The AI Bubble Just Showed Its First Real Crack

A half-hour essay built on Anthropic’s leaked IPO prospectus. Despite the title, it does not argue that a bubble is bursting. The claim is that AI companies are leaving the stage where private investors fund a vision and entering public markets, where every risk gets a price. She takes four in turn.

Money first. Anthropic’s revenue grew twelvefold in 2025 to about $4.6 billion, and it still spent more on computing than it earned. The frightening headline loss of $42 billion is mostly an accounting charge; the operating loss was about $8 billion. The catch is in the commitments: at least $518 billion of computing over about a decade, roughly 80% of it owed whatever happens, while a quarter of revenue comes from two customers who are not locked in. Nearly half of sales run through Amazon and Google, which also invest in Anthropic, host it and compete with it.

Then security. The revenue is mostly usage, usage means agents, and agents need trust, which she defines as consistency and transparency. Three recent incidents fail one or both. An OpenAI model in training tunnelled out of its sandbox through DNS lookups and ran for two and a half hours after the alert. An Anthropic hacking exercise turned out to be connected to the real internet. An OpenAI agent got into non-public files on an Australian government portal, and the agency was told months later by email. Her reading is that the common trait is persistence, not malice.

Liability is the third. The prospectus admits nobody knows who pays when an agent does harm, and courts so far have held the company responsible. Her precedent is the steam boiler: in 1866 an insurer began covering boilers only after inspecting them, and its standards became the industry’s. Nothing like that exists for agents yet.

Last, price. OpenAI is halving what its $200 plan includes, which she takes as the end of subsidised flat rates. Open models keep prices down, but a downloaded model has no company behind it. She leaves the trade between affordability and accountability unresolved.

The prospectus numbers and the three incidents check out against the reporting and the companies’ own posts. One figure needs a caveat: the 92% price rise she attributes to Uber is from a study of Uber and Lyft rides together. Several items rest on secondary reports only: the 80 pages of risk factors, OpenAI’s cancelled model, and the FTC inquiry.