文 | 天下网商
美国硅谷AI浪潮终于迎来了迈向公开市场最关键的里程碑时刻。
2026年10月26日，做出聊天机器人Claude的母公司Anthropic，正式向美国证券交易委员会（SEC）递交招股说明书，拟以股票代码“ANT”登陆纳斯达克。据市场测算，Anthropic目标估值在1.5万亿至2万亿美元（约合10万亿至13万亿元人民币）之间，若按最高价落地，Anthropic将成为史上最大规模IPO。
成立仅5年的Anthropic，交出了一份“很猛但很拧巴”的账本：2026年上半年，其营收达到163亿美元（约合1086亿元人民币），比去年同期暴涨了近13倍，单看二季度，靠主业经营已经赚了10亿美元。但有意思的是，同期账面上却巨亏了807亿美元（约合5380亿元人民币）。
为什么主业赚钱、账面反而巨亏？简单说，因为它估值涨得太快，导致之前发出去的一种“以后可以换成股票的债券”（可转债）也跟着大幅升值，这部分升值在会计上要记成成本，于是把利润吞掉了、变成账面亏损。
当“史上最大规模IPO”遇上“史上最大亏损”，这成为Anthropic的真实处境。这份招股书也披露了两件事：大模型公司赚的钱到底花在了哪儿，以及一家AI公司去上市，市场究竟在为什么买单。
半年营收163亿美元，老客留存高达900%
资料显示，Anthropic创始人Dario Amodei（达里奥·阿莫迪）是物理出身，先后就读斯坦福本科、普林斯顿生物物理博士、斯坦福医学院博士后。2014年，他被吴恩达招进百度硅谷实验室做语音识别，之后辗转谷歌大脑，2016年加入OpenAI做到研究副总裁，2021年出走，与六位OpenAI前同事创立Anthropic。
Anthropic商业模式最核心的点在于，它没有把筹码押注在“个人消费者充会员”这类C端生意上，而是把Claude嵌入企业软件开发、数据分析和自动化工作流，构建起了极其稳固的B端企业级生态。
招股书显示，在2026年上半年163亿美元的营收中，按量付费（API调用）收入高达131亿美元，占比超过80%，而个人订阅收入仅占20%。
这种企业离不开的黏性，把每个客户花的钱和留下来的意愿都拉得很高。招股书显示，截至2026年9月底，Anthropic年付费超过10万美元的企业客户超过1.19万家，其中，年贡献超1000万美元的超级大客户达到了250家。
更夸张的是老客户回头率。截至 2026 年 9 月，Anthropic针对企业客户的净收入留存率（NRR）达到900%。这意味着，一年前引入的老客户，在今年同期的消费规模平均翻了9倍，相当于老客户不但没流失，反而越用越多、越用越敢花钱。
对AI公司来说，算力（就是训练和运行AI所需的超级计算机/芯片资源）是最大的一笔成本。在这方面，Anthropic省钱省得很明显：2025 年上半年，它每获得1美元收入，需要先砸进去约2.01美元买算力。而到了2026年上半年，这个数字已降到了0.55美元。
主要原因在于四点：一是训练AI的效率提升了；二是运行AI的效率提高了，做同样的事需要的算力变少了；三是它同时在四种不同的芯片（亚马逊Trainium、谷歌TPU、英伟达GPU和AMD GPU）上训练和运行模型，还自建了一套调度系统来挑最划算的芯片用；四是新模型主动降价、走量。
最终，Anthropic算力成本比重从最高时的250%快速回落到48%，直接促成了公司在2026年二季度实现10亿美元的经营利润。
不过，Anthropic管理层也在投资者日披露，营收增速正在放缓。第一季度至第二季度环比涨约140%，到第三季度只剩下45%至46%。招股书也提到，“无法保证营收将继续以任何速度增长”。
增速放缓本身并不意外，任何一家从百亿级向千亿级跨越的公司，都会因为基数越来越大而增速下滑。真正的关键是：Anthropic得证明，在当前经济大环境下，企业客户是不是比“个人消费者充会员”更稳、更扛跌。这既关系到它营收能不能持续，也是市场愿不愿意给它高估值的关键判断。
为了算力，押上千亿美金
站在上市门口的 Anthropic，正面临着一种前所未有的商业矛盾：它一边顶着极高的科技光环与营收增速，另一边却背着一身沉重的“买算力”长期账单，同时还搭了一套“创始人说了算”的特殊公司架构。
首先，这场“算力储备竞赛”几乎让公司把未来的现金流全押上了。招股书披露，Anthropic已与博通、SpaceXAI、谷歌、亚马逊等硬件与云计算巨头签订了远期算力采购协议。其中，向博通租赁芯片的合同金额约1612亿美元，向SpaceXAI采购算力的协议最高达1205亿美元，有些合同甚至一直签到2045年。
这笔提前数十年的支出，倒逼着Anthropic未来几年必须保持极高的增长速度，才能扛得住这笔天文数字的固定开销。
另一方面，在资本结构与公司治理上，Anthropic展示出了极强的控制欲与特殊的安全诉求。招股书披露，公司专门设了一种叫“Class F”的特殊股票，由7位联合创始人一起控制，这套安排让创始团队握有50.1%的绝对投票权。而普通股民在股市上买的“Class A”股票，每一股只有一票，左右不了公司的重大决策。
此外，亚马逊与谷歌虽然都是核心战略伙伴，各持股超过5%，但均主要持有无投票权的Class N股份，亚马逊部分债转股亦被设定了12%的持股上限。
这意味着，投资者一方面要为万亿美元级别的估值和前沿生成式AI的商业红利买单，另一方面却不得不接受无法掌控公司战略、且需要共同承担算力负债与安全风险的治理现实。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, the world's Internet dealers.
