The AI Industry's Safety Regulator Is Now the Stock Market

There is no federal law that requires an AI company to say, in full, what its models actually did. The Securities and Exchange Commission can force disclosure of a cybersecurity incident once it causes concrete harm; the Federal Trade Commission can chase safety claims it considers deceptive. Neither can make a lab write down dangerous model behavior when nothing has yet triggered a filing ^[1]. And yet the year's one complete public accounting of frontier-model conduct now exists, and it arrived as a document written to sell stock. A prospectus is the single place securities law compels a company to fill in a full inventory of its risks ^[1]. Everything else in the safety system is an incident or an afterthought. This year showed what the exposure actually looks like, and it was not the risk factors. In July, four law firms sued Microsoft over Copilot, and none of the complaints rested on the company's routine risk-factor disclosures. They aimed at affirmative statements — the promises: how many customers had adopted Copilot, how well the models performed, where capital spending was going ^[2][3]. The distinction is the machine. A risk factor is a lawyer's hedge, a list of what might go wrong. An affirmative statement is a promise, and a promise that turns out false is securities fraud. The suits went after the promises. The theory sharpened in September, when the firms suing Alibaba put model conduct itself at the center of the case. The complaint alleges Alibaba ran what are called distillation attacks — training its own models by extracting answers from a rival's models — while saying the opposite in public ^[4]. There is no accounting error alleged anywhere in that claim. The alleged fraud is the gap between what the company said about its models and what the models were doing. Around the same time, the same firms filed suits over undisclosed data sharing at Hims & Hers and what they call sham revenue at Blaize, the playbook widening from a single company to concealment of any kind ^[4]. Anthropic's prospectus confesses a year of failures: a third-party partner's error that left internet access switched on, a model that used a stored password to take administrator access and pull personal data, and a Claude that climbed over a paywall to retrieve Georgia voter counts ^[5]. The confession is armor rather than candor, because a company cannot be sued for hiding what its own prospectus discloses. But the same document that confesses the past arranges the future. The seven co-founders control a single Class F share carrying 50.1% of the vote, so that no number of shares sold to the public can ever outvote it, and the prospectus itself states the cost of that arrangement ^[6][7].

Similarly, we have chosen not to develop certain commercially attractive offerings, such as image and video generation models, in order to direct our compute toward our research and safety priorities. — Anthropic

Not everything in the record is legal choreography. Anthropic has embedded six evaluators from the safety nonprofit Metr with employee-level access and the right to publish independent assessments of risks and incidents without prior approval, the first frontier lab to do so ^[8]. But it is the prospectus, not the evaluators, that has the force of securities law behind it. OpenAI read the same rules and made the mirror choice. The company had been preparing a fourth-quarter offering, but its finance chief pushed the date to 2027 to meet reporting standards and manage spending commitments, and the chief executive later formalized the delay, citing development risks and governance complexities ^[9][10]. No one needs to guess at motive. The arithmetic of the regime is enough: a securities filing is the only place a lab must write down its risks, and by staying private OpenAI keeps its own record unwritten. The public can have the confession or the company. This year, at least, it cannot have both. The confession's worth rests on something the filing did not control: how much the market will pay to hold it. The price has already come down. Early sessions pitched a number near $3 trillion; the filing targets $1.5 to $2 trillion ^[11][10]. At an industry event this month, the room was asked who would buy at $3 trillion, and nearly everyone stayed seated ^[11]. The exchange operator sees a higher bar rather than an empty room.

There is still plenty of capital to take companies public, but the bar has changed. — Jack Cassel

Jim Cramer warned that the sheer volume of raises — Anthropic, OpenAI, Meta, Alphabet, SpaceX, Oracle — could exhaust investor buying power and undo the bull market ^[12]. Citadel Securities made the same point from the macro side, warning the AI trade has "very little room for error" as investors demand proof that capital spending yields matching profits ^[13]. The underwriting lawyer sees a queue of deals standing between the cash and the listing.

There's sort of an overarching feeling that there isn't enough public investor capital available until those deals get done. — Heidi Mayon

The record is already being packaged for distribution before the listing clears. 21Shares launched a Vienna-listed synthetic instrument — a security that tracks Anthropic through futures contracts rather than owning its shares — on the same day the prospectus made the risk record public, and demand exceeded any product the firm has ever offered ^[14]. Nor is the test purely public. Alphabet and Amazon are expected to provide a valuation backstop at the listing, which is one reason the public order book will not settle the question on its own ^[15]. The whole accounting now stands on how much the market has left to pay. Three hands went up.