SAN FRANCISCO, 15 AUG 2026 — Anthropic told prospective investors this week that it took more than US$11.5 billion in revenue last quarter, against US$787 million in the same quarter of 2025. That is a fourteen-fold increase in a year, and it is the number driving the story.
It also reported, for the first time, positive adjusted operating income. The two halves of that sentence are not equally precise. Revenue arrived as a figure; profit arrived as an adjective.
What was reported
Revenue roughly doubled quarter on quarter and rose fourteen-fold year on year. Both are extraordinary and neither is in dispute.
The number everyone is quoting is the old one
Coverage this week has widely attached a specific profit figure to the quarter: US$559 million, on revenue of US$10.9 billion, a margin of about 5.1 per cent.
Those are not the results. They are the projection Anthropic gave investors in May, reported at the time by the Wall Street Journal. The company said then what it expected the quarter to look like; this week it said what the quarter actually was.
| May projection | August preliminary | |
|---|---|---|
| Revenue | US$10.9bn | More than US$11.5bn |
| Adjusted operating income | US$559m (≈5.1% margin) | Positive — no figure |
| Status | Told to investors | Preliminary, unaudited |
Revenue beat the projection by more than half a billion dollars, which is the good news and is being under-reported. The profit line went the other way: from a specific number in a forecast to a directional statement in a result.
There are innocent explanations. Preliminary figures are often given directionally before the books close, and a company with no public reporting obligation has no duty to publish a profit-and-loss statement at all. The asymmetry within a single disclosure is telling. If revenue can be stated to the nearest hundred million while it is still preliminary, the constraint on the profit line is a choice about what to say, not an accounting limitation.
What "adjusted" is doing
Adjusted operating income is not a defined term. Each company decides what it adjusts for, and the adjustments are where the interesting information sits.
For a frontier AI laboratory, the list of likely candidates is large and specific. It would include share-based compensation (substantial at a company competing for researchers), the amortisation of prepaid compute, one-off funding round costs, and any restructuring. None of these is illegitimate to exclude, and all of them are real money leaving the business.
Anthropic has not published what it adjusted for. Until it does, "positive adjusted operating income" establishes that one particular constructed figure crossed zero, and not much else. This is not a claim of profitability. The difference is not pedantry; it is the entire point.
The timing question
The most substantive challenge to the profitability story came from the writer Ed Zitron in May, and it is about when the costs fall rather than how they are labelled.
His argument is that Anthropic secured a large compute arrangement whose fees were reduced during a ramp-up period, and that the reduction lands inside precisely the quarter being described as profitable. If a discount is temporary, a margin that depends on it is temporary too, and the following quarters would show costs resuming their previous relationship to revenue.
We could not verify the counterparty or the contract terms, and we are not asserting them. The mechanism is worth considering because it is testable. It predicts that the third quarter will look materially worse than the second at similar revenue. That is a falsifiable claim, which is more than most commentary on this company offers.
The supporting figure that is on the record is the cost of compute, reported as falling from 71 cents per dollar of revenue in the first quarter to 56 cents in the second. That is a very large move in three months. Efficiency gains, better utilisation and cheaper inference all plausibly contribute. So would a discount.
The valuation is a separate argument
Anthropic is reported to be closing a US$30 billion round at a US$900 billion valuation, with backers discussing a figure closer to US$2 trillion ahead of an expected listing.
Take the quarter at face value and annualise it: something above US$46 billion of revenue. At US$900 billion that is roughly twenty times revenue, which is high but not unprecedented for a business growing fourteen-fold. At US$2 trillion it is above forty times revenue on a business whose profitability is, by its own description, adjusted and preliminary.
The gap between those two numbers is not a rounding error. It is the difference between pricing a fast-growing company and pricing an outcome that has not happened yet.
Why this matters from here
The reason to read the profit line carefully in this region is pricing.
Anthropic withdrew a scheduled price rise on Sonnet 5 last week, making introductory pricing permanent rather than lifting it on 1 September. A company that reports a profitable quarter and cancels a price rise in the same fortnight is signalling which constraint it is managing. It is not margin.
Set that beside DeepSeek moving in the opposite direction, and the split across the market is now clear. The Western laboratories are buying share and can afford to, because capital is abundant and a listing rewards growth. The Chinese tier is passing through a compute squeeze that reaches down to the foundry.
For a team here choosing a model, that means the price of Western frontier capacity is currently a strategic decision by a company raising money, not a reflection of what it costs to serve. Strategic prices move when the strategy does. Build on the assumption that today's rate card is a floor rather than a ceiling.
What we could not establish
The actual profit figure, the adjustments behind it, the accounting basis, and whether any of it has been audited. The company is not subject to public reporting requirements and has published no statements.
Also unestablished: the identity and terms of the compute arrangement at the centre of the timing critique, whether the discount characterisation is accurate, what drove the fall in compute cost per revenue dollar, the split between consumer and enterprise revenue, and whether the funding round has closed. Figures described here as preliminary were shared with prospective investors rather than published.
What to watch
The third quarter, and specifically whether compute cost per revenue dollar keeps falling. That single ratio will settle the timing argument more decisively than any amount of commentary.
Whether a profit figure ever appears with its adjustments listed. An IPO process would eventually force that, which is the strongest reason to expect it.
And whether the price of Claude moves after the round closes. That is the tell that costs you money, showing up on an invoice long before it appears in a filing.
View author profile → · Editorial policy
