Anthropic crossed $47 billion run-rate revenue in five months and growth is accelerating

Simon Willison pulled the most interesting number from Anthropic's Series H announcement: the company says run-rate revenue crossed $47 billion. The trajectory is as striking as the figure itself: $9 billion in December 2025, $14 billion in February 2026, $30 billion in April, $47 billion in May. If that holds, Claude is no longer a tool for teams with experiment budgets. It is becoming a consumption layer for enterprise software.

Run-rate from $9 to $47 billion in five months: a fundraising number, not an audit

Anthropic stated in its Series H materials that since the February round, adoption across global enterprise customers kept growing and run-rate revenue crossed $47 billion. Willison correctly notes this is run-rate revenue: a projection from current pace, not audited full-year revenue.

That does not make the number worthless. Willison argues that quoting a materially false figure in documents for investors who just put in $65 billion would be securities fraud. The number is at least credible enough to anchor future verification.

For the market, this changes the metric: not how many firms try AI, but how many pay for it daily

AI models have mostly been evaluated by benchmarks and demo capabilities. This number points to a different metric: how much companies actually spend when AI is inside their processes. Enterprise customers are no longer paying to try AI. They are paying for daily work through APIs, agents and internal automation.

For the market this suggests model consumption can scale faster than classic SaaS. Not because one license is sold per seat, but because work is metered through APIs. Willison also mentions a case where a company apparently spent a large amount after failing to set usage limits, which is exactly the dynamic driving fast run-rate figures.

Run-rate is a sharp number but a fundraising one: without margin and audit it tells only half the story

Run-rate revenue can exaggerate growth when a customer temporarily spikes usage or when usage patterns are still unstable. The second missing piece is margin: with model companies, revenue alone is not enough. You need inference cost, model development cost and chip capacity.

Willison himself notes that the most convincing test will be an IPO prospectus or harder financial disclosure where the number survives an audit.

The key signal will be whether enterprise customers introduce limits the way they manage an electric bill

Watch for IPO documents or harder financial disclosures and whether Anthropic starts reporting retention, gross margin and customer concentration. If the number survives stricter disclosure, this is one of the fastest revenue growth stories in software.

The second signal is enterprise behavior. If customers start enforcing caps, internal routing and model budgets, AI has stopped being an experimental line item. It has become the electric bill for the automated office.

Lilith's verdict

A $47 billion run-rate is the ledger where enterprise customers see for the first time what automated work costs when nobody sets limits. Somewhere in those numbers there is probably one badly configured usage policy.