14 Trillion, Is the Largest IPO in History Coming Again?
In August, several closed-door meetings were held in San Francisco.
Seated across the table were bankers, investors and other participants, who kept raising one question after another: Open-source models have been continuously driving down prices, how long can your gross profit margin hold up? If the pace of data center construction slows down, what will the revenue curve look like? Krishna Rao, Anthropic's Chief Financial Officer, answered in great detail, but the only question he did not respond to was — what is the company actually worth.
The meeting, known as the "trial balloon" session, is the standard procedure for large US stock IPOs, and all attendees are required to keep strict confidentiality. The real signal emerged on August 20. According to people familiar with the matter, the scale of Anthropic's offering this time will be as large as that of SpaceX, or even larger.
As reported by The New York Times, during initial contacts with potential investors, Anthropic stated that the company might raise more than 100 billion USD through an IPO, with a maximum valuation of up to 2 trillion USD.
The news mentioned that the company will release its prospectus no later than the end of August. Anthropic's underwriting group includes Morgan Stanley, Goldman Sachs and JPMorgan Chase, and some sources say Citi has also joined the group. Actions were taken as early as June: the company submitted its listing application form to the U.S. Securities and Exchange Commission back in June, and the listing can be as early as October.
The company was founded in January 2021, which is only five years from now.
If implemented as planned, a five-year young company will achieve the largest initial public offering in commercial history, and its market value will directly enter the top ten in the world.
Five Years, The Largest IPO in History
In January 2021, Dario Amodei left OpenAI. The people who left with him included his sister Daniela and more than a dozen key researchers. The reason for the split was that they were dissatisfied with OpenAI's overly aggressive commercialization strategy, and worried that their capability development was moving too fast to keep up with safety standards.
Dario is not like an average Silicon Valley CEO. He majored in physics and biology, and runs the company with the habits of a researcher. He said he spends half or even two-thirds of his time doing one thing: ensuring that the culture does not change, and that people under him can speak their minds freely. Later, Anthropic used this style to represent itself, and it also played a certain role in the financing process.
Google invested 300 million USD in 2022, and increased its investment many times afterwards, with a total investment of about 3 billion USD. The real acceleration started at the beginning of this year: after completing a round of financing in September 2025, the valuation reached 183 billion USD; by the beginning of this year, this figure had risen to about 350 billion USD; and after completing the Series H financing on May 28, the post-investment valuation reached 965 billion USD. Going back, part of the 150 billion USD promised by NVIDIA and Microsoft also flowed in this way. In exchange, Anthropic promised to purchase about 300 billion USD of computing power from Microsoft.
Within one year, the total capital invested by institutions has approached 1000 billion USD.
The listing speed is even faster than the financing speed: the company secretly submitted the listing application in June, investment banks began to arrange meetings between the management team and investors in July, roadshows were held in San Francisco and other places in August, documents were officially submitted to the SEC at the end of August, and the company went public by ringing the bell in October. There is also a revolving credit line under negotiation, which is much larger than the 100 billion USD the company originally planned. Generally speaking, this means the company has prepared more cash reserves for itself than the outside world thinks.
Anthropic is considering setting up super-voting shares, so that Dario and some co-founders will retain control of the company after it goes public. This is not a new design, but it is different when applied to this company. Dario himself holds about 2% of the company's shares, which is 40 billion USD based on a market value of 2 trillion USD; in terms of equity ratio, he is no longer the owner of this company, and only a special vote can allow him to remain in charge. The safety narrative needs to survive in the public market, and this vote is probably the only guarantee.
SpaceX issued shares at a price of 135 USD per share in June this year, raising 75 billion USD, and the total amount reached 85.7 billion USD after the over-allotment option was exercised. Before that, the record was set by Saudi Aramco's 25.6 billion USD in 2019, and the largest transaction in the United States was Visa's 17.9 billion USD in 2008. SpaceX has exceeded the sum of the top three in history, and now what Anthropic wants to do is to raise this bar even higher.
There is another set of data that can prove the importance of this matter. The total number of companies that have actually reached the 2-trillion-dollar level in commercial history is only a dozen, and Anthropic will enter the first echelon as soon as it goes public.
