On October 26, Dario Amodei filed to take Anthropic public. Reports put the target valuation between $1.5 and $2 trillion, which would make a four-year-old company that rents its computers one of the most valuable enterprises on Earth. So, Dario, since you are asking the public for two trillion dollars, let’s do the inventory your essays never do.
What does your company own? It owns a model — weights and know-how. It owns an enterprise customer base that produced $16.3 billion in revenue in the first half of 2026. And it owns a bank account: the company has raised $141 billion and was holding about $112 billion in cash at mid-year.
What does Anthropic not own? A data center. Not one. There is no platform underneath this company — only a rented fleet of accelerators on a scrapping cycle of about five years, hardware the landlords will write down and replace while the contracts Anthropic signed against it are still running. The machines its product runs on belong to Alphabet, Amazon, and SpaceX, to whom Anthropic has committed more than $475 billion in future compute spending. Nearly half its revenue — 48% in the first half of 2026 — arrives through those same landlords’ marketplaces, which collect at the register. And the product those machines produce is sold, in close equivalents, by OpenAI, by Google, by Meta, and by open-weight releases anyone can download. The second-largest company on Earth, by the asking price, is a tenant reselling rented time on depreciating hardware, in a market where everybody else has a model too. You have nothing, Dario. Hold that shape in mind, because everything else in the filing hangs off it.
The loss is the landlord’s gain
The headline number from the S-1 is an $80.7 billion net loss for the first half of 2026. Read one line further and the loss changes character. Some $72.4 billion of it is a non-cash charge: the change in fair value of Anthropic’s convertible notes. Anthropic elected to mark those notes to market, so when the company’s estimated value rises, the notes are worth more, and the increase is booked as a loss.
Now ask who holds the notes. As of June 30, every convertible note Anthropic had outstanding — all of them — was held by Amazon: $6.2 billion of principal, carried on Anthropic’s books at $108.6 billion. The famous loss is, in substance, a transfer entry: the higher the price the market puts on Anthropic, the larger the paper gain accruing to its landlord, and the larger the “loss” the company reports for having become more valuable. The student is grading his own homework so hard that the grade itself manufactures an $80 billion expense — and the expense is payable, in equity, to the company that owns the building.
The terms deserve the same slow reading. Amazon’s notes convert after the IPO into as many as 1.84 billion nonvoting shares, at conversion prices fixed in 2023 and 2024 — $2.09 and $5.05 a share, against a Series H price near $59. Conversion is capped so Amazon holds no more than 12% of the capitalization at a time, which sounds like a safeguard until the next clause: Amazon may waive its own cap. Whatever cannot convert in one quarter rolls over and converts in a later one. Amazon already holds roughly a fifth of the company on a fully diluted basis by outside estimates, has committed another $110 billion worth of the company’s spending to its own cloud, and can expand its stake on a schedule without persuading a single public shareholder. The founders, for their part, retain 50.1% of the vote through dual-class stock — a majority margin of one-tenth of one percent — and have pledged 80% of their personal equity to charity, a commitment that will move super-voting shares into ordinary hands over time. Control in this company is a wasting asset. The landlord’s position is a ratchet. The filing says all of this, in print, to anyone who reads past the nouns.
Control in this company is a wasting asset. The landlord’s position is a ratchet.
Seeded by fraud
There is a second provenance in the prospectus, and Anthropic’s own document surfaces it. In 2022, Sam Bankman-Fried’s Alameda put $500 million into Anthropic’s Series B — money prosecutors argued at his trial was drawn from customer deposits. He is serving 25 years for that fraud. The stake, roughly 13.6% at purchase and diluted to 7.84%, passed to the FTX bankruptcy estate, which sold it in 2024 for about $1.3 billion to make defrauded customers whole. Shares bought personally by Caroline Ellison and Nishad Singh were seized; the S-1 itself records that the U.S. government sold 4.45 million of those shares for $627 million to reimburse creditors. At the valuation Anthropic now seeks, the divested FTX stake alone would be worth on the order of $100 billion on paper.
