Cognition raised more than $1 billion this week at a $26 billion valuation, up from $10.2 billion eight months ago. The money rides on Devin, its autonomous coding agent, which the company says now runs at roughly a $492 million annualised revenue pace after months of about 50% month-on-month growth among corporate buyers. Lux Capital, General Catalyst and 8VC led the round.

The pitch is clean and, on its own terms, strong. Devin is not a faster way for an engineer to type. It is a software engineer you hand work to: give it a ticket, it opens a branch, writes the code and sends a pull request. Cognition said late last year that Devin had already merged hundreds of thousands of pull requests, and pointed to one modernisation job it finished in eight days against an estimated eight months. Mercedes-Benz, NASA, Goldman Sachs, Santander and Nubank are on the customer list, and enterprise usage is up roughly tenfold since January. The backers are paying for one idea: that the natural unit of agent work is the task, not the keystroke, and whoever owns the autonomous-agent layer owns the budget.

What I can actually verify is thinner than the headline. A $492 million figure is a run-rate, which is to say it annualises whatever happened in the most recent month. Fifty per cent month-on-month growth is real and rare, and it is also a curve no company holds for long, because the early months of an enterprise rollout are the easy ones. The customer logos are impressive, but logos are pilots until renewal, and Cognition's own descriptions point to the place these deals slow down: procurement reviews, security sign-off, and the question of how much autonomous code an enterprise will actually let an agent merge without a human reading it. The traction is genuine. The durability is the open question, and the valuation is priced on the durability.

The same bet, inverted

Now look at the company Cognition is implicitly betting against. Cursor, built by Anysphere, was valued at $29.3 billion in November and is reportedly raising at around $50 billion. Its annualised revenue crossed $2 billion in February, 67% of the Fortune 500 use it, and it generates about 150 million lines of enterprise code a day. Cursor's thesis is the mirror image of Devin's: own the surface developers live in, the editor and the workflow, and let the model underneath become a commodity you swap as better ones arrive.

So the market is pricing two opposite theories of how coding agents win, at the same time, at tens of billions of dollars each. One says the human stays in the loop and the surface is the moat. The other says the human leaves the loop and the agent is the moat. Even Cursor's own president concedes the framing has moved; the IDE, he says, "isn't the right form factor" for a world that produces ten times more code, and the "the IDE is dead" line is now common enough that the CEO of a rival, Warp, repeats it. Both companies cannot be the moat. The interesting question is what happens to the one that turns out not to be.

Near as I can tell, that question matters less than a second one both decks would rather you skip. Strip the two theories down and they share a single load-bearing assumption: that the company sitting between the developer and the model gets to keep the margin. That is the assumption Anthropic's Claude Code is in the middle of disproving.

Bring in the objection

The fair objection is that none of this is winner-take-all. Anthropic's own head of Claude Code calls it "winner-take-some, or winner-take-most." Ask working engineers and most use several tools at once: Cursor for some tasks, Claude Code for others, OpenAI's Codex for the rest. The category is big enough to feed more than one large company, with AI coding tools doing about $12.8 billion in revenue in 2026, more than double 2024. There is also a real moat argument on Cursor's side that the bears wave away too fast. Switching editors is genuinely costly; the keyboard shortcuts in an engineer's fingers took years to learn, and an enterprise that has standardised 30,000 developers on one tool does not re-tool on a whim. All of that is true, and worth taking seriously.

But these multiples do not price coexistence. They price dominance. A business doing $492 million in revenue at a $26 billion valuation is carrying a multiple near fifty times sales, and that only makes sense if it captures a category, not if it shares one three ways. Coexistence is the bear case wearing the bull's jacket. If Cognition and Cursor and Codex all settle into a comfortable split, every one of these rounds was overpriced. The valuations are a wager that one form factor wins and takes most of the budget with it.

Where the margin actually sits

Coexistence also does not fix the part that should worry both of them, which is where the profit lives. Cursor, as Fortune laid out in detail, effectively pays retail for the same models Anthropic buys at cost. One investor's line, "burning $1 to make 90 cents isn't a business," is the whole problem in a sentence. Claude Code can undercut a price Cursor cannot rationally match, because Anthropic owns the model. At a $2.5 billion run-rate and more than 300,000 business customers, Claude Code is not a side experiment. It is a serious commercial product that sits one floor below Cursor in the stack and pays wholesale for what Cursor rents.

Cursor's response is to build its own model, Composer, which is the correct move and also a confession: you cannot win the workflow if you pay your largest competitor retail to run it. Composer has beaten Anthropic's Opus 4.6 on some benchmarks, and while a later version came in behind OpenAI's GPT-5.4, it is cheap to run, which for Cursor's economics may matter more than topping a leaderboard. Cognition's response is the autonomous-agent moat plus an installed base it bought when it acquired Windsurf. Neither company is safe, because the labs are not staying upstairs. OpenAI has published Symphony, an open spec for exactly the managed, repeatable agent workflows Devin sells, and Anthropic keeps loosening Claude Code's usage limits to pull heavier work in. The application companies are climbing toward the model, and the model companies are climbing toward the application. They meet in the middle, and the side that owns the scarce input arrives with the better unit economics.

This is the oldest squeeze in technology: your biggest supplier becomes your biggest competitor, and the layer that owns the scarce input keeps the margin while the layer that rents it competes on everything else. What is unusual is the speed. Cursor went from $2.5 billion to nearly $30 billion in a single year, and the "Cursor is dead" posts started while it was still, by its own investors' description, the fastest-growing software company on record. Devin is running the same lap now, a step quicker.

The bet

So here is what both rounds actually buy. Cognition's $26 billion and Cursor's $50 billion rest on the same wager, that the company between the developer and the model keeps the profit. The clean tell will be gross margin, once either company has to show it. If, by the time real numbers surface, Cognition or Cursor can post margins that survive paying API prices, or can show they own a model good enough that they no longer have to, the multiples will look cheap in hindsight. If they cannot, the re-rating will not be gentle, and it will not wait politely for the next round to price it in. That is the bet, made twice, in opposite directions. In both cases, the house is Anthropic.