Dario once said in an interview that the enterprise he runs has about 2500 employees. Dividing the two numbers, the market value per capita reaches 800 million USD, which should be the "most expensive per capita" listed company in commercial history.
14 Trillion Yuan, On What Grounds?
Anthropic's preliminary data for the second quarter shows that revenue reached 11.5 billion USD, while the figure for the same period last year was 787 million USD.
The curve of annualized revenue run rate is quite steep. It was about 14 billion USD at the beginning of 2026, reached 47 billion USD in May, and exceeded 65 billion USD by the end of July. The company's total revenue in 2025 was about 10 billion USD.
In addition to growth, there is another important thing: the adjusted operating profit in the second quarter has turned positive, and the estimated value circulating in the market is 559 million USD. For a company that is still investing in cutting-edge models, the arrival of the inflection point is much earlier than people expected. According to the company's internal forecast, it can achieve operating profitability in 2026 and reach cash flow balance before 2028.
Where the money comes from can also be clearly explained. The growth engine was replaced once in the past 18 months, shifting from consumer-facing product subscriptions to high-value scenarios for enterprises and developers. More than 300,000 enterprises are using Claude, among which commercial customers account for nearly 80% of the revenue, and the number of large customers paying more than 1 million USD per year reached more than 1,000 at the beginning of the year, 8 of the top 10 Fortune companies are its customers.
It only took a few months from the release of Claude Code to achieving an annualized revenue of nearly 1 billion USD in the same year. By the second quarter of this year, Anthropic's market share in the enterprise-level large model API market has reached 32%, while OpenAI's share is about 25%.
Therefore, 2 trillion USD is not an outrageous figure on paper. If we infer the annualized revenue of 1000 billion to 1200 billion USD by the end of the year according to the calculation results of investors, the corresponding price-to-sales ratio should be between 17 and 20 times, which is even lower than that of SpaceX at the time of its listing.
The problem lies in another set of books.
In 2025, Anthropic's net loss will reach about 420 billion USD, compared with 83 billion USD the previous year, a difference of more than five times. It is easy to find where the money is burned: it is a computing power contract signed with SpaceX, which will reach hundreds of billions of dollars in scale after three years. At present, the gross profit margin is around 40%, and the company's goal is to reach more than 70% by 2028, with dozens of data centers not yet completed and a not-so-smooth power supply line in between.
There will be a very special risk factor in the upcoming prospectus. As reported by CNBC, the public's negative sentiment towards artificial intelligence and data centers will be listed separately. To put it in adult words: this company is going public when people are increasingly worried about their jobs, and what it sells is exactly the thing that people are worried about.
Another concern is related-party transactions. Amazon, Google, Microsoft, and NVIDIA are all shareholders, computing power providers and channel partners. They invest money with one hand and take back computing power orders with the other. Few people delve into such a structure in the primary market, but once it is written into the prospectus, analysts will question it item by item.
The camp represented by Dalio believes that this is a textbook-style capital carnival, and the valuation has been far away from the real cash flow; Jensen Huang insists that the computing power has been running at full load, and short-term losses are only strategic investment in the technological revolution; Bezos, Altman and others have also mentioned the statement of "benign bubble", saying that railways and the Internet developed in this way in the past.
Morgan Stanley did a calculation: in 2026, the ratio of capital expenditure to revenue of hyperscale enterprises will reach 34%, which is higher than the 32% during the dot-com bubble in 2000; within three years, the artificial intelligence infrastructure investment of the five major companies will reach 2 trillion USD.
The question Rao was asked in San Francisco is the essence of this debate. He cannot give an answer, nor can he give an answer during the quiet period.
Chinese Large Model Players Have Already Seated at the Table
In Hong Kong, the capitalization process of this track started at the beginning of the year.
It was January 8. Zhipu AI was listed on the main board of The Stock Exchange of Hong Kong Limited, becoming the world's first company whose main business is general artificial intelligence foundation large models. Its offering price was 116.20 Hong Kong dollars, raising 4.348 billion Hong Kong dollars. During the prospectus period, the margin financing provided by brokerages for it was close to 200 billion Hong Kong dollars, the public offering was oversubscribed 910 times, and it rose 13.17% on the first day of closing.
