Product1 publisher3 min readPublished
Investors repriced Cognition at $48bn without touching the revenue multiple they paid in May
The Series E takes company-reported run-rate from $492M to close to $900M in four months, but nothing in the announcement says how much of that is Devin agent work and how much is Windsurf editor seats.
The Product Desk · Product desk
What happened
- Cognition said it has raised more than $2 billion in late-stage funding at a $48 billion valuation, nearly double the price investors put on the company in its previous round.
- The company puts run-rate revenue close to $900 million, up from the $492 million it reported in May when it raised more than $1 billion at a $26 billion valuation.
- Devin is named in chip design at Nvidia, aviation work at GE Aerospace, financial services at Citigroup, automotive engineering at Mercedes-Benz and AI infrastructure at Modal Labs, with Goldman Sachs also a customer.
- New platform pieces include Auto-Triage for production incidents, Security Swarm for hunting vulnerabilities across a codebase, and Automations that start agent work from events in Slack, GitHub and Linear.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint With only a blended run-rate published, a procurement team cannot price the renewal risk it is actually taking on: editor seats churn one developer at a time, an agent with commit rights churns after a single bad incident.
- precedent Because the multiple held rather than expanded, the next round needs another near-doubling of revenue to clear the same bar, and that pressure lands on the enterprise renewals being signed this year.
- exposure One of the flagship reference deployments belongs to an investor in the round, which lowers what Nvidia's Devin usage can prove to a buyer treating it as independent validation.
- decision Any team that switches on event-triggered agent work is deciding, usually before it is anyone's named job, who reviews the change and who owns the revert when the fix lands unattended.
Somewhere inside one of those accounts, a staff engineer is looking at a pull request an agent opened against a service she owns, deciding whether to read every line or trust the test run. That review queue is where a run-rate gets renewed, and it is invisible in a funding announcement.
The arithmetic of the two rounds is worth setting side by side. In May, more than $1B came in at a $26B valuation against $492M of run-rate [2][3], about 53 times revenue [3]. This round prices $48B against a figure the company itself puts close to $900M [1][4], which is also about 53 times [4]. The price per dollar of revenue did not move [5]; what moved is roughly $408M of run-rate in about four months [1][7], near $102M a month [6], or 83% growth [2]. Andreessen Horowitz and Accel led as new investors, with more than 30 other firms in the round [5]. They bought more revenue at last round's price, which is a bet that this revenue is the same quality as the revenue priced in May.
Whether it is depends on a split the announcement does not give. Cognition sells Devin, which takes high-level requirements and then plans, writes and tests code, debugs and pushes deployments inside a sandbox [8], and it owns Windsurf, the editor it bought in July 2025 after OpenAI's takeover bid collapsed and Google hired Windsurf's founders in a technology licensing deal [9]. An editor seat and a standing agent with commit rights are different purchases with different renewal logic: one churns when a developer leaves, the other churns after one bad incident. No product-level revenue split, retention rate, seat count or contract value appears anywhere in the disclosure [8].
The company's own framing is that engineers "should operate more like architects and delegate execution to swarms of agents" [13], and that it can "choose and combine the models best suited to the work, including our own" [12]. What is checkable is thinner and still substantial: six new offices, in Washington, Tokyo, Singapore, London, Sao Paulo and Madrid, added to hubs in San Francisco, New York and Austin [11]. That footprint is what a vendor builds to sell into and support long procurement cycles. It carries no information about how deep the work goes once inside an account, and no depth measure has been published.
Devin's March 2024 debut was followed by a widely circulated Internet of Bugs video arguing the demo overstated what the agent had actually accomplished on an Upwork task, which fed developer skepticism about the company's benchmark claims [14]. In August 2025, three weeks after the Windsurf deal, Cognition laid off 30 staff and offered the roughly 200 who remained either nine months of salary to leave or six-day weeks of 80 hours or more, with chief executive Scott Wu writing "We don't believe in work-life balance", according to TechCrunch [15][16]. A buyer signing a multi-year support commitment is buying that staffing model along with the product.
The measurement that separates the pitch from the work fits on one repository in one month. It comes down to the share of merged pull requests an agent opened, the share of those reverted or hotfixed within a week, and median ticket-to-merge time on that same repository before agents were switched on. Out of those fall the two axes worth more than any benchmark claim, cheap revertibility and clean attribution. Work that has both is where a swarm earns its keep now, while work with neither leaves a 53x multiple, rather than evidence, doing the arguing.
What to watch
- A revenue split between Devin and Windsurf, or any net retention figure, in the next raise or an audited disclosure.
- Whether a named account publishes deployment depth: merged agent pull requests, revert rates, or the scope of repositories an agent can write to.
- Headcount growth behind the new international offices, which would show whether the August 2025 staffing terms survived the expansion.