For most of the past year the question about Anthropic was how fast it could grow. Monday’s filing asked a different one: how much it owes.
The maker of the Claude models publicly filed its registration statement with the Securities and Exchange Commission, setting up a Nasdaq listing under the ticker ANT. Goldman Sachs and Morgan Stanley are leading the offering. The company has not fixed a share count or a price range, so the document reads less like a sale than a disclosure of arithmetic.
The arithmetic runs in two directions. Revenue for the six months through June was $16.3 billion, against $1.2 billion a year earlier. Operating losses narrowed to $693 million over the same period from $2.1 billion. The net loss was $80.7 billion, a number larger than the revenue line and one the company attributes mainly to non-cash accounting adjustments tied to its convertible notes.
That last figure will draw the first questions on the roadshow. A loss of that size does not represent cash leaving the business, but it does reflect what the capital structure behind the company’s growth costs to carry. The filing carries the standard warning that the company may not achieve or sustain profitability, a sentence that appears in nearly every technology prospectus and reads differently when the operating loss has already narrowed to a few hundred million dollars.
Control is settled in advance. Seven co-founders hold 50.1% of voting power through an entity called Founder LLC, a structure that keeps the founding group in charge of decisions public shareholders cannot outvote. Buyers of the stock would get economic exposure to the business and little say in how it is run, a trade investors have accepted at other founder-led technology companies and complained about later.
The table that drew more attention than the loss statement was the one listing supply contracts. Anthropic has committed to pay SpaceX’s artificial-intelligence arm, SpaceXAI, roughly $3.1 billion a month under an agreement whose cap runs to $120.5 billion through 2029. It has promised Microsoft at least $44.9 billion between November 2026 and May 2033, on terms the filing describes as non-cancelable. Commitments to Google and Amazon stand at $114.5 billion and $110 billion. Added together, the caps come to about $807 billion.
Set that total against first-half revenue and the ratio is roughly fifty to one. Set it against the operating loss and the comparison stops being useful at all.
What the company has signed are capacity agreements, not debt. They oblige Anthropic to buy computing over years at prices negotiated in advance, and they appear in a prospectus as purchase obligations rather than borrowings. That distinction matters to a credit analyst and very little to a treasurer. A purchase obligation consumes cash in the quarter it comes due, whether or not the models trained on that capacity earn anything. The practice became standard as model developers stopped renting capacity by the hour and began reserving multiyear blocks of data-centre space, power and silicon, on the reasoning that a training run which starts late is worth less than one that starts on schedule.
The counterparties sit on both sides of the ledger. Microsoft, Google and Amazon are investors in Anthropic, hosts of its workloads and, through these contracts, its largest suppliers. Each has an interest in the company’s success and a claim on its spending. Regulators in Washington and Brussels have spent the past year asking about arrangements of this shape across the industry.
For the public market, the listing would be the first direct claim on a frontier model developer of this size. Until now, investors who wanted exposure to the largest private AI companies could buy shares in their backers or their suppliers and hope the correlation held. A listed Anthropic converts that proxy into a price, updated every second, and gives the private rounds at its rivals something to be measured against.
It also hands the company a currency. Employees hold equity that has been valued by venture rounds and secondary sales; a public price settles what it is worth and gives recruiters a number to quote. Bankers involved describe the timing as dependent on how quickly the commission finishes its review, not on demand, which is the usual phrasing and, in a year when several technology offerings traded below their issue price within weeks, a defensible one.
The disclosure does something else that is easy to miss. It makes the cost of staying in the model race legible to people who do not work in the industry. A reader can now see revenue, an operating loss, a net loss and a stack of multiyear purchase commitments on consecutive pages, all audited to the same standard, and decide whether the business looks like a software company with unusual expenses or an industrial project with a software margin.
That question will be answered by the price, and the price will be set in a market that has spent the year arguing about it. The filing’s contribution is narrower and more useful: it puts the numbers in one place, in public, where the argument can be had with arithmetic instead of adjectives.
