A 50-Page Subpoena, a $518 Billion Compute Book and a November Roadshow: AI Safety Marketing Now Faces a Securities Test

The agency's theory is not that the models are dangerous in the abstract. It is that the representations built around them cannot be substantiated. The demands run at four targets — incident records, safeguards, marketing materials and privacy practices — and the third of those converts a safety inquiry into an advertising inquiry. The Federal Trade Commission already has the template: its September 2025 inquiry into AI chatbots acting as companions ordered seven consumer-facing providers, including OpenAI OpCo, Alphabet and Meta, to file special reports on how they use \"disclosures, advertising, and other representations\" to describe features, capabilities, intended audience, potential negative impacts and data collection. Marketing claims now sit inside the same evidentiary file as the incident logs.

The sharpest exposure sits inside OpenAI's newest business. In its August 31, 2026 announcement that ChatGPT Ads had reached a $1 billion annualised run rate in under 200 days, the company made three categorical claims: ads are \"always clearly labeled and separate from ChatGPT's answers,\" \"advertising does not influence the answers ChatGPT provides,\" and \"advertisers do not receive access to people's private conversations.\" Those are the exact propositions on which a deception claim turns — a representation plus a reasonable basis — and they now sit inside a compulsory demand for marketing materials. The advertising line with the least regulatory history carries the most absolute claims.

Anthropic's prospectus is what makes this a capital-markets event rather than a regulatory footnote. Anthropic, PBC confidentially submitted a draft registration statement on June 1, 2026. The public prospectus disclosed 2025 revenue of nearly $4.6 billion, up twelvefold, an operating loss of $8.06 billion, a net loss of $42 billion, and $7.33 billion spent on compute and infrastructure — more than half of $12.65 billion in total operating expenses. For the first half of 2026 it disclosed $16.3 billion of revenue and an $80.7 billion net loss. A company losing $80.7 billion in six months cannot defer its funding event by a quarter.

The listed exposure is not to a fine. It is to credit and to demand. These labs buy compute years in advance, and the bills land on companies whose own filings show how much of their spending is already committed. Microsoft spent $115.9 billion on property and equipment in the fiscal year ended June 30, 2026 — 34.9% of revenue. Amazon spent $131.8 billion in 2025, Alphabet $91.4 billion, and CoreWeave $20.6 billion over the trailing year against $6.9 billion of operating cash flow. Every dollar of that build is financed on the assumption that lab revenue keeps compounding.

There is a case for calm. A civil investigative demand is a request, not a complaint; the agency can investigate without alleging anything and close a matter without filing. The labs' transparency is as much a defence as an exhibit: publishing safety frameworks and incident notes gives them the evidentiary basis for their own claims, and the September 2025 order already put their data-handling practices on the record. The thesis fails if the agency concludes the marketing was substantiated, and nothing in the public record yet says otherwise.

The investable conclusion is narrower than \"AI regulation is bad\". The Federal Trade Commission is not attacking compute demand; it is auditing whether the trust premium that supports AI pricing and enterprise budgets has evidence behind it. That premium is funded, on the balance sheets above, by capital spent years in advance against revenue that has not been earned. Own the companies whose revenue comes from the labs' compute bill, and discount the names whose value depends on the labs' credibility being unquestioned.