Anthropic's Q3 shows AI margins hinge on an 18-month forecast

Anthropic swung from a $974 million operating profit to a $567 million loss in one quarter after buying $5 billion of compute at short-term prices.
Anthropic grew revenue to $17.0 billion in the third quarter and still swung from a $974 million operating profit to a $567 million operating loss. The amended prospectus it filed on 2 November, ahead of a planned Nasdaq listing, puts the swing on two lines, and the larger one is $5.0 billion of compute bought under short-term commitments, at prices well above its multi-year contracts.
Demand did not fail. The forecast did, in the direction every lab hopes for. Anthropic contracts most of its compute 12 to 18 months before the chips arrive, and when customers outran that plan it paid short-term rates for the gap. At a frontier lab, operating margin is now set less by the price of a token than by how well the company guessed its demand a year and a half earlier. The purchasing calendar has become the main lever on valuation.
The ratio moves fast. Compute and infrastructure took 48% of revenue in the second quarter, the low point of the first public filing. In the third quarter the line reached $9.9 billion, or 58% of revenue, ten points in a single quarter, and the company expects to lean on short-term compute even more in the near term. Its answer is the commitment stack: $895.4 billion of purchase commitments and leases, with about $69.8 billion due within twelve months of 30 September and the tail running to 2045. A lab that underbuys pays the short-term premium. One that overbuys owes take-or-pay bills whatever demand does. Either way the risk sits on the balance sheet, and the suppliers on the other side, the clouds whose contracted backlogs already passed a trillion dollars, get paid in both cases.
For companies built on frontier models the read-across is concrete. Anthropic did lower prices in the quarter, but through new model versions and contracted discounts, not through cuts on the models customers already run. A supplier paying a premium for scarce compute has every reason to ration capacity at peak before it discounts. An AI-native application should budget for cheaper tokens arriving with the next model rather than with lower prices on the current one, and keep the engineering that lets it move workloads between models when one supplier tightens. The application companies Fanaura looks for keep their gross margin through a quarter in which the model supplier runs short of chips, because the third quarter showed that such quarters reach a lab at Anthropic's scale.
The strongest objection is that the loss is noise. The $2.0 billion non-cash charge for a donation-matching programme on its own exceeds the operating loss, and short-term compute is a bridge until contracted capacity lands. The company itself, though, expects that bridge to grow before it shrinks. The settling number is the compute share of revenue in Anthropic's first quarterly report as a listed company. Back under 50% means the forecast caught up with customers. Above 58% means the short-term premium has become part of the business.
Public investors will price the listing on revenue growth. The better gauge is the distance between what Anthropic contracted 18 months ago and what its customers want today, because that distance now decides who keeps the margin.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building software that cuts inference cost or moves workloads between models? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.