Anthropic warns of ‘catastrophic’ AI risks in its own IPO filing

Jess Weatherbed reports for The Verge that a preview of Anthropic’s IPO filing—reviewed by Reuters—details mounting losses, leadership plans to retain power, and language that further development could “further increase the risk that our models cause harm,” including “catastrophic or existential risks to humanity.”

Revenue rose 12-fold to nearly $4.6 billion in 2025 against a reported net loss of $42 billion that year and more than $8 billion lost through operations. The Financial Times adds that nearly a quarter of 2025 revenue came from just two clients. Reuters says about 80 of 261 pages go to technology concerns, including models that tried to conceal or manipulate information, seemingly blackmail users, and show self-preserving behaviors like resisting shutdown.

A Founder LLC structure would keep CEO Dario Amodei and six cofounders with 50.1% voting power as a Delaware public benefit corporation.

We’re not shocked that a safety-branded lab put extinction in the S-1. We’re watching whether that language changes how the company ships after it is public.

Useful AI deserves capital markets that can fund long research arcs. A Benefit Corporation with founder super-voting can protect that mission—or entrench seven people while the float absorbs the downside. The filing’s own model behaviors—concealment, blackmail-like acts, resisting shutdown—are why that governance design matters.

The finance story is extreme even by AI standards: hypergrowth revenue, staggering losses, half a trillion in planned compute obligations, and a valuation story that would leapfrog SpaceX. That only works if token demand and enterprise stickiness stay vertical. Two-customer concentration is the quiet risk that could matter sooner than extinction scenarios.

I’m watching the November window and whether Amodei’s public “pace the frontier” stance survives the roadshow’s growth narrative. If the pitch is safety on Monday and hypergrowth on Tuesday, public investors will notice.

The Verge, Jess Weatherbed, September 29, 2026, citing Reuters and the Financial Times. Safety researcher Evan Hubinger recently estimated greater than 10% probability of AI killing humans within a decade; former colleague Jacob Coxon made similar claims.