On May 27, 2026, Cognition closed a $1 billion Series D at a $26 billion post-money valuation, co-led by Lux Capital, General Catalyst, and 8VC. Peter Thiel's Founders Fund — which led the first round in March 2024 and every round since — wrote into the new one. Ribbit Capital, Atreides Management, Layer Global, Elad Gil, Soma Capital, and Omri Casspi rounded out the list. The valuation is 2.5x what the company was worth eight months ago and 53x its current annualized revenue, per TechCrunch's report and The Next Web's writeup.

The number behind the multiple is the part worth pausing on. Cognition's annualized run rate was about $37 million in May 2025. It is $492 million now. That is a thirteenfold jump in twelve months, on a product whose initial March 2024 launch demo was accused of being staged. Whatever else is true, somebody is buying Devin.

The pitch is different now

What got priced this week is not the pitch Cognition was selling two years ago. In March 2024, Devin was "the world's first AI software engineer" — a single autonomous agent that would replace the junior developer. The demo went viral, then a series of debunkings followed, then the trade press largely moved on while Anthropic shipped Claude Code and OpenAI shipped Codex and Google shipped Jules. By the time those three were in general availability, the "one autonomous engineer" framing had a problem: the autonomous engineer Cognition actually shipped was, in production, riding on the same frontier-lab APIs as everyone else's.

The Series D pitch deck — relayed in CEO Scott Wu's quotes to outlets covering the round — solves that problem by no longer being the same pitch. Cognition is now, in Wu's framing relayed by The Next Web, "an orchestration layer rather than a model company" — a router that picks the right model from the right provider for the task in front of it. It uses multiple providers as a feature, not an embarrassment.

That is a different company. A model company sells the engine. An orchestration company sells the dispatcher. The dispatcher is worth $26 billion if, and only if, the labs that make the engines are unable or unwilling to ship their own dispatchers. Anthropic shipped Claude Code in 2025. OpenAI shipped Codex. Google shipped Jules. All three are dispatchers. All three are free at the edges and metered at scale. The orchestration pitch is the pitch that this is the wrong layer to compete on — that the moat is workflow and integration, not the routing logic itself.

The 19 investors who wrote in to OpenAI's Deployment Company three weeks ago had a related thesis, from the other direction. They paid OpenAI $4 billion to build a forward-deployed-engineer business above the model. Cognition is selling its investors a forward-deployed product above multiple models. Same insight, opposite seat at the table.

The 90% number

The most quotable line out of the round is Wu's claim, reported by The Next Web and The Decoder, that "more than 90% of the company's internal code is now written by Devin." In December 2025, the figure was 13%. That is a six-month jump from a token contribution to near-total, on the codebase the engineering team uses to ship the product the customers buy.

There are two ways to read that number. The generous read is that Cognition is the most aggressive practitioner of its own product in enterprise software, that the dogfood ratio is the strongest available signal on whether the product is real, and that 90% is the line above which the eat-your-own-dogfood criticism stops working. The cynical read is that 90% is the number you say to investors during a fundraising cycle, and that the audit of what "written by Devin" means in practice — does it cover tests, configuration, comments, scaffold, or only diffs in the commit history? — will arrive later, if ever. The TechCrunch writeup did not press on the definition. Neither did the others.

Either read leaves a useful question open. If a 100-engineer team genuinely needs to write only 10% of its own code, the implication for engineering hiring at every Cognition customer is the headline. If the 90% is the marketing number and the real share is materially lower, the headline is that Wu said it anyway and a Series D priced it in. Both are interesting. The fundraise does not require either to be definitively true; it requires investors to believe one of them.

What the customers actually pay for

The list of named enterprise customers is the part of the announcement that is the hardest to dismiss. Goldman Sachs, Citi, Santander, Mercedes-Benz, NASA, and the U.S. Army and Navy show up across the TechCrunch and The Decoder reports. Enterprise usage has reportedly grown 50% month-over-month for six straight months, per the TechCrunch number. Compounded, that is a 11x ramp over the period — which lines up with the revenue trajectory, give or take.

The Windsurf acquisition, closed last year, is the other half of the explanation. Windsurf brought roughly $82 million in run-rate and 350+ enterprise accounts into the Cognition stack, per Sacra's company profile. Stripping that out, the underlying Devin business is still growing fast but not 13x; what changed between May 2025 and May 2026 is partly Devin shipping more, partly Cognition buying a distribution channel that was already in the room. The 53x multiple does not distinguish between the two. The buyer of $1 billion of Series D paper does, presumably, ask the question. The press release does not.

What this means

Three takeaways.

1. The orchestration layer is now a financeable category, priced at frontier-lab multiples. $26 billion is in the range where a company gets compared to the labs whose APIs it consumes. Anthropic and OpenAI are still 30–40x larger by valuation, but Cognition has stopped being a curiosity and started being a counterparty. The next round — and there will be a next round — will price whether the orchestration thesis survives Anthropic and OpenAI building their own forward-deployed businesses on top of their own models. The labs have a structural advantage; Cognition has a head start measured in quarters, not years.
2. Dogfood-as-marketing is now a fundraise lever, and 90% is the new ceiling. Cognition is the first AI company to put a number above 85% on its own internal AI authorship and have it quoted as a positive in the round announcement. Every other lab in the space will be asked their equivalent figure on the next earnings call or pitch meeting. Some will produce numbers; some will refuse; the refusals will mean something. If Wu's 90% turns out to mean "90% of the lines that ended up in main" and not "90% of the engineering work," the next CEO who tries the line will get a sharper question. For now, the line worked.
3. The Founders Fund position is the through-line. Thiel's firm led the March 2024 seed, the April 2024 round, the August 2025 round, the September 2025 round, and now wrote into the Series D. Five rounds in two and a half years, on a position that opens at the $200 million valuation and closes at $26 billion. That is not a thesis trade; that is a portfolio company. The Decoder noted the participation flatly. Sacra put the position in writing. When the orchestration thesis is tested — by an Anthropic Claude Code 2.0 or an OpenAI Codex Pro — the question of whether Founders Fund stays for the next round will be the cleanest available signal on whether the thesis is still working.

The round is real. The customers are real. The revenue is real. The bet that priced it — that there is a $26 billion business sitting above the frontier labs, in a layer those labs cannot or will not build themselves — is the part still being tested. Lux, General Catalyst, and 8VC say it has been. Anthropic, OpenAI, and Google will spend the next four quarters trying to make sure it has not.