On September 9, Cognition closed a Series E of more than $2B at a $48B valuation, led by Andreessen Horowitz, Accel, Founders Fund, General Catalyst and Avenir.
The valuation is the number everyone will repeat. The number worth keeping is the other one Cognition published: run-rate revenue has gone from $492M to nearly $900M since the May round. Four months.
What was actually named
The follow-on list runs long: Benchmark, Bessemer, Kleiner Perkins, Greylock, Lightspeed, Altimeter, Bond Capital, Meritech, Atreides, Valor, T Rowe Price, Lux, 8VC, D1 and 137 Ventures, among others.
Price and revenue moved at nearly the same rate. In May the company raised $1B at $26B. In September it is $48B, up 85%. Revenue over the same window is up 83%. That coupling is rare in this year’s AI rounds, where valuation usually runs several quarters ahead of the income statement.
The cohort point sharpens it. Harvey priced at $15.5B on the same day against reported ARR past $400M, roughly 39 times revenue. Cognition is roughly 53 times. Both are expensive. Only one of them nearly doubled its revenue base between rounds.
Devin’s customer list has moved outside technology: the US Army and Navy, NASA, ServiceNow, Infosys, Goldman Sachs. Cognition opened offices in Washington DC, Tokyo, Singapore, London, São Paulo and Madrid over the past year, on top of San Francisco, New York and Austin. Last year it bought what was left of Windsurf’s assets and team, after Windsurf’s founders and core staff went to Google in a $2.4B deal that included a technology licence.
The category is consolidating
The Windsurf detail is not trivia. A year ago that company’s founders and core engineers left for Google under a licensing deal, and Cognition bought the remainder. Coding agents are collapsing from a crowded field into a handful of vendors large enough to sell to a defence department.
Consolidation matters here for a reason specific to labour. A fragmented tool market gets bought by individual teams on individual budgets, which produces pilots. A consolidated one gets bought centrally, on a standard, at a negotiated seat or usage price, which produces policy. The nine offices Cognition now runs across three continents are the shape of a company selling to procurement, not to engineering managers.
The job description in the blog post
The line worth stopping on is not financial. In its Series E post, Cognition described where the product goes next: Devin becomes more proactive by default, and humans act as overseers — ordering tasks, setting goals, defining priorities.
That is a job description, published by the vendor. It does not say engineers get replaced. It says the doing gets moved, and what remains is assignment and acceptance. Those two things are priced very differently. People who write the implementation are paid for output. People who assign and accept are paid for judgment, and a team needs far fewer of the second kind than the first.
This is the third instance of the same move in two weeks. Salesforce took its pricing off the seat at Goldman Sachs yesterday. Harvey shipped a legal agent benchmark before it raised. Three companies, three industries, one mechanism: take output that used to live inside a person and turn it into a unit that can be priced on its own.
The case against
The strongest counter-evidence is a month old. In August, Meta planned to shrink teams by 60% and cancelled the project when the agents did not deliver. Budget was not the problem. Capability was.
A revenue curve proves that enterprises are willing to pay to try. It does not prove that what they tried covered a person’s work. So read this round as two separate claims. First, nearly $900M in run-rate means procurement has already happened, including inside the Army, NASA and Goldman Sachs, which are among the slowest buying processes in existence. Sign-off there implies an internal acceptance test was passed. Second, whether it was passed on acceleration or on substitution is a question nobody has published a denominator for.
Exposure sits on the ticket-takers
Inside software engineering, the first thing to go was never architecture or trade-off. It is the stretch that runs: pick up a ticket, implement to the description, submit, wait for review. Devin’s proactivity roadmap points directly at that stretch.
The hiring line moved first. In August we logged CBRE’s count: AI roles are 31% of US tech postings, and non-AI postings fell 60% year over year. A large share of that 60% is junior and mid-level implementation work.
The other half of that trade is hiring, and it is worth naming precisely rather than waving at. Two days ago Accenture committed to staffing 1,000 forward-deployed engineers against Google’s agent stack, in the same week YouTube reported a 37% cut in support handle time. Those 1,000 roles are real jobs and they are engineering jobs. They are also a different job: the work is deploying, configuring and babysitting somebody else’s model inside a client’s systems, not building the system.
So the honest framing is not destruction, and it is not redeployment either. It is a swap at an unfavourable ratio, with a skills gap in the middle. The implementation roles disappearing are entry-accessible and numerous. The deployment roles appearing are fewer, more senior, and require exactly the production experience that the disappearing roles used to confer. That gap is where the next two years of career damage lives.
If you write code for a living, three questions test your own exposure. How much of last month’s work started from a ticket someone else had already written clearly? How much of your judgment gets exercised before the ticket exists? And when your team buys AI coding tools, does the business case say faster or say fewer? The answer to the third usually shows up in which budget line the purchase comes out of — tooling or headcount.
There is no need to call it early. The result of this year’s acceptance tests will surface in next year’s IT budgets and engineering hiring plans, not in funding announcements.
Sources