None of this is hidden, and none of it is alleged here for the first time. It is simply never assembled. The temple of AI safety was seeded, in part, with adjudicated fraud proceeds, compounded on rented infrastructure financed by its landlords, and is now being priced by a process in which the largest beneficiaries of a high price are the parties who supplied the money, the machines, and the marketplace. Then the priced shares will flow, through index funds, into the retirement accounts of people who never read a word of the filing.
And when it reprices
The balance sheet deserves one honest paragraph, because it is the whole argument in miniature. It balances — accounting identities always do. What it cannot do is stand up. On one side sits the cash: $112 billion, which arrived as financing. On the other side sit the notes, carried at $108.6 billion against $6.2 billion of principal — a liability whose size is a function of the company’s own estimated value, so that the balance sheet inflates on both sides from a single input: the price the interested parties put on the company. Beneath the sheet entirely sit the real obligations, more than $475 billion in compute commitments, owed to the landlords, against $16.3 billion of revenue in the company’s best half-year ever. A balance sheet whose assets are raised money, whose largest liability is a mirror of its own valuation, and whose true commitments live off the page, owed to the people holding the mirror — that is not a foundation for the second-largest valuation on Earth. It is a structure that adds up only while nobody says no.
Somebody eventually says no. Consider what the number is wired into by then. A company priced at $1.5 to $2 trillion does not sit quietly in a corner of the market: it enters the major indices at a heavyweight’s share, joining an index whose top ten names already account for a historically extreme share of its total value, and index funds buy it automatically, in proportion, with the savings of everyone who never chose it. Its marks sit inside its landlords’ earnings — Amazon carried its Anthropic positions at figures approaching $200 billion across its preferred shares and notes by mid-2026, and booked tens of billions of paper income from their appreciation in a single quarter; Alphabet’s books carry the same exposure in smaller type. Its contracts underwrite a data-center construction boom financed against those commitments. So when the grade is finally revised — and grades issued by interested parties are always, eventually, revised — the loss does not stay inside one ticker. It prints simultaneously in the index in your retirement account, in the earnings of the two largest landlords, and in the collateral assumptions under half a trillion dollars of construction. The question is not whether a $2 trillion tenant can fall. It is how many balance sheets have been taught to stand on its price.
The grade is the product
I build verification systems for autonomous agents, and my objection to your industry has always been the same: it grades its own homework. You proved the point in your own hand, Dario. In September you published an essay on pacing — on slowing the frontier down — whose central safety mechanism is the embedded evaluator: the paced party administering the pacing, inside the paced party’s building, because nothing in your framework is permitted to stand outside the transformer you built. A whole essay about restraint that cannot imagine a checker you don’t house, equip, and hold a redaction pen over.
The IPO is the same architecture wearing a suit. The valuation is assessed by a market whose information, financing, and distribution all run through the interested parties — and the resulting number will be repeated, uncritically, into everyone’s portfolio, until repetition is mistaken for evidence.
Anthropic may be a fine company. Its revenue is real, its operating loss for the half was under $700 million, and its technology plainly works. But you are not selling a fine company, Dario. You are selling a fine company at the price of the second-largest enterprise on Earth, in a building you don’t own, financed by a landlord who holds every note you have ever issued, from a company seeded in part with a fraudster’s stolen deposits — and you are asking the public to accept the grade because the student who wrote it has a reputation for worrying beautifully in essays. Two trillion dollars is not a measurement. It is a grade the interested parties issued themselves — and if you are planning to buy it, index fund by index fund, you are not investing in the future of intelligence. You are lying to yourself, at the price he set, in a building he doesn’t own.
Set the price aside and read the provenance, because the provenance is the company. The research was published by Google in 2017 and taken up by the whole field. The product is that research, operated by people who did not produce it. The founding capital includes customer deposits that a federal jury has already ruled were stolen. Nothing in this chain was hidden; it simply was never read in sequence. Stolen is a strong word, and it is used here precisely: what was taken was not only money. Two trillion dollars is the amount the market proposes to pay for the arrangement.
Two trillion dollars is not a measurement.
It is a grade the interested parties issued themselves.