Anthropic's compute margins worsened in the third quarter as demand outran supply

Anthropic filed an update to its S-1 registration statement on Monday. It adds a quarter of financial results to the IPO filing the company released last week. The original filing ran through June 30. The update runs through September 30, and it shows one number moving the wrong way: the share of revenue that goes to compute.

According to TechCrunch's reading of the filing, revenue reached $17 billion in the third quarter, up from $11.5 billion in the second. Compute took 58% of that revenue, up from 48% the quarter before. The operating loss also grew, largely because of compute costs. A company preparing to sell shares to the public would normally want to show margins getting better. These got worse.

What the filing shows

The compute bill for the quarter was $10 billion. The filing includes a chart of revenue earned per dollar spent on compute. TechCrunch reports that this ratio rose steadily from the start of 2025 and peaked at just over 2x in the second quarter of 2026. In the third quarter it fell to 1.72x.

The filing gives a direct reason. Of the $10 billion, it says, "$5.0 billion of compute under short-term commitments, which currently costs significantly more than comparable compute procured under our multi-year commitments." Half the quarter's compute, then, was bought at short notice and at higher prices. Demand grew faster than the capacity Anthropic had locked in under long-term contracts, and the company paid more to close the gap.

The company does not expect the bill to shrink. The filing states: "We expect compute and infrastructure expenses to increase in absolute dollars and remain our largest operating expense for the foreseeable future as additional committed compute becomes available to us under multi-year arrangements with CSPs, and as we continue to invest in additional compute."

Where the extra compute comes from

Some of the short-term capacity has come from deals already made public. TechCrunch reports that Anthropic has been paying SpaceX $1.25 billion a month for compute. It also has multi-billion dollar agreements with Nscale, Akamai and Volta. In industry terms these providers are neoclouds rather than hyperscalers. They are smaller operators that offer more flexible terms at higher rates.

Anthropic is also trying to build its own compute base. The filing discusses a project it calls "Anthropic Direct". TechCrunch's assessment is that with demand growing this fast, that effort cannot keep pace for now, and the neoclouds will remain the main option for marginal capacity.

Why the margin moved the wrong way

The usual picture of AI economics says that costs per unit of work fall as models get more efficient and hardware improves. Over most of the past two years, Anthropic's chart supported that picture. The third quarter shows a second force that can override it: the price of capacity bought in a hurry.

The arithmetic is simple. Revenue grew by about 48% in one quarter. Long-term compute contracts are negotiated well in advance and cannot be expanded on that timescale. Every unit of demand above the contracted base has to be served from the spot or short-term market, where prices are higher. Faster growth in sales can therefore produce worse margins, at least until the long-term capacity arrives.

TechCrunch describes the shift as more likely a blip than a lasting problem. If demand slows, the multi-year contracts will cover a larger share of usage and the ratio should improve. If demand keeps rising at this pace, Anthropic will keep buying at the margin. The publication also points to risks outside the company's control. A slowdown in data centre construction, whether from nervous bond markets or local opposition to new sites, would tighten supply further and raise the price of the capacity Anthropic depends on.

The filing makes one structural point clear. Delivering frontier AI at this scale requires capital commitments measured in billions of dollars per month, and access to that capacity is concentrated among a small number of hyperscalers, neoclouds and well-financed buyers. A lab's gross economics depend less on its models than on how well it forecast demand a year or more in advance.

What it means for work

Much of Anthropic's revenue comes from businesses and developers paying to have models carry out tasks: writing and reviewing code, drafting documents, running agents. The filing shows that each of those tasks currently costs Anthropic more to serve when demand spikes, because the extra work runs on rented capacity at premium rates. For customers who build workflows on these models, the cost and availability of the service are tied to a physical supply chain of chips, power and buildings that can tighten without warning. Organisations moving substantial work onto AI systems may want to treat price and capacity as variables rather than givens. What remains open is whether long-term contracts and Anthropic Direct will bring the cost of serving a task back down, or whether demand will keep growing faster than the supply built to meet it.