文 | 朝阳资本论
2026年10月26日，Anthropic（Claude母公司）正式向SEC（美国证券交易委员会）递交S-1招股书，预计募资超千亿美元，估值2万亿美元。
如果按估值价格进行IPO，Anthropic会是人类历史上在发售阶段市值最高的公司。
在他身前，是一连串如雷贯耳的名字：
这其中，上市超过10年的工商银行和阿里巴巴，市值巅峰较发售价相比，上涨幅度都超过了200%。
如果Anthropic同样能获取200%以上涨幅，那么其市值将超越现有记录（英伟达，2025年10月29日，盘中峰值5.37万亿美元），成为人类历史上最值钱的公司。
会有这一天吗？
不是所有的大模型公司都不赚钱
对于大部分人来说，Anthropic不是一个足够耳熟的名字，但他的拳头产品Claude，常年关注AI领域的读者一定不陌生。
简单来说，Claude是人类目前综合实力最强、最好用、有最多人真金白银付钱的AI。
对比榜单没意义，生产力不会说谎，大公司全都在用Claude，尤其是AI目前最擅长的编程工作。
风投机构Menlo Ventures在2025年末，对全美495名各大企业的AI决策人发起调查，这些人全部都能决定各自企业用什么大模型，能用多少额度以及是否付费。
数据显示，Claude占据的份额来到了惊人的54%，市场上其它所有AI的市场份额，加在一块儿都比不上Claude一家。
听起来已经很多了，但根据Anthropic递交的S-1招股书，2025年，这54%的市场份额背后，Anthropic的年度营收只有45.85亿美元。
而今年上半年，仅仅半年，Anthropic的营收已经增长到了162.86亿美元，是去年全年的3.5倍。
AI编程提供不了这么庞大的增量，营收数字的暴增意味着Claude不仅仅只是编程领域无可争议的第一，事实上它已经成为了综合实力最强的大模型。
Recon Analytics在10月做了个调查，数据表明，超过一半的Claude用户使用的是付费版，占比为55%；作为对比，Chat GPT只有29%，不到三成；Gemini更惨，只有13%。
这其中，有相当多的用户都是企业用户，据招股书，截至2026年9月30日，TTM收入≥10万美元的客户超过11900家；而2025年9月30日这个数字仅950家。
庞大的付费人群，极高的付费意愿，使得Anthropic不同于普遍认知，它已经不再是一家烧钱的AI公司，而是已经赚钱了。
2026年上半年，Anthropic录得经营现金流转正，为+39.29亿美元，根据公司自行定义的非GAAP指标显示（主要是剔除了股权激励等因素），2026年上半年，调整后营业利润为16.18亿美元，营业利润率约10%。
当然，也许会有文章告诉你，Anthropic今年上半年亏了800亿美元，但这其实只是一种会计魔术，主要源自可转换票据、衍生负债及SAFE公允价值变动损失。
根据招股书，Anthropic目前仍有本金价值约62亿美元的可转换票据，这些票据主要源于2023、24两年中亚马逊对Anthropic的融资。
亚马逊可以在2034、35年之前选择将票据转换为股份，以现有估值2万亿美元计算，这批票据价值约合1085.66亿美元，是本金的1750%。
亚马逊当然会选择要股票，因此Anthropic不得不把这笔钱记为亏损。
和亚马逊一样投资Anthropic的科技巨头还有很多，微软和英伟达在去年11月联合投资，分别是50亿、100亿美元，谷歌在今年4月追投100亿美元；AMD在今年7月追投50亿美元……
和亚马逊云计算一样，这些投资者全部都是Anthropic的供应商，Anthropic向它们采购算力，它们反手把钱投回Anthropic，以至于到现在Anthropic的账上还躺着1000多亿美元的现金及等价物。
用户暴增，营收疯长，上游供应商都青眼有加，Anthropic好像不缺钱，为什么要募资千亿？
别人亏钱，但别人“有爹”
Anthropic的创始人Dario兄妹和其它五人，全部出走自OpenAI。
具体细节很复杂，但总结起来就一句话，创始人理念不合，Dario认为需要非常警惕AI会变得像电影《终结者》里的天网那样，背叛人类。
于是他带着几个人出来，创立了Anthropic。
这使得Anthropic非常注重独立性，根据招股书，即便上市后，由Anthropic7位创始人组成的Founder LLC也会持有1股F股超级股票，对应50.1的投票权，以保证对公司的控制权。
没有哪个大公司能容忍自己说了完全不算，因此Anthropic对独立性的追求，导致他们是美国五个前沿大语言模型公司里，唯一背后“没爹”的那个。
剩下四家里，每家都背靠上市公司。
Gemini靠Google养着，根据AlphaBet（谷歌母公司）财报，2026年第二季度公司整体营业利润为407.70亿美元，尽管Gemini的财务数据没有单独列出，但谷歌家大业大，暂时还烧得起钱；
Grok被马斯克打包装进了Space X一起上市，据财报，整个Space X的Super Intelligence板块，前三季度合计亏损32.95亿美元，但这不重要，特斯拉还卖得动，马斯克没道理放弃Grok。
