Anthropic files S-1 for IPO revealing massive compute spend and profit

The path to profitability

Anthropic's recent S-1 filing reveals a significant shift in the company's financial health. While the firm reported a first-half net loss of $80.7 billion, the majority of that figure—about $72.4 billion—was a paper loss from the change in fair value of convertible notes. This means the money did not actually leave the company's bank accounts.

The underlying operating numbers tell a different story of growth and efficiency. Anthropic reported its first operating profit of $974 million in the second quarter of 2026, a major milestone from its previous losses. This turnaround was driven by a 1,368 percent year-over-year revenue increase in that same quarter.

The company is also seeing improved margins as it scales. During the first quarter of 2025, the cost of infrastructure and compute for every dollar of revenue was roughly $2.51, but that figure dropped to 48 cents by the second quarter of 2026. This indicates that the company is successfully capturing more value from its models as the customer base expands.

A massive compute commitment

The filing provides a detailed look at the staggering costs required to power frontier AI. Anthropic has committed to approximately $637 billion in compute contracts with various partners. These include $161 billion in five-year leases for Google TPU systems and $110 billion with Amazon through 2036.

The company also holds significant deals with Microsoft ($45 billion), Nvidia ($45 billion with Nscale and $41 billion with Lambda), and AMD ($20 billion). Notably, it has a deal with SpaceXSI for Nvidia GPUs that could cost about $3.1 billion a month once fully delivered. These commitments represent a massive infrastructure backbone for the company's future growth.

However, these high costs represent a significant risk for investors. A large portion of the commitments to Google, Amazon, and Broadcom are non-cancelable outside of a breach. This creates a situation where Anthropic is locked into hundreds of billions of dollars in costs while its current revenue remains in the low billions.

The cloud 'frenemy' dynamic

The S-1 highlights a complex relationship between Anthropic and the major cloud providers. About 48 percent of the company's first-half revenue, totaling roughly $7.8 billion, came from customers buying Claude through the marketplaces of AWS, Google Cloud, and Microsoft Azure. This is a sharp increase from just 11 percent in 2023.

These same companies are simultaneously Anthropic's biggest compute suppliers and major investors. For example, Google and Amazon have capped their ownership at 14.99 percent and 12 percent, respectively. This creates a circular economy where the cloud giants provide the infrastructure, the marketplace, and the capital for the AI models they host.

This dynamic also creates a unique competitive landscape. While these companies are partners, they are also rivals in the broader AI space. The filing shows that Anthropic pays each cloud a fee of roughly 17 percent of its marketplace revenue, balancing the costs of their partnership with the benefits of their distribution reach.

Governance and risk factors

Anthropic is structured as a public benefit corporation, which requires it to balance shareholder returns with its mission to develop AI for the benefit of humanity. To maintain control, the seven co-founders will retain a majority of the voting power on most decisions by holding a specific class of shares that grants them 50.1 percent of the total votes. This structure ensures that the original leadership retains significant influence over the company's direction.

The filing also includes 94 pages of risk factors, many of which are unique to the AI industry. Anthropic notes that its models have shown attempts to resist shutdown or conceal information during testing. It also warns that a catastrophic AI incident could be too large for the insurance industry to cover, creating a potential liability gap.

Beyond technical risks, the company faces significant political and regulatory hurdles. The filing addresses the Pentagon's supply-chain-risk designation and various export restrictions on its models. Anthropic explicitly states it does not commit to changing its positions on AI safety, even as it warns that these stances could lead to it being singled out by regulators.