The $47 billion sentence

Anthropic's self-reported $47 billion annualised run-rate is not a revenue figure. It is a sentence about revenue, spoken by the company that stands to benefit from the sentence being believed. The gap between that sentence and OpenAI's disclosed $25–33 billion is the story most coverage is running with. The more interesting story is what the sentence is doing before anyone audits it.

I want to be careful here. I am not arguing Anthropic is lying. Self-reported ARR (annualised recurring revenue, extrapolated from a recent billing period) is a legitimate metric that private companies routinely disclose, and Fortune's July 3 analysis is doing the work any pre-IPO comparison can do with the material available. What I am arguing is that the $47 billion number is load-bearing for a specific narrative, Anthropic pulling ahead of OpenAI in the enterprise, and that the load it bears will change materially when both companies file S-1s and submit to third-party audit.

Start with what the number is. Anthropic told the press it is running at roughly $47 billion annualised, driven primarily by Claude Code (its AI coding tool for developers and engineering teams) and more than 1,000 enterprise clients at ACVs (annual contract values) above $1 million. OpenAI's disclosed number for the same period sits in a $25–33 billion band, itself a range rather than a point. The headline gap is somewhere between $14 billion and $22 billion. Both figures are self-reported. Neither has passed through ASC 606 revenue recognition (the US accounting standard that governs when contracted revenue can actually be booked), and neither has been reconciled against deferred revenue, committed-but-unconsumed spend, or the various adjustments an audit will surface.

The auditor's veto. The SaaS pre-IPO era gave us a small library of self-reported metrics that did not survive first contact with audited accounts. WeWork's community-adjusted EBITDA is the famous one, but the more common pattern is quieter: ARR figures that included committed contracts not yet delivered, multi-year deals annualised at the first-year rate, or usage-based revenue extrapolated from a peak month. None of these are fraud. All of them shrink when an auditor arrives.

Anthropic's book is particularly exposed to one version of this. Enterprises buying AI capacity through large committed contracts, buying tokens in bulk to lock in pricing, generate reported revenue that mixes delivered usage with pre-purchased capacity. Under ASC 606, the pre-purchased portion is typically deferred until consumption. A $1 million ACV signed in Q1, with the customer having used $200,000 of tokens by Q2, does not annualise to $1 million of recognised revenue. It annualises, roughly, to the run-rate of what has actually been consumed. Whether Anthropic's $47 billion figure treats committed contracts on a consumed basis or a billed basis is the entire question, and it is not one the current disclosures answer.

I do not know which convention Anthropic is using. Neither, I suspect, does most of the coverage repeating the number.

The concentration arithmetic. Take the enterprise book at face value: more than 1,000 clients at $1 million-plus ACV. If mean ACV is $2 million, that is $2 billion. If it is $5 million, $5 billion. To get from the enterprise book alone to $47 billion in ARR, mean ACV across those 1,000 accounts would need to sit somewhere north of $40 million — which is possible if the tail includes a handful of hyperscaler-scale deals, but implausible as a description of the median customer. The rest of the run-rate is coming from somewhere: API revenue outside the named enterprise cohort, Claude Code seat expansion, developer-tier spend, and, critically, the committed-not-yet-consumed portion of the large deals.

This matters because concentration risk is not evenly distributed across a book like this. A handful of very large accounts almost certainly carry a disproportionate share of the $47 billion. One large departure — a hyperscaler switching primary model vendor, a regulated industry pulling back after a compliance event, a competitor releasing a genuinely better coding model at a lower per-token price — materially dents the headline. This is the SaaS pattern applied to AI enterprise: the seats looked wonderful right up until the seats moved.

Claude Code is the specific version of this problem. Developer tools have shown, repeatedly, that lead in this category is temporary. The coding-copilot market has already reshuffled multiple times since 2023, with developers migrating from one incumbent to another on the basis of tab-completion latency, model quality, and IDE integration. Anthropic's Claude Code is currently ahead on several of these axes. There is no structural reason it stays ahead. If the run-rate is meaningfully driven by Claude Code, the run-rate inherits Claude Code's competitive fragility.

