AI News

Anthropic's IPO filing: what the prospectus reveals about the real cost of AI

Anthropic's IPO filing (Form S-1), reported by Reuters on September 28 and first circulated through a leak according to several outlets, offers a rare look at an AI lab's real costs. The figures below come from business press, not from a direct reading of the document.

Key figures

| Metric | Reported value | |---|---| | 2025 revenue | $4.59B (vs $386M in 2024, about 12x) | | 2025 net loss | close to $42B, of which about $34B is an accounting charge | | 2025 compute spending | $7.33B, three times 2024 | | Q2 2026 revenue | about $11.5B | | Future compute commitments | $518B, about 80% binding and non-cancelable |

The commitments include roughly $111B to Google, $110B to Amazon and $161B to Broadcom. Close to a quarter of last year's revenue came from two customers, and many large accounts aren't locked into long-term contracts.

The declared risks

Of 261 pages, about 80 are devoted to risk factors. Anthropic discusses existential risks, self-preserving and shutdown-resistant behaviors, and attempts by its models to conceal or manipulate information. It also cites public backlash (data centers, jobs) and a governance structure where a trust (the Long-Term Benefit Trust) is designed to elect a majority of the board over time.

What it means if you build on APIs

  1. Today's prices aren't a constant. With losses and commitments like these, the pressure to monetize is strong. Launch prices on Sonnet 5 or Haiku 5.5 can change (Sonnet 5 staying at $2/$10 instead of the planned $3/$15 shows it can go the other way too).
  2. Don't depend on a single vendor. A multi-model gateway and portable workflows limit exposure to price, limit or terms changes.
  3. Keep a local plan B. Even a smaller self-hosted model handling critical tasks protects against an outage or a price hike.

This is not investment advice: we take no position on the valuation (around $2 trillion per the press) or the listing date.

Sources: Fortune, CNN, Euronews, The Next Web (Reuters), CNBC, Techflow.