Behind Anthropic's confidential IPO filing: The urgent need to redeem itself with computing power

In June 2026, Anthropic secretly submitted its S-1 prospectus to the SEC, with a valuation of $965 billion. Just a week ago, it completed a $65 billion Series H financing round. Almost at the same time, OpenAI was also preparing to file its prospectus. These two leading AI companies are rushing to become the "first large model enterprise to go public in the United States."

During the same period, Zhipu AI and MiniMax are intensively preparing for their A-share listings. It has only been five months since they rang the bell on the Hong Kong Stock Exchange in January 2026. In the capital market, the listing cycle of a company is usually measured in years, but these two Chinese AI companies clearly can't wait.

Why are global AI companies flocking to go public in 2026? The answer is neither the complete maturity of technology nor the accidental opening of the secondary market window. There is only one core truth: The high - cost computing power bills are due, and the scale of private financing is no longer sufficient to cover the continuous iteration costs. The industry must turn to the vast capital in the public market to survive and break the deadlock.

1

Anthropic's S - 1 did not disclose the complete financial details, but SpaceX's prospectus directly revealed the real survival situation of leading large models.

Regulatory documents show that Anthropic pays SpaceX $1.25 billion per month to exclusively rent all the computing power of the Colossus 1 data center in Memphis, covering more than 220,000 AI chips. The annual computing power bill is nearly $15 billion. Based on the industry - disclosed annualized revenue run - rate of $47 billion, nearly one - third of Anthropic's revenue continuously flows to Musk's data center.

Compared with the financial pressure, the stability of computing power supply is even more important. Although the paper term of the contract between the two parties is until 2029, Musk publicly stated on the X platform that this is not a long - term and stable cooperation. Once SpaceX's own xAI business faces a shortage of computing power, it has the right to take back the computing power resources at any time. The initiative of the contract is completely in the hands of SpaceX, and Anthropic is always in a passive and dependent state.

This is the real reason why Anthropic, despite having a valuation of tens of billions and huge financing, is still in a hurry to rush for an IPO. The $65 billion Series H financing may seem large in scale, but due to the continuous high - cost and rigid procurement of computing power, most of the funds cannot stay on the books for a long time. Within 12 months, they will continuously flow back to upstream computing power providers such as SpaceX, AWS, and Google Cloud in the form of computing power procurement.

It's like just earning a transfer fee.

The scale of a single round of private financing can never catch up with the exponential growth of computing power costs. Moreover, after each round of financing, the scale of computing power procurement and the degree of external dependence increase simultaneously. In essence, the more you try to survive, the more you get tied up, and the more passive you become. The public capital pool in the open market can provide a sufficient amount of funds to support the long - term investment in building a self - owned computing power base at one time. It can also help enterprises establish independent capital credit and gradually get rid of the passive dependence on a single computing power supplier.

Private financing is for survival, while going public is for a new life. Anthropic's decision to file its prospectus at the time when its book funds were most abundant actually shows that it has seen through the industry's fate: If the $65 billion financing is still only used to pay the computing power bills, a larger - scale financing and more demanding cooperation conditions will be required in the next round. It will become more and more difficult to break free from the computing power shackles in the future, and it will be impossible to recover the early investment in the foreseeable future.

OpenAI's situation is highly similar to that of Anthropic, only with a different commercialization path. Altman once admitted that "the outcome of the AI competition does not depend on who goes public first," but the real financial data of the industry is extremely cruel.

According to The Information, in Q1 2026, for every $1 of revenue generated by OpenAI, it incurred a loss of $1.22. With the overall ecological layout of the entire track and huge costs for model training and inference computing power, the overall rate of burning money is even higher than that of Anthropic. The product forms of these two top - level AI enterprises are very different. One focuses on the general intelligent entrance, and the other focuses on programmer productivity. However, their underlying dilemmas are completely the same: They do not have independent control over the core computing power and are always working for upstream computing power giants.

2

In the Chinese market, the second - round listing rush of Zhipu AI and MiniMax is a regional variation of the global AI computing power - driven logic. In January 2026, the two enterprises went public on the Hong Kong Stock Exchange. Zhipu's stock price rose by 13% on the first day, and MiniMax soared by 110%, once setting off a market upsurge. But just five months later, the two enterprises quickly launched the A - share listing guidance, which is enough to show that the first - round fundraising on the Hong Kong Stock Exchange has already failed to keep up with the pace of business expansion - in essence, it has failed to keep up with the speed of capital consumption for computing power procurement.

Compared with AI enterprises in the US stock market, the computing power environment of domestic AI companies is relatively stable. There is no pressure of a single monthly computing power rent in the tens of billions of dollars, and the risk of extreme dependence on a single supplier is significantly lower. The competition pattern of domestic cloud providers is fragmented, and the stability of computing power supply is stronger. However, the scarcity of high - end computing power and the continuous rise in the cost of technical talents are common bottlenecks in the global AI industry.

The valuations of Zhipu and MiniMax in the tens of billions of Hong Kong dollars after their listing on the Hong Kong Stock Exchange, corresponding to the fundraising scale, are under pressure to support the long - term development in the face of the continuous computing power investment in top - level large models. They are just entry - level tickets. Rushing for an A - share listing essentially means raising funds for the second - round computing power iteration and global expansion, and making up for the capital short - board for long - term development.

The differences in the surface rhythm of the listing race between Chinese and US AI companies are clearly visible: US stock enterprises are competing for the first - round listing seats in the industry to seize the industrial discourse power; domestic enterprises are conducting relay financing on the A + H dual platforms to consolidate the foundation for long - term development. However, the underlying driving forces are completely the same: The growth rate of AI computing power costs has completely exceeded the supply rate of private financing.

According to the prediction of Anthropic CEO Amodei, the cost of a single top - level model training may approach $10 billion by the end of 2026, while the estimate of the academic institution Epoch AI is about $1 billion in 2027. No matter which set of parameters is used, no single round of private financing can independently bear such a high - cost iteration investment. Going public on the open market and relying on public capital to break through the computing power shackles has become the only way out for all leading AI enterprises. Going public is not the end but the starting point of breaking free from the computing power shackles.

3

Anthropic's public announcement did not disclose the issue price or a clear listing schedule, but it completely changed the competition rules of the global AI industry. Whoever completes the capitalization in the public market first will break through the dual shackles of computing power and capital first and set the subsequent development passing line for the industry.

It is not the player with the largest industry scale or the highest market voice, but the first one to see through the essence of the industry: The computing power bills will not wait for enterprises to grow. The longer the path of relying on private financing for survival, the deeper the dependence on external computing power, and the smaller the industry's fault - tolerance space.

In 2026, it is by no means an industrial celebration of the maturity of AI technology, but a collective action of global AI enterprises to break free from the shackles of computing power costs. US stock enterprises use listing to break free from the computing power bondage and compete for industrial autonomy; domestic enterprises rely on dual - platform relay financing to support the global computing power layout and long - term model iteration.

The global tracks are different, and the rhythms are different, but the ultimate law has never changed: Behind all the glamorous AI capitalization narratives, there is a life - and - death breakthrough about computing power and capital that has to be made.

This article is from the WeChat official account "AI Goes Against the Grain", author: Changqing, published by 36Kr with authorization.