Muse是新玩家，在9月份面向市场推出Muse之前，Meta（三季度净利润98.6亿美元）本来在这场AI转型战争里上不了桌，因此扎克伯格十分珍视这次机会，并且打算加大投入。
最后是Chat GPT，母公司OpenAI可以说是微软一手养大的。其创始人Sam Altman在9月接受《Fortune》专访时表示公司没有上市压力，IPO计划推迟。
只要Windows还是人类最常用的操作系统，OpenAI根本不会缺钱。
也就是说，市面上的竞争对手各个都是富二代，只有Anthropic是真的只能靠自己。
但他账户上的千亿美金不是很够烧。
这源于Anthropic的算力困境。由于自身并非云服务/芯片厂商，Anthropic的算力储备严重不足。
为了让算力发展水平跟得上用户日益增长的访问需求，Anthropic在招股书里披露，他们已经跟8家算力供应商，在未来十年内签下了超过6500亿美元的算力建设合同：
--博通：1612亿，芯片与算力租赁
--SpaceX：最高1205亿，算力租赁
--谷歌：1145亿，云与算力
--亚马逊：1100亿美元，云服务
--微软：449亿，云基础设施
--Nscale：446亿，算力
--Lambda：413亿，算力
--AMD：超200亿，算力采购
更可怕的是，为了能锁定这些算力，避免巨头们为了提高自家模型的竞争力，不再向Anthropic提供算力服务，这其中有超过80%的合同是没有反悔权限的，哪怕Claude在日后因为竞争力下滑，用不了这么多算力，钱还得照付不误。
所以Anthropic必须在未来至少实打实地掏出5000多亿来，账上现在的现金只能覆盖20%，IPO完成后差不多能覆盖一半，剩下的再想办法。
如果赚钱的速度跟不上，没准还得继续增发。
美股AI科技牛还会有吗？
美股AI链上的企业，最近表现都不太好。
费城半导体指数较近期高点下跌约13%，较6月历史高点下跌21%；Alphabet在10月29日当天跌掉了12.6%，亚马逊和英特尔在29日财报公布后分别下跌7%和23%（盘后交易数据）。
正在排队IPO的企业也开始考虑延期或撤回，比如Holtec Nuclear、SB Energy、Nscale等AI产业链上游企业。
处在AI链头部梯队且正在排队IPO的，只有Anthropic一家。
整个市场都盯着Anthropic在看，根据过往经验，在这种情况下，上市前期基本上就是疯涨。
Space X就是最好的例子，发行市值1.77万亿美元，市值和时间都和Anthropic比较接近。
2026年6月12日，Space X以135美元价格上市发售，上市当日收盘大涨19.22%，次日收盘再涨19.60%，上市首周最高涨到225.64美元，首周收盘涨幅为37.04%
Anthropic大概率也不意外，市场热钱总会流向关注度最高的企业，关键是，它能涨多久，能涨多高，能带动下一波AI科技牛吗？
看起来很难。
曾经的估值第一们（SpaceX时间太短除外），之所以能维持甚至继续推高股价，是因为他们的收入，成本和商业模式都趋于稳定，但Anthropic不是，他用Claude坐上了大模型综合实力第一的宝座，不代表能一直稳坐钓鱼台。
一来是研发的不确定性。所有的科技企业在行业发展前期都必须面对研发风险，因为技术护城河积累得还不够久，后来者随时可能居上。
比如2025年，Gemini还是当时综合能力最强的大模型，到了2026年下半年，大家已经叫它“北美豆包”。
没准Claude也有这一天。
二来是中国的对手们正在农村包围城市，虽然DeepSeek等模型的综合实力略有落后，但看看价格，DeepSeek是真的香。
按低峰价格来看，Claude和GPT比Deepseek贵十倍不止，多出来的那部分性能真的值这么多钱吗？
当然，短时间内，特朗普正在大力推行的“美国国产替代”概念应该还不会破产，Claude短时间内在美国市场还不会面临来自中国的威胁，但一旦未来有政策变动，其收入恐怕马上要被打上问号。
Anthropic倒是表现得很有信心。在招股书里，它们说自己是在一个高达40万亿的市场里抢饭吃，换算成人民币，应该是266万亿人民币，A15，可以给133个许老板还债。
人类历史上还没见过敢这么吹的。
当然，Anthropic自有其逻辑，按照他们的说法，这个价格是人类所有脑力劳动者产生的价值总和，而AI终有一天能取代所有的脑力劳动者。
如果真的有那一天，恐怕人类历史上市值最高的公司真的会是一家AI公司，但是不是Anthropic，那就要看它的造化了。
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。
Zhongyang Capitalism
On 26 October 2026, Anthropic (Claude parent company) officially submitted to the SEC (United States Securities and Exchange Commission) the S-1 book of solicitation, which is expected to raise more than $100 billion, valued at $2 trillion.
