed and to receive email updates.
- Elizabeth Lopatto