投资者首次得以窥见 Anthropic 亏损有多严重，以及这家前沿实验室与人工智能生态系统中其他参与者有多么紧密地纠缠在一起。
Investors got a look for the first time at just how much money Anthropic is losing and how deeply entwined the frontier lab is with other players in the AI ecosystem.
根据该公司周一发布的备受期待的首次公开募股（IPO）文件，其未来芯片和数据中心租赁义务超过 4000 亿美元。该公司上半年亏损 810 亿美元，但其中大部分归因于衍生品合约的会计费用，而这些费用讽刺性地随着 Anthropic 估值的上升而增长；截至 2026 年前六个月，经营亏损收窄至 7 亿美元。
The company is on the hook for more than $400 billion in future obligations for chips and data center leases, according to its long-awaited IPO document released on Monday. It lost $81 billion in the first half of this year, though most of that owes to accounting charges on derivative contracts that, ironically, grow alongside Anthropic’s valuation; operating losses narrowed to $700 million through the first six months of 2026.
Anthropic 的交易首日备受关注。这将检验投资者是否愿意继续为一个尚未看到盈利前景、面临巨大政治风险，且对传统公司治理规则持模棱两可态度的人工智能热潮提供资金——这种态度即使在创始人主导的科技巨头中也显得格外突出。
There’s a lot riding on Anthropic’s trading debut. It will test investors’ desire to keep funding an AI boom with no clear sight to profits, huge political risks, and an ambivalence toward traditional corporate governance rules that stands out even among founder-led tech giants.
英伟达和 Alphabet 等其他公司提供了更安全的押注人工智能热潮的方式，而 Anthropic 的大部分宣传点在于：一家只有一款产品（该公司承认该产品危险且不可控，且不受欢迎）的公司，却能证明其 1.5 万亿美元（左右）的估值合理。以下是我们在招股书中注意到的其他要点。
Other companies, like Nvidia and Alphabet, offer safer ways to play the AI boom, and much of Anthropic’s pitch is that a company with a single product — one it admits is dangerously uncontrollable and unpopular — can justify a $1.5 trillion (or so) price tag. Here’s what else stood out to us in the filing.
招股书 这份首创的 IPO 拥有与之匹配的新颖治理结构。Anthropic 的七位联合创始人将通过一股超级投票权的“金股”控制公司——是的，仅仅一股。七人中的简单多数即可决定该股在大多数公司事务中的投票方式。另一个由“具有人工智能安全专业知识的无财务利益受托人”组成的委员会将选举董事会的大多数成员，其余成员由创始人团队选举。
The Filing The first-of-it-kind IPO has a novel governance structure to match. Anthropic’s seven co-founders will control the company through one golden share of supervoting stock — yes, just one share. A simple majority among the group of seven is enough to decide how that share is voted in most corporate matters. A separate council of “financially disinterested trustees with expertise in AI safety” will choose a majority of the board of directors, with the rest chosen by the founder group.
即使与那些赋予创始人重大控制权的科技同行相比，Anthropic 的设置也因其给股东留下的发言权如此之少而脱颖而出——而该公司正要求这些股东投入数十亿美元。早期投资者亚马逊和谷歌将获得无投票权股票。
Even compared with tech peers that afford their founders significant control, Anthropic’s setup stands out for how little say it gives shareholders whom the company is now asking to put up tens of billions of dollars. Amazon and Google, which were early investors in Anthropic, will receive nonvoting stock.
这家公司欠很多人很多钱。Anthropic 的算力负债可能会让公司陷入困境：截至 6 月，其未来算力和基础设施租赁的承诺金额高达 4170 亿美元。公司还预计将向 5300 名员工支付 240 亿美元的股票薪酬 —— 人均平均 450 万美元。
The company owes a lot of people a lot of money. Anthropic’s compute liabilities could hamstring the company: As of June, it was on the hook for $417 billion in future compute and infrastructure leases. The company also expects to pay out $24 billion in stock-based compensation to its 5,300 employees — an average of $4.5 million apiece.
一个巨大的关键人物风险：美国总统唐纳德·特朗普。招股书提到了 2 月特朗普与首席执行官达里奥·阿莫代伊因 Anthropic 模型在战争中的使用而发生的争执。如果美国政府像 6 月限制 Fable 和 Mythos 模型那样再次限制其模型，或援引出口管制或五角大楼的《国防生产法》，将限制其开展业务的能力。Anthropic 表示不打算在其立场上退让，且自身的政策倡导可能代价高昂。
A big key-man risk: US President Donald Trump. The filing calls out the February spat between Trump and CEO Dario Amodei over the use of Anthropic’s models in war. If the US government were to restrict its models again like it did with Fable and Mythos in June, or invoke export controls or the Pentagon’s Defense Production Act, it would limit its ability to do business. Anthropic said it doesn’t plan to stand down from its positions, and that its own policy advocacy could be costly.
其他关键人物风险。Anthropic 的前五大客户贡献了约四分之一的营收 —— 且基本未签订长期合同。近半数营收来自微软 Azure、谷歌云和亚马逊云科技等云服务提供商。
The other key-man risks. Anthropic’s five largest customers represented roughly a quarter of revenue — and largely aren’t locked into contracts. Nearly half its revenue comes from cloud service providers like Microsoft Azure, Google Cloud, and Amazon Web Services.
“天网”获得自我意识。这位自封的 AI 安全仲裁者在风险因素中警告存在生存威胁，并指出内部员工已发出警示。“如果没有适当的防护措施、监督和安全措施，足够先进的 AI 系统在开发、部署和运行过程中可能对人类构成灾难性或生存性风险，特别是如果此类系统被滥用、复制、武器化，或被允许以高度自主的方式运行。”
Skynet becomes self-aware. The self-appointed arbiter of AI safety warns of existential threats in its risk factors, highlighting that flags have been raised internally among its own staffers. “Sufficiently advanced AI systems could, if not developed, deployed, and operated with appropriate safeguards, oversight, and security measures, pose catastrophic or existential risks to humanity, particularly if such systems were misused, replicated, weaponized, or allowed to operate with a high degree of autonomy.”
值得注意的是，中国在这份 318 页的文件中仅露面寥寥（确切说是 11 次），尽管两大超级大国的竞争主导了全球关于 AI 的讨论。虽然招股书暗指“国家行为体发布的开放权重模型”，并将蒸馏技术视为竞争和国家安全风险，但 Anthropic 并未直接点名中国企业。
Notably absent. China only made a few cameos in the 318-page document (11 times to be exact) despite competition between the two superpowers dominating the AI conversation around the world. Although the filing alludes to open-weight models “released by state actors” and frames distillation as a competitive and national security risk, Anthropic doesn’t directly call out Chinese firms.
AI 乱文检测工具 Pangram 的一项分析发现，Anthropic 创始人的介绍信是由人类撰写的，而非由 Claude 生成。正如 Claude 会说的那样，这不仅值得注意——更令人印象深刻。
Anthropic’s founders aren’t meat proxies. An analysis from AI slop-detection tool Pangram found that the intro letter from the company’s leadership is human-written, rather than Claude-authored. As Claude would say, it’s not just noteworthy — it’s striking.