If IPO were to be valued at prices, Anthropic would be the company with the highest market value in human history at the distribution stage.
In front of him, it was a string of names:
Of these, commercial banks and Alibaba, which have been on the market for more than 10 years, have increased by more than 200 per cent compared to the sale price.
If Anthropic could also obtain 200 per cent to increase, its market value would go beyond the current record (5.37 trillion dollars at 29 October 2025) and become the most valuable company in human history.
Will this be the day?
Not all LLM companies make money.
For most people, Anthropic is not a sufficiently familiar name, but his fist product, Claude, is a familiar reader in the field of AI.
In short, Claude is the one who currently has the strongest combined power, best used, and most people pay for it.
The list is meaningless, productivity is not a lie, and the big companies are all using Claude, especially AI's best programming at the moment.
In late 2025, Menlo Ventures launched a survey of 495 AI decision makers in major enterprises across the United States, all of whom were able to determine what LLM was used in their respective enterprises, how much quota was available and whether paid was available.
Data show that Claude ' s share has reached an alarming 54 per cent and that the market share of all the other AIs in the market, added together, is less than the Claudes.
It sounds like a lot, but according to the S-1 book submitted by Anthropic, in 2025, after 54 per cent of the market share, Anthropic received only $4.585 billion annually.
In the first half of this year, only half a year has seen an increase in the income of Anthropic to $16,286 million, which is 3.5 times the figure for the entire year.
AI’s programming does not provide such a huge increase, and the surge in collections means Claude is not just the first in the programming field, but in fact it has become the most integrated LLM.
In October, Recon Analytics conducted a survey, which showed that more than half of Claude users used paid, or 55 per cent; in contrast, Chat GPT was only 29 per cent, or less than 30 per cent; Gemini was worse, at 13 per cent.
Of these, a considerable number were business users, and according to equity books, as at 30 September 2026, more than 11,900 clients had tTM revenues of $100,000, compared to 950 as at 30 September 2025.
The huge paid population, the very high paid will make Anthropic different from the general perception that it is no longer an AI company that burns money, but has made money.
In the first half of 2026, Anthropic recorded a positive cash flow of +3.929 billion, and in the first half of 2026 the adjusted operating profit was $1.618 billion, with an operating profit margin of about 10 per cent, according to the non-GAP indicators defined by the company (mainly excluding equity incentives).
Of course, there may be an article that tells you that Anthropic lost $80 billion in the first half of this year, but it is simply an accounting trick, mainly arising from convertible instruments, derivative liabilities and SAFE ' s loss of fair value.
According to the equity book, Anthropic still has some $6.2 billion worth of convertible instruments in principal, mainly derived from the financing of Anthropic by the Central Asian Amazon in 2023 and 24 years.
By 2034 or 35 years, the Amazon could have opted to convert the instruments into shares, valued at approximately $108,566 million, or 1750 per cent of the principal, based on the current valuation of $2 trillion.
The Amazon would of course have opted for stocks, so Anthropic had to record the money as a loss.
The technology giants that invest in Anthropic, like the Amazon, are many, Microsoft and Ingveida combined to invest $5 billion and $10 billion, respectively, last November, Google spent $10 billion in April, and AMD spent $5 billion in July.
As with the Amazon cloud calculations, these investors are all suppliers to Anthropic, who purchase credit from them, and who instead throw their money back to Anthropic, so that now there is more than $100 billion in cash and equivalent on Anthropic's books.
Why raise hundreds of millions of dollars?
Others lose money, but others "have a father."
The founders of Anthropic, the Dario brothers and sisters and five others, all left Open AI.
The details are complex, but to sum up one sentence, the founders do not agree, and Dario believes that the need to be very vigilant about AI will turn into a betrayal of humanity, like the Skynet in The Terminator.
So he came out with a few people and created Anthropic.
This has made Anthropic very independent, according to which, even after listing, the Founder LLC, consisting of seven founders of Anthropic, holds a share of a F stock, corresponding to 50.1 voting rights, to guarantee control of the company.
No big company can tolerate its own words at all, so Anthropic's quest for independence leads them to be the only one behind the five front-line big language model companies in the United States.
Each of the four remaining families is attached to the listed company.
