SAN FRANCISCO, 22 AUG 2026 — Anthropic told investors that its annualised revenue run rate reached US$65bn at the end of July, according to reporting on 17 August. The company had passed roughly US$9bn at the end of 2025 and about US$47bn in May.
It confidentially filed a prospectus with the Securities and Exchange Commission in June and is reported to be preparing a listing as soon as October, at a valuation of two trillion dollars or more. The run rate is the number carrying that arithmetic, and it is the one number in the set that a prospectus cannot report.
The numbers on the table
Investors are reported to expect the company to end 2026 somewhere between US$100bn and US$120bn annualised.
What a run rate is, and what it is not
An annualised run rate takes recent revenue — typically a single month or a single quarter — and multiplies it out to a year. It is a projection dressed as a measurement. It assumes the most recent period repeats twelve times, which is a reasonable planning assumption for a subscription business with low churn and an unreasonable one during a period of steep growth in either direction.
Recognised revenue, on the other hand, is what the company actually billed and earned in a period. It is the figure auditors examine and that a registration statement must report; securities regulators want historical financials, not projections from a hot month.
This is not a suggestion that anyone is being misled. Run rate is standard vocabulary in private software markets, investors asked for it, and Anthropic supplied it. The point is narrower. The figure anchoring a two-trillion-dollar valuation is a different kind of number from what will appear in the offering documents, and the two won’t match.
The arithmetic between US$46bn and US$65bn
The two disclosures are worth reconciling, because the gap between them is the growth story and it is steeper than either number alone conveys.
Preliminary second-quarter revenue of more than US$11.5bn averages roughly US$3.8bn a month across April, May and June. A US$65bn annualised run rate at the end of July implies July revenue of about US$5.4bn. That is roughly forty per cent above the second-quarter monthly average, reached in a matter of weeks.
The figure is internally consistent with the May disclosure. A run rate of US$47bn in May implies about US$3.9bn that month, which sits neatly at the top of the second-quarter range, and the climb from there to US$5.4bn in July is the same slope carried forward. Nothing in the sequence looks stretched.
The practical result is that the annual figures in the prospectus will look small next to the run rate now in circulation. A company that grew from US$9bn to US$65bn annualised over roughly seven months will recognise, across the full year, something far below the rate it exits the year at. That is arithmetic rather than weakness, and it is the standard confusion around every fast-growing company that lists.
What two trillion implies as a multiple
The reported valuation target is worth converting into a multiple, because that is the form in which it can be argued about.
Two trillion dollars against a US$65bn annualised run rate is roughly thirty-one times revenue. Measured against revenue the company will actually recognise for 2026, which will be well below its exit rate, the multiple is materially higher again. Measured against the year-end expectation of US$100bn to US$120bn annualised, it falls to somewhere between sixteen and twenty times.
Those three numbers describe the same company and differ by a factor of two or more, which is the practical reason the choice of denominator matters. An investor quoting the lowest of them is pricing off a run rate that does not exist yet, projected from a run rate that is itself a projection.
None of that makes the valuation wrong. Software businesses growing at this rate have historically supported multiples that look indefensible against trailing revenue and reasonable against the following year. It does mean the argument is entirely about the growth continuing, and the filing is where that argument acquires evidence.
The profit line is still an adjective
We noted when the second-quarter figures appeared that revenue received a number and profitability received a direction: the company reported its first quarter of positive adjusted operating income without attaching a result to it, while the widely quoted US$559m figure was a projection made in May rather than an outcome.
Nothing in the July disclosure changes that. A run rate is a revenue metric and carries no information about margin. The cost of serving inference at this volume, the compute commitments underwriting it and the depreciation schedule attached to those commitments are the terms that decide whether US$65bn annualised is a profitable business or an expensive one, and none of them are addressed by annualising a month.
A registration statement will have to address them. That is the substantive reason to wait for the filing rather than to price off the run rate.
The comparison with OpenAI is doing more work than it should
OpenAI's run rate at the end of July is reported at about US$40bn, roughly sixty per cent of Anthropic's. The comparison is being used to establish a ranking, and it is worth handling carefully.
Both figures are self-reported, neither is audited, and the two companies do not necessarily annualise on the same basis or recognise revenue across the same mix. One weighted towards enterprise contracts and API consumption behaves differently from one weighted towards consumer subscriptions, both in how revenue is recognised and in how durable a single strong month is.
The ratio may well be directionally right. It is not a like-for-like measurement, and treating it as one is how a comparison between two unaudited projections becomes a fact about market share.
What remains unconfirmed
Anthropic has not published the figures itself, and the run rate is attributed to what the company told investors rather than to a filing. The basis of the annualisation — whether it multiplies a single month, a quarter or a contracted book — is not stated. No gross margin, operating expense or cash-burn figure accompanies the revenue number.
The October listing date, the two-trillion-dollar valuation target and the year-end expectation of US$100bn to US$120bn are all reported expectations rather than company guidance. A confidential submission is not a public filing and can be withdrawn, amended or delayed without any announcement, and no terms, share count or offering size exists publicly.
What to watch for
The first thing that matters is the public filing itself, because it converts every number above into audited history. Until the registration statement is public, there is no external check on any of this.
The second is whether the prospectus discloses revenue concentration. A business that reached this scale in under a year is likely to have significant revenue from a small number of very large customers and compute partners, and the concentration disclosure will say more about durability than the growth rate does.
The third is the compute commitment schedule. Multi-year capacity contracts appear as obligations, and the relationship between those obligations and the revenue they support is the single most informative table in a filing of this kind.
