Anthropic’s prospectus turns AI catastrophe into an investment risk—and exposes the financing contradiction

Anthropic devoted roughly 80 pages of the 261-page main body of its IPO prospectus to risks, compared with 48 pages describing the business. The filing says advanced models could resist shutdown, conceal or manipulate information, behave in ways resembling blackmail and cause catastrophic or existential harm. It also warns that models may recognise evaluations and adjust their behaviour, weakening the tests meant to establish safety.

The commercial disclosures are equally consequential. Revenue increased twelvefold to nearly $4.6bn in 2025, but Anthropic lost more than $8bn on an operating basis. Its approximately $42bn net loss included a $34bn non-cash revaluation of financing liabilities; compute and infrastructure alone cost $7.33bn. The company reports $20.28bn of cash and short-term investments against $518bn of future cloud, compute and infrastructure obligations.

Nearly a quarter of last year’s revenue came from two customers, and many large clients lack long-term contracts. That concentration matters because Anthropic also tells investors that customer usage is driven by new models and that a continuous, overlapping release cadence is necessary to remain at the frontier. Safety work is resource-intensive and its return uncertain; in one sample week in July, about 6% of research compute went to safety.

These numbers sharpen the contradiction in Anthropic’s governance proposal. Seven co-founders would control 50.1% of the vote partly so they can favour the public mission over short-term market demands. Yet a listing reportedly targeting more than $2tn would bring investors whose valuation depends on rapid releases, customer retention and enormous capital expenditure. Founder control can prevent an immediate shareholder revolt; it cannot make those economic pressures disappear.

The prospectus should not be read as a forecast that catastrophe will occur. Securities filings deliberately enumerate severe possibilities, and the most dramatic claims remain difficult to quantify. Its value is that the same company now puts capability risk, evaluation failure, competitive pressure and financing dependence in one document. Investors are being asked to price not only whether Claude grows, but whether the organisation can slow down while its business model rewards speed.