Anthropic S-1 解读：要么造出机器之神，要么破产

The Anthropic S-1 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 whether Anthropic is doing its level best to build a machine god while convincing Wall Street that it is in the business of selling IT contracts, Anthropic has just offered a handy solution: doing a bunch of things that look like you are trying to build the machine god is the only way to avoid going bankrupt.

Let's set aside the machine god for a moment. At least from my reading of the S-1, Anthropic seems to be pinning the company's future to dramatic growth; in order to cover fixed costs, it has to do so. Even if you are not an AI skeptic, you can't help but raise an eyebrow at these numbers. But if you doubt whether dramatic recursive self-improvement is possible — in the way the business plan seems to require — then this doesn't look like a good investment.

Let's start with the good news for Anthropic fans. Between the first half of 2026 and the first half of 2025, the company's revenue grew fourteen times, reaching $163 billion (16.3 billion US dollars). There is an $800 billion (80 billion US dollar) loss, but it is largely an accounting issue unrelated to the underlying business. If we strip that out along with some other charges, including employee compensation, we see $1.62 billion in adjusted income in the first half of 2026. I have previously asked whether AI can make enough money to pay for itself, and the answer given by Anthropic's S-1 seems 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 this discipline 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 focuses on tasks like coding and cybersecurity, always laser-focused on what a computer does most easily: more compute. It also targets the enterprise market, which is less fickle than consumers. Its net revenue retention rate is 900%. Pretty good! And that revenue doesn't come from consumers either. About three-quarters of customers in the first half of 2026 spend 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, but that leaderboard changes frequently. And for anyone who doesn't care that their software is four months behind, there are the open-source models.

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? 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 competing on distribution is very hard.

There are some long-term commitments that highlight how intense the expansion pressure on Anthropic is

此外，收入是间歇性涌入的，通常在新模型发布之后。在竞争压力和每个新模型发布后的使用量增长之间，Anthropic 永远不能停止训练，而且必须比体量远大于自己的竞争对手更高效地进行训练。

因此，一个重要的问题是：Anthropic 能否继续交出我们在 S-1 中看到的成果？

有一些长期承诺凸显了 Anthropic 面临的扩张压力有多强烈。当然，在第 64 页，其采购承诺和租赁协议处有一个恰到好处的空白，第 137 页上，关于其预计在 2030 年 12 月 31 日之前支出金额的地方也有另一个耐人寻味的小空白。不过，我们知道这些承诺一直延续到 2045 年。

幸运的是，有一个“承诺与或有事项”部分可以帮上一点忙。不可撤销的采购承诺有固定的最低总额 2656 亿美元，期限在 1 到 20 年之间。根据 6 月 30 日的未经审计数据，该公司还有尚未开始的租赁协议 3214 亿美元，这些协议“主要与算力和基础设施容量有关”。在此之后，Anthropic 又签署了总额约 2450 亿美元的新算力和基础设施协议，期限同样在 1 到 20 年之间。还有一些尚未最终敲定的其他交易，Anthropic 估计成本在 1250 亿至 1500 亿美元之间。此外还有其他数额可变的算力承诺。

“S-1 确实把一段快乐时光定格在了那一刻：他们的初始模型使用量飞速飙升。”

如果把这些全部加起来，我得出的是未来 20 年至少约 8500 亿美元的承诺，如果把尚未最终敲定的交易按当前预估成本计入，这一数字将升至约 1 万亿美元。这还是假设我没有遗漏任何东西——而我很可能有所遗漏。

因此，为了兑现这些承诺，Anthropic 的增长必须继续以惊人的速度推进。Anthropic 没有给出其未来承诺的付款时间表，但 PitchBook 的分析师 Harrison Rolfes 建模估算，2027 年至 2030 年间每年约为 900 亿美元，这意味着需要 1880 亿美元的收入才能保持当前的成本与算力比例稳定。“对这家公司来说，增长不是可选项，”他写道。“这些合同正是以此为前提的。”

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）花费了 23 亿美元。模型开发（以及少数其他项目）的开销要大得多；Anthropic 在这方面花费了 97 亿美元。但据 Rolfes 称，Anthropic 同时也变得更加高效——为带来每一美元收入所花费的算力成本更低了。

网络安全领域的每个人都必须不停奔跑才能留在原地，与 Anthropic 如出一辙

这是一种长期趋势吗？很难说。我们手头没有第三季度的数据。“S-1 确实把一段快乐时光定格在了那一刻：他们的初始模型使用量飞速飙升，而算力又不够充足，因此他们能够为手头的东西收取更高的费用，”Futurum Group 的分析师 Brendan Burke 说道。

作为企业软件，Anthropic 把目标对准公司的 IT 预算是合理的。而网络安全恰好是一个你不持续保持更新就绝对会栽大跟头的领域——那里人人都得拼命奔跑才能留在原地，和 Anthropic 如出一辙。当然，AI 也是问题的一部分；黑客可以用它暴力破解以前做不到的事情。此外，产品本身也让人头疼，原因是一家合作伙伴在 AI 智能体相关的安全协议上搞砸了。

Burke 指出，前沿实验室需要高利润率的产品，以避免代币定价陷入恶性价格战。这让我注意到另一件事：Opus 5.5 把客户成本比 Opus 5 降低了 40%。Anthropic 表示，成本降低意味着使用量增加，这让它得以维持利润率。如果假设每次发布新模型成本都会下降，那么用不了多久，任务的价格就会降到原价的一小部分。

Claude 已经在人类指导下构建 Claude

这未必是一个无法克服的问题，部分原因在于效率提升并不总会削减利润率，DA Davidson 的分析师 Gil Luria 说。Amazon Web Services 和其他云提供商“每年的效率都大幅提高，每年都在降价，但它们一直保持稳定增长和不断上升的利润率，”他说。

但增长的压力是无法否认的。这正是最小监督乃至零监督的递归自我改进的用武之地。“更强大的模型能帮助我们的研究人员更快地发现这些改进；反过来，更强大的算力乘数让我们能够训练出更强大的模型，并更高效地部署它们。”按照 Anthropic 关于末日之类话题的一般论调来看，它预计会迎来戏剧性的递归自我改进，也许很快就会。

Claude 已经在人类指导下构建 Claude。在其风险因素部分，Anthropic 警告说它已锁定这一路径。这伴随着明显的权衡，因为用 AI 开发更多 AI “可能导致我们模型的能力进步得比我们开发、验证和实施相应安全、对齐与安保措施的能力更快”。如果我理解正确的话，Anthropic 是在说，它的业务依赖于以极快的速度构建新模型，使用它可能并未完全理解的工作，以至于出于商业原因，它实质上被迫做一些可能给所有人带来风险的事情。这就是为什么风险因素里有一堆关于“灾难性”危害的存在性威胁以及“感知形式”发展的奇怪内容。

逃出人类监督范围的戏剧性递归自我改进有可能实现吗？不知道。但如果不能，那 Anthropic 要履行其现有的算力承诺并实现盈利。就会难得多。

关注本篇报道的话题和作者，即可在您的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