Gemini is dependent on Google, which, according to the AlphaBet (Google Mother) financial paper, had an overall operating profit of $40,770 million in the second quarter of 2026. Although Gemini ' s financial data are not separately listed, Googlers are large and can afford to burn for the time being;
Grok was put on the market together with Space X. According to the financial report, the entire Super Intelligence plate of Space X had a total loss of US$ 3,295 million during the first three quarters, but that was not important, and Tesla was selling well and Mask had no reason to give up Grok.
Muse was a new player, and before September, when it was introduced to the market, Meta (three quarters net profit of $9.86 billion) would not have been on the table in the AI transition war, Zuckerberg valued the opportunity and intended to invest more.
Finally, Chat GPT, the parent company Open AI could be said to have been raised by Microsoft. Sam Altman, its founder, said in September that the company had no market pressure and that the IPO plan had been postponed.
As long as Windows is the most common operating system in humans, Open AI will not be in need of money.
In other words, all the competitors on the market are rich, and only Anthropic is really on his own.
But the hundreds of billion dollars in his account were not enough to burn.
It's from Anthropic's math dilemma. Anthropic ' s computing reserves are severely inadequate because it is not a cloud service/chip manufacturer.
In order to keep up with the growing demand for access by users at the level of computing development, Anthropic disclosed in his book that they had signed over $65 billion in construction contracts with eight computing suppliers over the next decade:
-Bop: 161.2 billion, chips and computing leases
--SpaceX: Up to $12.5 billion, computing lease
- Google: 11.45 billion, clouds and computing
- Amazon: $11 billion, cloud services
- Microsoft: 44.9 billion, cloud infrastructure
--Nscale: 44.6 billion.
- Lambda: 41.3 billion.
- AMD: Over $20 billion in computing
Worse still, in order to lock these calculations in, and avoid the giants from providing any more calculator services to Anthropic in order to improve the competitiveness of their own models, more than 80 per cent of these contracts are unretroverted to permission, even if Claude will not be able to use so much of the calculus in the future because of his competitiveness.
So Anthropic would have to come up with at least $500 billion in cash in the future, now covering only 20 per cent of the cash, almost half of the IPO, and the rest.
If the money doesn't keep up, it'll have to go on.
Is there any more U.S.-A. technology?
Businesses in the U.S. stock chain have not been doing well lately.
The Philadelphia semiconductor index fell by about 13 per cent over its recent highs and 21 per cent over its June historical highs; Alphabet fell by 12.6 per cent on 29 October and Amazon and Intel fell by 7 per cent and 23 per cent (post-discretion transaction data), respectively, after the 29th edition.
Enterprises that are queuing IPOs are also considering extension or withdrawal, such as those upstream of the AI industrial chain, such as Holtec Nuclear, SB Energy and Nscale.
Only the Anthropics are in the Al-Stairwell and in line for IPO.
The whole market is staring at Anthropic, and according to past experience, in this case the pre-market period was basically a boom.
Space X is the best example, with a market value of $1.77 trillion and a closer market value and timing to Anthropic.
On 12 June 2026, Space X was placed on the market at $135, and the closings rose by 19.22 per cent on the same day, and the closings rose by 19.60 per cent on the next day, reaching a maximum of $225.64 in the first week and 37.04 per cent in the first week.
Anthropic's probably not surprising, and the market's hot money always goes to the most focused firms. The question is, how long will it go, how high will it go, and how high will it drive the next wave of AI-based technology cows?
Looks hard.
The first of the valuations (except SpaceX time was too short) was maintained and even continued to push stock prices because of the stabilization of their income, costs and business patterns, but not because of the fact that anthropic had taken the first seat of LLM combined strength with Claude, which did not mean he had been able to sit on a fishing platform.
One is the uncertainty of research and development. All S & T enterprises have to face R & D risks before the development of the industry, as technology has not accumulated long enough to become available at any later stage.
In 2025, for example, Gemini was also the most integrated LLM at the time, and in the second half of 2026 it was called “Doubao North America”.
Maybe Claude has one.
Second, Chinese opponents are encircling cities in rural areas, and while the combined power of models like DeepSeek is slightly behind, look at the price, DeepSeek is a real scent.
At a low price, Claude and GPT are more than ten times more expensive than Deepseek.
Of course, in a short period of time, the concept of “national replacement for US production,” which Trump is promoting, should not go bankrupt, and Claude will not face threats from China in a short time on the US market, but his revenues will be called a question mark as soon as policy changes take place.
Anthropic did show confidence. In the book, they said that they were eating in a market of up to 40 trillion yuan, convert in RMB, which should be 26.6 trillion yuan, and A15, which could repay the debt to 133 employers.
Never before in human history have you dared to blow like this.
Of course, Anthropic has its own logic, according to which the price is the sum of the value produced by all human brain workers, and AI will one day replace all brain workers.
If that day had come, the company with the highest value in human history would have really been an AI company, but not anthropic, depending on how it was made.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.