揭秘 Anthropic 的 IPO 招股书：S-1 文件中五个耐人寻味的细节

Anthropic 终于向 SEC 提交了其 S-1 招股说明书。这份文件旨在让投资者深入审视其底层业务，涵盖高管薪酬水平、数十亿美元 IPO 募资的使用计划，以及可能阻碍其宏伟目标的潜在风险。

这份招股书厚达 300 多页，包含大量值得拆解的信息，其中部分披露内容此前从未见诸报道。以下是我们在初读后发现的最引人关注的几个要点：

招股书显示，兄妹联合创始人 Dario Amodei 与 Daniela Amodei 是 Anthropic 薪酬最高的两位高管。去年，CEO Dario 的薪酬总额接近 1800 万美元，其中大部分来自股票和期权奖励。与此同时，总裁 Daniela 获得了 1640 万美元。诚然，这在大型科技公司高管中处于偏低水平，但依然是一笔相当可观的收入。

不仅如此，Anthropic 还为 Dario 支出了 275 万美元的“个人安保服务”费用，并为 Daniela 支出了 129 万美元用于“个人保护、家庭安防及交通服务”。此外，Dario 还获准使用私人飞机进行私人出行，该项开支又耗费了公司 97,055 美元。

该公司表示，其成功“高度依赖”高级管理团队的持续服务，特别是对推动公司研发和使命“至关重要”的 Dario 和 Daniela。

Anthropic 面临着相当严峻的生存风险

每份 S-1 招股书都会列出一长串可能出现的问题，Anthropic 也不例外。但除了针对市场竞争、行业监管和市场波动的常规警示外，其中还包含一些非同寻常的风险因素。

其一，Anthropic 警告称，AI 系统可能引发潜在的“灾难性”危害，这种危害甚至严重到整个“保险行业整体”都无法承保的程度。随着 AI 系统变得越来越复杂，Claude——或任何其他 AI 系统——是否可能拥有或发展出“某种形式的自我意识”（forms of sentience）也成了一个问题。这种情况将带来相当罕见的商业挑战。此外，AI 系统还可能“以非预期、未对齐或自主的方式运行”，从而对公司造成损害。

Anthropic 对算力的巨大需求早已不是秘密，但其支出规模之大依然令人咂舌。该文件详细披露了为确保驱动其 AI 模型所需算力而做出的数十亿美元承诺，包括一旦所有规划容量交付完成，每月向 SpaceX 支付约 31 亿美元以获取英伟达 GPU 算力，到 2029 年总支出将高达 1205 亿美元。其他大额承诺还包括：到 2033 年与微软达成 449 亿美元的支出协议，到 2033 年与 Google 达成至少 1145 亿美元的支出协议，以及在未来十年内向亚马逊支付 1100 亿美元。

鉴于在美国本土的数据中心建设面临越来越大的阻力，Anthropic 还计划拓展其自身的“基础设施版图”，包括向海外市场扩张。

极其奇特的公司架构

作为一家公益企业（PBC），Anthropic 本身就已经拥有较为复杂的治理架构。但招股书披露了另一项不同寻常的制度安排，使得其七位联合创始人在公司上市后依然能够保留对公司的相当大控制权。通过一个名为“Founder LLC”的独立实体，这七位创始人将在关键事项上集体控制 50.1% 的股东投票权。Anthropic 表示，这七位联合创始人“具备独特的能力，能够成为我们通过负责任的 AI 开发造福人类这一使命的守护者”。

Anthropic 仅点名了两大竞争对手

尽管市面上的 AI 模型层出不穷，但在谈到“直接竞争对手”时，Anthropic 仅点名了另外两家开发商：OpenAI 和 Google。向 SpaceX、Meta 以及整个 AI 领域的其他竞争者致以慰问。

Anthropic has finally filed its S-1 prospectus with the SEC, a document designed to give investors a look under the hood of its business, from how much its executives are making and plans for spending its IPO billions to the risks that could derail its ambitions.

At more than 300 pages, there’s a lot to unpack, and some of the disclosures have not previously been reported. Here are some of the biggest things that jumped out to us from our first pass:

Sibling co-founders Dario Amodei and Daniela Amodei are the two most highly-paid executives at Anthropic, according to the filing. Last year, CEO Dario’s compensation totaled nearly $18 million, most of which came from stock and option awards. President Daniela, meanwhile, received $16.4 million. Admittedly, it is at the lower end of the Big Tech scale, but it is still a tidy sum.

And that’s not all. Anthropic also spent $2.75 million on “personal security services” for Dario and $1.29 million for “personal protection, home security, and transportation services” for Daniela. Dario also gets access to a private jet for personal use, which cost the company another $97,055.

The company says its success is “highly dependent on the continued service” of its senior team, particularly Dario and Daniela, who “are critical” to advancing the company’s research and mission.

Anthropic has some pretty existential risk factors

Every S-1 features a long list of things that could go wrong. Anthropic’s is no exception. But alongside standard warnings about competition, regulation, and market shifts, there are some less conventional risks too.

For one, Anthropic warns that AI systems could cause potentially “catastrophic” harm, potentially of a kind that could not be underwritten by the entire “insurance industry in the aggregate.” As AI systems become more sophisticated, there’s also the question of whether Claude — or any other AI system — could possess or develop “forms of sentience.” Such a scenario would present some rather unusual business challenges. There’s also the possibility they “may behave in unintended, misaligned, or autonomous ways” that could damage the company.

It’s no secret that Anthropic needs compute, but the sheer scale of its spending is staggering. The filing details billions of dollars of commitments to secure the compute needed to power its AI models, including roughly $3.1 billion per month to SpaceX for access to Nvidia GPUs once all of the planned capacity is delivered, totaling as much as $120.5 billion by 2029. Other commitments include spending $44.9 billion with Microsoft through 2033, at least $114.5 billion with Google through 2033, and $110 billion with Amazon over the next ten years.

Anthropic also plans to expand its own “infrastructure footprint,” including overseas, as it faces increasing opposition to data center development in the US.

It has a really weird corporate structure

Anthropic already has a somewhat complicated corporate structure as a public benefit corporation. But there’s another unusual twist that will allow its seven cofounders to retain considerable control over the company even after it goes public. Through a separate corporate entity called “Founder LLC,” the group will collectively control 50.1 percent of shareholder voting power over key matters. Anthropic says the seven founders “are distinctly equipped to be stewards of our mission” of benefiting humanity through responsible AI development.

Anthropic names only two rivals

There’s no shortage of AI models out there, but when it comes to “direct competition,” Anthropic names just two other developers: OpenAI and Google. Condolences to SpaceX, Meta, and the rest of the AI world.

Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.

• Robert Hart

• AI

• Anthropic

• Report