Anthropic is growing fast — and paying more than ever to keep up. In an updated IPO filing released on Monday, the Claude maker reported $17 billion in revenue for the third quarter, up from $11.5 billion in the previous quarter. But compute — the chips and data centres needed to train and run its models — took 58% of that revenue, up from 48% the quarter before, and pushed its operating result back into loss.
What the filing shows
The figures come from an amendment to Anthropic's S-1, the registration statement filed ahead of a planned initial public offering. The original filing covered results through 30 June; the update, reported by TechCrunch on 3 November, extends them through 30 September:
- Q3 revenue: $17 billion — up from $11.5 billion in Q2, and, according to The Hindu BusinessLine, up 1,127% from the same quarter a year earlier.
- Compute bill: about $10 billion for the quarter — TechCrunch's rounded figure; BusinessLine puts compute and infrastructure expenses at $9.9 billion.
- Compute share: 58% of revenue, up from 48% in Q2.
- $5.0 billion of that compute was bought under short-term commitments, which the filing says "currently costs significantly more than comparable compute procured under our multi-year commitments."
- Revenue per dollar of compute fell to 1.72x, after peaking at just over 2x, according to a chart in the filing cited by TechCrunch.
- Operating loss: $567 million in Q3, compared with operating income of $974 million in Q2, as reported by The Hindu BusinessLine.
Why the bill jumped
The filing's explanation is demand. Anthropic says it had to scramble for extra capacity to keep up, and short-notice capacity costs more. TechCrunch notes deals already visible in public, including paying SpaceX $1.25 billion a month for compute, alongside multibillion-dollar deals with Nscale, Akamai and Volta. In industry terms these are "neoclouds" — smaller providers that offer more flexible deals at higher rates than the big hyperscalers.
BusinessLine adds a second factor behind the operating loss: about $2 billion in non-cash charges tied to the company's employee donation-matching programme. It also reports aggregate compute and infrastructure purchase and lease commitments of around $895.4 billion disclosed in the filing.
Anthropic is trying to build its own compute base under a project it calls "Anthropic Direct," TechCrunch reports. But the filing itself expects the pressure to continue: "We expect compute and infrastructure expenses to increase in absolute dollars and remain our largest operating expense for the foreseeable future."
The honest caveats
These are Anthropic's own figures, published in its own IPO filing — a document written, in part, to present the company to future investors. The $10 billion compute figure is a rounded number in coverage; the underlying compute and infrastructure expense reported elsewhere from the same filing is about $9.9 billion. The headline quarter also mixes cash costs with non-cash charges, so the operating loss is not the same as cash burned. And one expensive quarter does not prove a permanent trend: TechCrunch notes the shift could be a blip caused by buying emergency capacity, and margins could look very different if demand slows or cheaper multi-year capacity comes online.
Why this matters to you
This is the clearest public look yet at the real economics behind a leading AI chatbot. For every dollar Anthropic earned in Q3, about 58 cents went straight back out on compute — before other costs. That helps explain why AI subscriptions, API prices and free limits are under constant pressure across the industry, and why the race to build data centres matters so much: the labs do not just compete on smarter models, they compete on who can buy computing power cheapest. If compute stays expensive, those costs shape what users pay — and what free tiers can offer.
Sources: TechCrunch, 3 November 2026; The Hindu BusinessLine, 3 November 2026 — both reporting Anthropic's amended S-1 filing covering the quarter ended 30 September 2026. Hero image: server room at CERN, by Florian Hirzinger, via Wikimedia Commons.
