The amended registration statement went to the Securities and Exchange Commission on Nov. 2, and the two figures that will decide how the listing is priced sit far apart in the document. One is revenue: $17 billion for the three months to September. The other is a queue of future payments: $895.4 billion of infrastructure commitments, some of which run until 2045.
The gap between them is the story. Anthropic lost $567 million on operations in the quarter, a number small enough against $17 billion of sales that a banker can present it as a rounding error on the way to scale. The commitment figure cannot be presented that way. It is a promise to pay, signed in advance of the customers who will eventually cover it.
Both numbers describe the same business from opposite ends. Anthropic sells access to models, and the capacity to serve those models has to be bought before a single query arrives. The filing shows how the company has chosen to buy it. Short-term compute spending reached $5 billion in the quarter, a sum that buys flexibility rather than the lowest available price. Multi-year contracts cost less per unit and bind for longer. Anthropic has been paying for the right to walk away.
Then there is the part of the document that changes fastest. The company’s commitment to Google Cloud stood at $114.5 billion a week before the filing. In the amended version it is roughly $121.3 billion. A single week added about $6.8 billion of contractual obligation, which is what a negotiation looks like when one side needs capacity sooner than the other side needs the deal.
Not all of the $895.4 billion is equally firm. About $213.8 billion carries contractual conditions or termination rights, roughly a quarter of the total. That slice is the part investors will read most closely, because it is the part the company can theoretically escape. The remaining three quarters cannot be escaped, only paid, and the payment schedule stretches across two decades.
That schedule matters more than the headline figure. A data center built to serve a model in 2027 has an economic life measured in a handful of years, because the chips inside it will be several generations behind by the time the last payment on the building clears. A commitment that runs to 2045 is therefore not a bet on demand in 2045. It is a bet that demand in every year between now and then will be strong enough to cover a fixed cost set in 2026. Contracts of that length are common in infrastructure and rare in technology, and the difference is who absorbs the obsolescence.
The operating loss is the smaller puzzle. Anthropic’s revenue is real, contracted and growing; the loss reflects what it costs to produce that revenue in the current market for computing capacity. Short-term contracts carry unit costs above those of long-term agreements, and the company has said as much in its own risk language. The loss is therefore a price paid for optionality, not a sign that the product fails to sell. Whether that trade was wise depends on a question the filing cannot answer: whether the cost of capacity falls before the commitments come due, or rises.
Public investors will be asked to underwrite the whole structure with a single share price. That forces a particular kind of arithmetic. Revenue has to grow fast enough to cover a fixed cost base that was locked in before the revenue existed. Every quarter that growth disappoints, the commitment does not shrink; it simply occupies a larger share of the cash the business generates.
The document also fixes the company’s position relative to its rivals in a way that marketing cannot. A competitor that signed decade-long agreements at lower rates carries a different risk: it pays less per unit and cannot leave if demand cools. Anthropic has bought the opposite exposure. In a shortage, the flexible buyer pays more and survives a downturn better. In a glut, the flexible buyer looks clever for about one quarter, until the long-term contracts signed by everyone else reset the market price downward.
There is a third possibility, which is that the shortage persists long enough for the flexible strategy to look expensive for years. The filing’s own framing suggests the company sees that risk clearly. It does not forecast relief. It lists the gap between short and long contracts among the things that could hurt results.
What the document cannot show is the demand curve that would justify the queue. Anthropic reported $17 billion of quarterly revenue and a loss of $567 million in the same three months. An investor who believes the first number keeps compounding will treat the second as a phase. An investor who does not will treat the $895.4 billion as the number that defines the company, and the revenue as the smaller of two facts.
The filing gives both, in the order the company chose, with the loss stated plainly and the commitments disclosed in full. Everything after that is a judgment about time: how long the buildout takes, how quickly the chips inside it age, and whether the customers arrive before the invoices do.