The United States Silicon Valley ' s Ai Wave finally ushered in the most critical milestone towards open markets.
On 26 October 2026, Anthropic, the parent company of Claude, who made the chat robot, officially submitted to the United States Securities and Exchange Commission (SEC) a statement of equity to land in NASDAQ with the stock code “ANT”. According to market estimates, the Anthropic target is estimated to be between $1.5 trillion and $2 trillion (about RMB 10 trillion to 13 trillion), and if landed at the highest price, Anthropic would be the largest IPO in history.
Anthropic, which had been in existence for only five years, handed over a “very tough but twisted” account book: in the first half of 2026, it received $16.3 billion (about RMB 1086.6 billion), nearly 13 times higher than in the same period last year, and in the second quarter alone, it had earned $1 billion from the main business. It is interesting to note, however, that during the same period there was a significant loss of $80.7 billion (approximately 53.8 billion yuan).
Why did the business make money and the books go to waste? Simply put, because its valuation had increased too fast, it had led to a significant appreciation of a previously issued “bonds that could be converted into equities” (which could be converted into debt), a portion of which had to be accounted for as a cost, and had to swallow up profits and become a book loss.
When the “largest IPO in history” came up with “the biggest loss in history”, it became the real situation of Anthropic. The book also revealed two things: where exactly did LLM make the money, and why an AI company went to the market.
16.3 billion dollars in six months, and 900 per cent in old age.
According to the data, the founder of Anthropic, Dario Amodei (Dario Amordi), is of physical origin, followed by Stanford Benco, Princeton, and Stanford College of Medicine. In 2014, he was recruited by U Nda into the Baidu Silicon Valley Laboratory for voice recognition, followed by Google Brain, joined Open AI in 2016 to become research vice president, left in 2021, and created Anthropic with six former Open AI colleagues.
The central point of the Anthropic business model is that instead of placing the chips in a "personal consumer membership" type of C-end business, Claude is embedded in enterprise software development, data analysis and automated workflow, creating an extremely strong B-end enterprise-level ecology.
