A RAAK.work project
Capex WatchAnthropic rents its computing from the builders and sells AI to businesses and the public, so its revenue is the nearest reading there is of what final customers pay. On 14 October it told prospective investors that it took in $16.8B in the third quarter, up about 45% on the second. That is fast, but slower than before, and some investors lowered what they would pay for it. None of these figures is filed; all are press-reported.
| Line | Expected | Delivered |
|---|---|---|
| Revenue growth on the quarter | the second quarter more than doubled the first | about 45%, to $16.8B (from $11.5B on the FT’s figure, $11.6B on Reuters’) |
| Annualised revenue | far more than $100B by the end of the year (some investors, in the summer) | no new figure from the company; $65B at the end of July; about $74B now on SemiAnalysis’s estimate |
| Profit | — | profitable on an adjusted basis for the second quarter in a row; operating margin 11% before stock-based pay |
| Cash | — | cash-flow positive |
| Revenue per dollar of computing | — | $1.70 |
| Revenue from rival labs training on its models | — | 5–7% of revenue, cut back in the quarter |
| Value at the listing | about $2 trillion | about $1.5 trillion, discussed by some investors who attended |
Press-reported, not filed. Every figure in this table is what Anthropic told prospective investors in San Francisco on 14 October, as reported by the Financial Times and Reuters from people who attended. Anthropic is a private company and has published no results. The two reports differ slightly on the earlier quarters: the FT gives $11.5B for the second and about $4.8B for the first, Reuters $11.6B and $4.7B. The three clocks above ask three questions: is AI revenue growing fast enough (demand), is the spending still covered (capex), and are investors still willing to wait (patience). Anthropic is not one of the five big AI builders the Bank for International Settlements models (Alphabet, Amazon, Meta, Microsoft and Oracle). It is one of their customers. Nothing on this page feeds the tracking page’s required pace of growth, its spending split or its gap chart.
The front page works out what the five builders must receive each year for their spending to pay off: the floor, $1.3T a year at its defaults. Most of that must come from firms like Anthropic, which rent the computing and sell AI products on. So Anthropic sits on both sides of the bet. What it spends on computing is the builders’ revenue. What it earns is part of what final customers pay for AI.
As a customer of the builders, it is still small against the floor. $16.8B of revenue at $1.70 per dollar of computing means it spent about $9.9B on computing in the quarter, or about $40B a year at that pace. That is about 3% of the floor, and not all of it reaches the five: some goes to builders outside them, such as SpaceX. Its commitments run much further: $518B of computing over the coming decade, of which $252.5B, less than half, is with Google, Amazon and Microsoft. From Anthropic’s draft prospectus, as reported by Reuters on 28 Sept 2026. Not an official filing.
As a seller to final customers, it gives the first measured view of a number the site has had to assume. The buyers’ bill works out what final customers must pay for the builders to receive the floor, once the firms in between add their other costs and profit. Its defaults assume such a firm spends 67.5 cents of each dollar of revenue on computing (computing 75% of its costs, a 10% margin), which gives a bill of $1.82T a year. Anthropic’s third quarter says 59 cents (1 ÷ 1.70). If every firm in the middle looked like that, the bill would be about $2.06T a year. That fits what the bill page says of its defaults: they are cautious, and the real bill is more likely higher than lower.
These are figures Anthropic chose to tell investors before a share sale, passed on by people who were in the room. They are not audited, and the reports leave open what they measure. They do not say whether revenue from Claude sold through cloud providers is counted in full or net of the provider’s share. They do not say what the adjusted profit leaves out beyond stock-based pay, which is a real cost to shareholders. They do not say whether “computing power” includes the cost of training new models or only of serving customers.
Nor does one quarter show a trend in profit. In 2025 Anthropic lost $8.06B on its operations on $4.59B of revenue, about $1.76 for every dollar it took in. From Anthropic’s draft prospectus, as reported by Reuters on 28 Sept 2026. Not an official filing. That loss and the third quarter’s 11% margin are not the same measure, so the two cannot be set side by side as a turnaround. What they show is how far the business has moved in a year, and how much depends on how the profit is counted.
And the reasons for the slowdown point two ways. The FT’s sources name three: a price war with OpenAI and cheaper Chinese models, less revenue from rival labs training their models on Claude’s answers, and two big customers, Meta and Microsoft, using more of their own AI. The first two are pressure on price, which is what the bet needs to survive: cheaper AI that more customers buy. The third is different. Meta and Microsoft are two of the five builders. When they replace a lab’s AI with their own, revenue moves from the lab to the builder’s own budget, and no new customer has paid for anything.
The public prospectus. Anthropic must file it with the US Securities and Exchange Commission before it lists, which Reuters expects after the US midterm elections in November. It will carry audited figures for these same quarters, and it will settle the questions above: gross or net revenue, what the adjusted profit leaves out, and what computing costs include. Its figures will replace the reported ones on this page, on the tracking page’s row and on the bill page.