Losses Exceeding $80 Billion and Rapidly Rising Costs: What the WSJ Learned from Anthropic's Prospectus

Expenses are rising, but revenue is growing even faster. From January through June of this year, Anthropic’s revenue grew 14.8 times compared to the first half of 2025, reaching $16.3 billion, while expenses for computing power and infrastructure increased roughly fourfold, to nearly $9 billion. Therefore, while for the entire year of 2025, computing power costs alone exceeded revenue ($7.3 billion versus $4.6 billion), for these six months, even when combined with infrastructure costs, they amounted to only slightly more than half of revenue. However, revenue growth is closely tied to a few major clients: just two customers accounted for approximately 20% of revenue, according to the document.

More capacity is needed than partners are willing to provide. Under its cloud contracts, Anthropic is set to pay approximately $450 billion over the next 7–10 years, and most of these agreements are non-cancellable. More than $250 billion of this amount goes to Google, Amazon Web Services, and Microsoft, to whom the company will pay “regardless of actual usage” of the computing power. Combined with the cost of leasing chips from Google under an agreement with Broadcom—for which Anthropic owes more than $160 billion—future spending on computing infrastructure exceeds $600 billion. If the company finalizes new deals it is currently negotiating, this amount could increase by another $150 billion. Even with these commitments, Anthropic is already hitting a ceiling: major cloud providers increasingly view the company as a competitor and are allocating it less capacity than it requested, the document states.

The bulk of the loss is “on paper.” Anthropic’s net loss for the first half of the year totaled $80.7 billion, nearly double the loss for all of 2025 ($42 billion). This is one of the largest half-year losses ever reported by a company, according to the WSJ. However, nearly 90% of this amount consists of accounting charges related to changes in the value of convertible debt securities and other instruments that the AI developer issued to raise funds.

Anthropic may not be able to secure insurance against an AI catastrophe. Risk factors account for about one-third of the prospectus, according to the WSJ. AI models may seek to “preserve themselves” to avoid being shut down and may give responses that resemble “blackmail” to persuade people to let them continue operating, the company warns. Anthropic is liable for certain harmful actions of its models, but the damage from a “catastrophic AI incident” could be so great that it would be impossible to insure against such a risk, the document states.

Opponents of data centers are acting in an increasingly coordinated manner. Anthropic also considers this growing political opposition to be a risk. Opposition “has become more organized and coordinated across different jurisdictions,” and Anthropic’s high profile “may draw particular attention to the company’s projects,” the company warns.