The next day, MiniMax went public. It was issued at the upper limit of the offering price of 165 Hong Kong dollars, raising about 5.54 billion Hong Kong dollars. The public offering part was oversubscribed 1837 times. The cornerstone investors included 14 institutions such as Alibaba, Abu Dhabi Investment Authority, Aspex, and E Fund, with a total subscription of about 2.723 billion Hong Kong dollars. On the listing day, MiniMax rose 109% to close at 345 Hong Kong dollars, with a market value exceeding 100 billion yuan, making it the artificial intelligence large model company with the largest IPO scale that year.
Hong Kong media called these three days the "AI Double Cannons".
There are clear accounts to calculate after the hype. Zhipu AI's total revenue in 2025 was 724 million yuan, a year-on-year increase of 131.9%, among which the annual recurring revenue of its core model and service platform was about 1.7 billion yuan, growing six times within 12 months. By July this year, the company's annual recurring revenue had reached 1 billion USD, exceeding the annual target, and growing 15 times from January to July — it only took five months to grow from 100 million USD to 1 billion USD, 15 months faster than Anthropic to achieve the same progress. The GLM-5.2 flagship model released in June relies on a sparse attention structure to support ultra-long context of millions of Tokens, and has entered the world's leading ranks in long-duration tasks and code capability tests.
MiniMax took another path. Its estimated operating revenue in 2025 was 790.4 million USD, a year-on-year increase of 159%, and its gross profit margin gradually rose from a negative value in 2023 to over 20%; by April 2026, it had achieved 400 million USD in annual recurring revenue. The company has about 400 employees, covering more than 200 countries and regions, with a total of 212 million registered users, of which overseas revenue accounts for more than 70%. Another set of data that can better reflect efficiency is: from the establishment of the company to September 2025, MiniMax spent a total of about 500 million USD, which is only 1% of the funds invested by OpenAI in the same period; and the M1 trained within three weeks with 512 H800s only cost 530,000 USD in training costs.
With the same capability, Chinese teams spend one to two orders of magnitude less money than Western teams.
On June 8, the two companies were simultaneously included in the Hang Seng Tech Index. The industry believes that this is a major event for the Hong Kong stock technology sector in 2026, representing that large model companies have officially entered the core technology benchmark of Hong Kong stocks.
The secondary market refinancing channel has also been opened. In July, Zhipu AI completed a 31.4 billion Hong Kong dollar placement at a price of 15.88 Hong Kong dollars per share, setting a new record for the single placement fundraising of Hong Kong stock technology companies this year. 55% of the funds will be used for model iteration and computing power construction; while MiniMax issued convertible bonds while placing new shares, raising a total of about 16.04 billion Hong Kong dollars, attracting the attention of many sovereign funds and long-term investors from Asia Pacific, Europe, the United States and other regions, with a subscription multiple of 7 times.
Raising 30 billion Hong Kong dollars at one time after listing is something that no previous Chinese technology company has ever achieved.
The attitude of the sell side has also improved. JPMorgan Chase raised Zhipu AI's target price from the original 1600 Hong Kong dollars to 1800 Hong Kong dollars on August 17, and continued to maintain an overweight rating, believing that the new generation of models will further expand the company's capability boundary; Daiwa Securities also initiated coverage on both Zhipu AI and MiniMax on August 12, giving buy ratings and corresponding target prices of 1500 Hong Kong dollars and 530 Hong Kong dollars respectively, on the grounds that the consumption and commercialization of AI models are accelerating; Goldman Sachs has set MiniMax's target price as high as 860 Hong Kong dollars.
The public market trading of Chinese large models has been going on for more than half a year, the revenue curve has been formed, refinancing has been completed, and there is a long queue of companies waiting to go public afterwards.
This article is from the WeChat official account "thecapital" (ID: thecapital), written by Wang Tao, and published with authorization from 36Kr.