Anthropic IPO Prospectus Reveals: Revenue Near $4.6B, But a Quarter Comes from Two Major Clients

According to a report by QbitAI citing Reuters, the latest leaked Anthropic prospectus shows that its 2025 revenue soared to 12 times the previous figure, approaching $4.6 billion; at the same time, operating losses more than doubled, exceeding $8 billion. Put together, these numbers do not paint a picture of a company that has already figured out a profitable model, but of one expanding at extremely high speed while trading larger-scale losses for position.

What deserves even more attention is the structure of revenue sources. The prospectus did not disclose the identities of the two customers, only stating that each accounted for 12% of revenue. Based on nearly $4.6 billion, the two contributed about $552 million each, or about $1.104 billion combined—meaning any change in procurement by either one would affect more than one-tenth of annual revenue. The report judges on this basis that this could directly reduce Anthropic's revenue stability and bargaining power.

It should be noted that the above figures on revenue, losses and customer concentration all come from QbitAI's paraphrase of Reuters and other sources, making them secondhand information; the report itself did not provide the original prospectus for independent verification. These figures should therefore be understood as "what the report claims," not as facts confirmed by audit or regulators.

The report explicitly states that the identities of these two customers have not been confirmed. The Information previously reported that when Anthropic's annualized revenue was about $4 billion, its top two customers together contributed about $1.2 billion; at the time, some investors speculated that these two were Cursor and GitHub Copilot, but the report emphasized that this was only speculation.

The reason this speculation keeps coming up is that it is not entirely groundless in terms of business relationships: Cursor said in 2025 that its high-tier subscription plan was built on multi-year cooperation with model providers such as Anthropic; GitHub said it began bringing Claude into Copilot in 2024, that Claude 3.5 Sonnet was then made available to all Copilot users, and that it would subsequently continue to expand Claude's coverage in Copilot. And AI coding happens to be one of Claude's strongest application areas.

But even if the speculation holds, the problem would not be eased. The report points out that Cursor also offers models from OpenAI, Google, xAI and others, and GitHub Copilot has long since shifted to a multi-model approach, letting users and enterprises choose among Claude, GPT, Gemini and other models. This means that once Claude loses its advantage in performance, price or supply stability, customers can adjust their procurement at relatively low cost.

The report also mentions a more general risk: apart from these two customers, many of Anthropic's large customers have not signed long-term contracts and can cut or even cancel their spending on Claude at any time according to their own needs. The combination of these two variables—revenue concentration and contract duration—is the part of this prospectus that most needs to be understood.