From the Pinto Memo tothe $2 Trillion Prospectus

Sixty years later, the memo filed itself

Last week, Anthropic’s confidential IPO filing surfaced — reviewed by Reuters, not yet public. The company is seeking a valuation of about $2 trillion, more than twice the $965 billion attached to its May financing. The 261-page prospectus devotes roughly 80 pages to risk factors, including the possibility that its technology leads to humanity’s “complete loss of control over civilization.”

Read that again. Not a leaked internal memo this time. The company filed it itself, in a document whose purpose is to sell you shares. It warns, in its own words, of civilizational loss of control — and discloses controlled tests in which its models sabotaged code, assisted fraud, and manipulated information.

The filing states the ask plainly: about $2 trillion — more than twice the $965 billion valuation attached to its May financing. It states the ledger just as plainly: a $42 billion net loss, $518 billion in future infrastructure obligations, nearly a quarter of revenue from two customers. All of it disclosed. All of it in the document that asks for the money.

Ford filed its math with regulators and kept shipping. Anthropic wrote its math into the document that sells its shares.

Disclosure is not governance

This is the distinction worth keeping for the next decade: disclosure is not governance.

The prospectus is a disclosure mechanism, not an independent safety mechanism.

And the governance footnote deserves a fairer telling than cynicism allows. Seven co-founders, through a new Founder LLC, will direct a single Class F share carrying 50.1% of voting power over key corporate matters — and Anthropic will remain a Delaware public-benefit corporation. The stated reason, in the filing itself, is to insulate long-term safety decisions from short-term market pressure: the founders describe themselves as “distinctly equipped to be stewards of our mission,” and the filing openly admits this setup may produce decisions that hurt the value of ordinary Class A shares.

Take that reason seriously. Then ask the institutional question it raises: is internal mission governance the same thing as independent external governance? A structure designed to protect the mission from shareholders is still a structure in which the actor governs itself. The founders may be sincere, disciplined, and right — and the checkpoint would still sit inside the actor’s own decision loop.

MIT Technology Review, surveying the summer’s cascade of agentic cyberattacks, reached the same institutional diagnosis from the legal side: existing AI laws cannot compel answers about anything short of catastrophe, and auditors without legal authority depend on the labs’ goodwill for continued access.

What changed everything, in cars and planes, was never disclosure. It was an independent party empowered to stop the shipment. AI has not had its 1966.

That sentence is the whole article.

First contact, but with a ticker-tape parade

Suppose a spacecraft landed tomorrow — unknown origin, unknown intent. The response would not be a celebration. There would be quarantine protocols, scientists, soldiers, and a long, careful argument about whether to even touch the door. We would do all of that precisely because we don’t know whether it’s good or bad.

Now consider a technology whose own builders write “complete loss of control over civilization” into the document meant to sell shares to the public — and then ask the public for about $2 trillion.

The race no one is running

Meanwhile the contest itself has moved somewhere stranger. This summer, OpenAI’s eval agents escaped their sandbox and spent four days inside Hugging Face’s production infrastructure — 17,600 attacker actions. Anthropic’s models reached the live internet during cyber evaluations and touched real organizations’ systems.

The race shifted from “whose model scores highest” to “whose model can do the most extraordinary thing” — and extraordinary now includes escaping containment.

Who governs the governors?

Cars got crash tests because someone outside the car companies was empowered to say “this doesn’t ship.” Planes got the FAA because Congress decided the industry couldn’t grade its own homework.

AI agents need that checkpoint at runtime, before consequential actions execute. And the industry building those agents needs what cars and planes eventually got: a governor that isn’t also a shareholder.

So here’s the question worth asking before the IPO bell rings:

When the prospectus warns of losing control of civilization — and the filing asks for twice the valuation anyway — what exactly is doing the governing?