Anthropic's prospectus reveals AI infrastructure “hard contract”: locking in 80% of Google, Amazon, and Microsoft's spending is “irrevocable”

The Zhitong Finance App learned that, according to the IPO prospectus, Anthropic is expected to jointly invest at least 518 billion US dollars in AI infrastructure construction with six partners within ten years. This plan is at the top of the scale of the disclosed AI infrastructure commitments. The AI lab stated in its prospectus that about 80% of the amount is non-cancellable or requires payment regardless of whether it is used or not. Anthropic told investors that these promises are necessary because access to computing power is becoming a key bottleneck in AI development, and future demand for advanced AI systems may exceed supply and will be “mainly limited by the availability of computing power.”

According to the IPO prospectus, Anthropic PBC's overall revenue in 2025 increased wildly to about 12 times that of the previous year, or close to 4.6 billion US dollars.

At the same time, the AI application leader that developed Claude may make the company's overall valuation exceed $2 trillion, surpass the overall valuation of SpaceX, which was founded and led by Musk at the time of the launch of about 1.77 trillion US dollars, and Anthropic's plan to raise close to 100 billion US dollars will also surpass the total amount of about 85.7 billion US dollars raised after SpaceX's overallocation. These all mean that Anthropic is expected to surpass SpaceX and become the largest initial stock offering in history.

Target Google, Amazon, Microsoft, and Broadcom

The company said it plans to pay at least 111.1 billion US dollars to Google, a subsidiary of Alphabet, 110 billion US dollars to Amazon, and 31.4 billion US dollars to Microsoft within the next 7 to 10 years, according to long-term infrastructure service obligations.

According to the prospectus, Anthropic is also responsible for approximately US$161.2 billion in equipment leasing obligations related to Broadcom, most of which are irrevocable. Anthropic secretly submitted an IPO application to the US Securities and Exchange Commission in June, but the documents have not been publicly disclosed. Anthropic did not immediately respond to requests for comment on the documents.

The company said that the promised payment period to Google is from April 2026 to July 2033, and the payment period to Amazon is from May 2026 to April 2036.

“If our actual expenses are insufficient, we must pay the difference to Google,” the company said, adding that the Amazon agreement also has similar provisions.

The company added that the $31.4 billion commitment period with Microsoft is from November 2026 to May 2033, and “cannot be cancelled unless Microsoft has committed a major unremedied breach of contract.”

Anthropic also said that the lease arrangement with Broadcom cannot be cancelled by either party unless there is a default.

Additional cooperation with xAI and AMD

The company also revealed the agreement with Musk's xAI, which may incur expenses of up to 84.5 billion US dollars for Nvidia's computing power capacity by 2029, and most can be cancelled 90 days in advance.

The documents also show that Anthropic's relationship with AMD is deepening. AMD has promised to buy up to $5 billion in Anthropic shares and provide AI computing capacity that is expected to exceed $20 billion.

The total scale described above is comparable to OpenAI's “Stargate” project. The latter is a $500 billion AI infrastructure plan, and the costs are expected to be shared by OpenAI, SoftBank, Oracle, and MGX.

Move to self-built infrastructure

Anthropic said the company is increasingly building its own AI infrastructure, moving from a pure cloud model to dedicated data centers and directly leasing chips.

The company listed its dependency on Amazon, Google, and Microsoft as a major risk. While providing computing infrastructure and distribution channels, these three companies are also developing competitive AI models.

The company said that these three tech giants simultaneously play multiple roles such as investors, customers, cloud providers, distributors, and competitors, and their interest motivations “may not be fully consistent with Anthropic.”

“If the computing power we obtain from third parties is curtailed, repriced, or terminated... our business, financial position, and operating results may be adversely affected,” the company said.