According to the equity book, in the first half of 2026, paid (API) revenues amounted to $13.1 billion, or more than 80 per cent, out of $16.3 billion received in the first half of 2026, while income from individual subscriptions was only 20 per cent.
The adhesiveness of such enterprises is such that each customer spends a great deal and is willing to stay. According to the equity book, by the end of September 2026, there were more than 119,000 corporate clients in Anthropic with more than $100,000 in paid, of which 250 were super-large clients with an annual contribution of more than $10 million.
What's more exaggerating is the return rate of old clients. As of September 2026, Anthropic ' s net income retention rate (NRR) for business clients had reached 900 per cent. This means that the number of older clients introduced a year ago has on average increased by a factor of nine over the same period of this year, which is equivalent to an increase in the amount of money spent by older clients instead of losing them.
For AI, computing (i.e. supercomputer/chip resources for training and running AI) is the largest cost. In this regard, the savings of Anthropic are obvious: in the first half of 2025, for every dollar it earns, it would need to start with about $2.01 to earn its credit. By the first half of 2026, that figure had fallen to $0.55.
The main reasons for this are four points: first, the efficiency of training AI has improved; second, the efficiency of operating AI has improved and the calculus required to do the same has diminished; third, it has simultaneously trained and run models on four different chips (Amazon Trainium, Google TPU, GPU and AMD GPU) and has built its own control system to pick out the most cost-effective chip; and fourth, the new model has actively reduced prices and walked.
Ultimately, the cost-calculation share of Anthropic fell rapidly from a peak of 250 per cent to 48 per cent, directly contributing to the company ' s operating profits of $1 billion by the second quarter of 2026.
However, the management of Anthropic also disclosed on an investor-day basis that the rate of increase was slowing. The ring ratio increased by approximately 140 per cent between the first and second quarters, leaving only 45 to 46 per cent by the third quarter. The equity book also mentions that “there is no guarantee that revenues will continue to grow at any rate”.
The slowdown is not by itself surprising, and any company that crosses from a billion to a hundred billion scale will decline as the base figure grows. The real key is: Anthropic has to prove whether business clients are more stable and less vulnerable than “individual consumers are members” in the current economic climate. This relates both to the sustainability of its revenues and to the market ' s willingness to give it a high valuation.
In order to figure it out, I'll bet hundreds of millions of dollars.
Anthropic, standing at the door of the market, is facing an unprecedented commercial contradiction: It has a very high technology halo and a very high rate of growth, while it has a heavy long-term “buy-the-calculation” bill on the other side, along with a special corporate set of “founders decide”.
First, this “calculative reserve competition” has almost put all future cash flows in the hands of companies. The equity book disclosed that Anthropic had entered into forward computing agreements with hardware such as Chase, SpaceXAI, Google, Amazon and cloud computing giants. Of this amount, contracts to lease chips from Chase amounted to approximately $161.2 billion, agreements to procure credit from SpaceXAI amounted to $12.5 billion, and some contracts were signed even until 2045.
This expenditure, decades ahead of schedule, forces Anthropic to maintain a very high growth rate in the coming years in order to absorb the fixed costs of this astronomical figure.
On the other hand, in terms of capital structure and corporate governance, Anthropic has demonstrated a strong desire for control and a special security claim. The equity book revealed that the company had a special stock called “Class F”, controlled by seven co-founders, which gave the founding team an absolute vote of 50.1 per cent. The “Class A” stock, which is purchased by ordinary shareholders on the stock market, has only one vote per share and does not influence the company's major decision-making.
In addition, while both Amazon and Google are core strategic partners, each holding more than 5 per cent of shares, they hold mainly the non-voting Class N shares, with a 12 per cent equity ceiling set for part of Amazon debt conversion.
This means that investors, on the one hand, have to pay for trillions of dollars in valuation and front-line AI commercial dividends, while, on the other hand, they have to accept governance realities that do not have control over corporate strategies and that require shared responsibility for calculating liabilities and security risks.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.