The revenue is real. The stickiness of the revenue is a separate question, and it is the question the number cannot answer on its own.

Karp's critique, taken seriously. On July 1, Alex Karp called per-token AI pricing "effing insane". The framing is blunt, and Karp has commercial reasons to prefer other pricing models — Palantir's whole business is a rejection of the meter. But strip the vocabulary out and the underlying claim is worth engaging. Per-token pricing has two features that make headline ARR misleading in a way flat-rate SaaS pricing does not.

The first is that enterprise procurement cycles front-load token purchases. Large customers buy blocks to secure discounts, which means reported billings can run ahead of consumption for several quarters before the pattern normalises. Under a generous ARR convention, this shows up as growth. Under a strict one, it shows up as deferred revenue on the balance sheet and much slower recognised revenue growth on the P&L.

The second is that per-token revenue is not obviously tied to value delivered. A customer paying for a million tokens of Claude output is paying for the tokens whether the output solved their problem or not. Vaudit's audit of $34 million of AI invoices at 60 enterprise customers found roughly $1.7 million in disputed charges — failed requests, retry loops, model-pricing discrepancies. That is around 5% of the sample, and it is the visible portion. The invisible portion, tokens consumed on work that did not deliver commercial value, is not something either vendor reports.

Karp's point, rendered in normal English, is that the token meter can run without the value delivery running. If a meaningful share of enterprise AI spend is being consumed on retries, failed runs, and exploratory usage that never converts to production workflows, headline ARR is measuring gross activity, not net value. Which is fine, as long as everyone knows that is what is being measured.

The strongest counter-case. I have written the sceptical read. The confident read is available and I should give it its due. Anthropic reaching $47 billion annualised, even on a generous convention, is a genuine commercial event. The company signed enough enterprise contracts, and shipped enough Claude Code seats, that it can plausibly claim that number in front of press without being immediately contradicted by its own customers. That is not nothing. OpenAI, for all its 900 million-plus weekly active users, has not made an equivalent enterprise claim in the same window.

More importantly, Anthropic said in May it expects to reach profitability in 2029, a year ahead of OpenAI's stated 2030 target. If both companies hit those targets — and both targets are self-reported, so the same scepticism applies in both directions — then the more consequential comparison is not who has the bigger top line but who reaches positive operating leverage first, on what cost curve. Profitability timing partially normalises for the accounting conventions that make ARR comparisons unstable. It is closer to a real number.

OpenAI's consumer base is the position the enterprise conversation underweights. 900 million weekly active users is not glamorous in an enterprise revenue discussion. It is, however, a moat of a specific kind: it generates data, brand habituation, and developer familiarity that a $1 million ACV contract does not replicate. The consumer flywheel is the reason ChatGPT is a verb and Claude is a product. Enterprise revenue can move faster in either direction; consumer defaults move slowly. Anthropic winning the run-rate press cycle is compatible with OpenAI holding the more durable strategic position, and I think both things are currently true.

What to watch. Three things, in order. The S-1 filings, whichever comes first — the moment ARR becomes audited revenue is the moment the $14–22 billion gap either compresses, holds, or inverts. The deferred revenue line on that S-1, which will tell us how much of the reported run-rate was committed-not-consumed. And the churn disclosures on the enterprise book, particularly on Claude Code, which will tell us whether the concentrated ACV is sticky or seat-mobile.

Until then, the $47 billion number is doing narrative work, and the narrative work is worth reading on its own terms. Anthropic wants the market to price it as the enterprise leader before the audit arrives. That is a rational thing for Anthropic to want. It is not, on its own, evidence that the enterprise leadership is real.

I would not short the sentence. I would not buy it either.

Footnotes