Cognition has raised US$1 billion at a US$25 billion pre-money valuation, US$26 billion once the new cash is counted, the company said on 27 May. Eight months ago, when it closed US$400 million, the figure was US$10.2 billion. The valuation has more than doubled in two quarters. What drove it appears one line down in the announcement. Devin, the company's autonomous software engineer, is on a US$492 million annualised revenue run-rate.
The round
Lux Capital, General Catalyst and 8VC led, with Ribbit Capital, Atreides and Layer Global joining as new backers and a long list of existing investors following on. The company has spent the past year buying its way to scale, folding in the remaining pieces of Windsurf after that startup's earlier breakup. A US$1 billion round at this valuation says the people writing the cheques think coding agents are past the demo stage.
What the revenue says
Run-rate is a forward number — last month's revenue times twelve — so treat US$492 million as a snapshot, not a banked figure. Even discounted, the trajectory is the rare part. Cognition says enterprise use of Devin has grown 50 per cent month on month for six straight months. Six straight months at that rate is the part the valuation is priced on, because it covers long enough to rule out a single quarter of trials.
The customers
Cognition names Mercedes-Benz, NASA, Goldman Sachs and Santander among its enterprise users, per Enterprise DNA. Logos like these matter for a coding agent because regulated industries are the slowest to let software write and ship code on its own. A car maker and a bank putting Devin into real workflows is a stronger signal than any benchmark score.
The valuation question
US$26 billion for a company at roughly half a billion in run-rate is a price that assumes the next few years look like the last six months. That is the bet across this whole cohort, from Cursor to Replit. The revenue is growing fast, but its durability is an open question as competition arrives from every model lab at once. The funding removes the cash constraint, but it does not settle whether enterprises will keep expanding their agent use once the novelty and free credits run out.
Eleven weeks later, the same question was asked at US$40bn
The closing argument here was that US$26bn for a company at roughly half a billion in run-rate prices in the next few years looking like the last six months. The market did not wait for the next few years. It repriced on the next quarter.
By 12 August, Cognition was reported to be in talks for a further round at a valuation of at least US$40bn, about eleven weeks after the US$26bn round closed. The basis reported for it is a target the company has stated publicly: crossing US$1bn in annualised revenue before the end of 2026, from the US$492m figure in this piece.
The caution above about run-rate applies with more force, not less, to the newer number. A US$1bn target is a forward projection of a forward metric. The important distinction will be whether it is reported as achieved or merely approached.
The free-credits question has a partial answer
The original piece questioned whether agent use would continue expanding after the novelty and the free credits wore off. Three months of data show expansion continuing, but leave the underlying question open.
Cognition reports enterprise usage of Devin growing more than tenfold since January 2026. That is a different claim from the month-on-month growth quoted above, and a more useful one, because it covers a long enough window to survive a single quarter of trials. Parts of the United States government have since been named alongside Mercedes-Benz, NASA, Goldman Sachs and Santander.
What that does not establish is margin. Expansion measured in usage says customers are running more work through the agent. It says nothing about what that work costs to serve, and the whole cohort is buying inference from suppliers who also sell competing products.
The cohort did not stay a cohort
The bet described here was said to run across the whole field, from Cursor to Replit. Within three months that field had resolved into three different outcomes rather than one.
Cursor reached roughly US$2bn in annual recurring revenue by February and was reported to be raising at around US$50bn. It never tested durability as an independent company: in August, SpaceX closed a US$60bn all-stock acquisition of its maker Anysphere, the largest startup acquisition on record. Anthropic's Claude Code passed US$2.5bn in run-rate revenue as a product of the lab that makes the model it runs on. Cognition is the one still scaling as an independent company that buys its models from other people.
These are three structurally different businesses, so the valuation logic for one does not transfer to the others. An acquisition price paid in a rocket company's stock is not a market comparable for an independent raise.
What has hardened since
Competition arriving from every model lab at once was named as the open risk here, and it has become the defining condition rather than a risk.
Warp shipped Factories in August, running coding agents as a staged pipeline with a different model allowed at each stage, and sold model neutrality as the feature. That is a direct answer to a market where the model suppliers are also the competitors. It also relocates the constraint to review capacity, and the evidence on human review is poor in both directions: a study of 409,000 approval decisions found reviewers missed 33.7 per cent of malicious commands while blocking safe ones at rates up to 59 per cent.
For an autonomous software engineer sold on working unattended, the reliability of human review is the real number under the valuation. Enterprises expanding usage tenfold are expanding their exposure to a control that does not measure well, and no funding round settles that.
View author profile → · Editorial policy
