Anthropic S-1 filing reveals massive compute spending and governance risks

The economics of AI scaling

Anthropic recently published its S-1 prospectus, a document filed with the SEC to prepare for an initial public offering. This filing provides a detailed look at the company's financial health and the underlying costs of running large-scale AI models. As Fortune reported, the company is currently generating more than two dollars in revenue for every dollar it spends on compute.

The company's unit economics have shown steady improvement, with a 50% increase in revenue per dollar of compute cost between the first and second quarters of 2026. In its most recent reported quarter, Anthropic maintained an adjusted operating margin of 19%. This margin figure notably excludes stock-based compensation, non-cash charitable gifts, and settlements for copyright claims from book authors.

Despite these positive trends, the company warns that profitability may face pressure in the future. Anthropic notes that as it brings more significant compute online, the expenses will be recognized before the corresponding revenue is fully realized. Additionally, the company warns that aggressive price competition from rivals could negatively impact its ability to sustain these margins.

Massive infrastructure commitments

The S-1 filing reveals that Anthropic is planning an enormous expansion of its physical infrastructure. In the first half of 2026, the company spent $5.6 billion on compute and infrastructure. To ensure future capacity, it has entered into several massive contracts with major technology and aerospace firms.

These commitments include $45 billion for data center capacity from Microsoft, $110 billion with Amazon, $121 billion with SpaceX, and $115 billion with Google. Furthermore, the company has agreements for an additional $125 billion to $150 billion in capacity, along with $45 billion with neocloud Nscale and $41 billion with Lambda. These figures represent a total of $602 to $627 billion in secured capacity.

Beyond data centers, Anthropic is also spending heavily on the hardware required to run its models. The company has at least $20 billion in commitments with AMD and a $161 billion leasing deal with Broadcom. When combined with the data center contracts, these total commitments reach as much as $808 billion. The company's 1,300% revenue growth rate in early 2026 was high enough to cover these costs, but the prospectus notes that such hypersonic expansion is unlikely to continue indefinitely.

Governance and shareholder power

Anthropic's corporate structure reveals a significant lack of trust in traditional shareholder democracy. The company places the authority over its board within its Long-Term Benefit Trust (LTBT). The five members of this trust have the power to appoint the majority of the corporate board, while the seven cofounders hold 50.1% of the voting power through a Founders LLC.

This structure means that regular investors who purchase Class A shares will almost always be out-voted by the cofounders. Their votes only count in the specific event that the cofounders are split on a decision. Additionally, major backers like Amazon and Google hold Class N shares, which carry no voting power at all.

This centralized control stands in contrast to the company's public-facing commitment to democratic values. While Anthropic works with free societies to provide a counterweight to authoritarian abuse, its internal governance ensures that the founders maintain ultimate authority. This setup allows the company to prioritize its long-term goals and safety over the desires of outside investors.

Unanswered questions and next steps

The S-1 filing leaves several major questions unanswered, most notably the absence of third quarter 2026 financial results. While the company included other recent events, critics suggest the missing data might show a less positive growth picture. For example, the prospectus states that revenue for the three months ending September 30 was 86 times higher than the same period in 2024, which equates to about $16.8 billion.

There is also concern regarding how much convertible notes are being used to mask true operating expenses. Some observers compare this to tactics used during the dot-com bubble to make cash flow look healthier. The S-1 does not provide a specific schedule for when these payments to infrastructure providers will come due, with some obligations spanning up to 2036.

Investors can expect the shares to begin trading sometime between November 9 and Thanksgiving. For those interested in how to use these technologies, Fortune is hosting a Brainstorm AI event in San Francisco from December 7-8. This event will feature a lineup of speakers to be announced soon to help companies harness AI successfully.