Expand menu

Cognition announced yesterday a Series E of more than $2 billion at a $48 billion valuation. Andreessen Horowitz and Accel led it, with Founders Fund, General Catalyst, Avenir, Benchmark, Bessemer, Kleiner Perkins, Greylock and more than thirty other funds behind them. In the same post, the company says its run-rate revenue went from $492 million to almost $900 million since May, and names NVIDIA, GE Aerospace, Citi, Mercedes-Benz and Modal among the customers of Devin, its coding agent.

Almost every headline stopped at the $48 billion. It is the biggest number and the least informative one. What matters is in the division.

In May, Cognition raised $1 billion at $26 billion on $492 million of run-rate revenue. That is 52.8 times revenue. Yesterday it raised at $48 billion on almost $900 million. That is 53.3 times revenue. In four months the valuation multiplied by 1.85 and the revenue by 1.83. The multiple did not move.

That changes the reading entirely. A round where the multiple expands is the market repricing the story: I believe in you more today than I did four months ago. A round where the multiple stays flat is something else. It is the market saying the thesis was already in the price back in May, and that the only thing that changed is the denominator. Nobody raised the bet. They put more chips down at the same price per chip.

That denominator deserves a careful look. Run-rate revenue, as this market uses the term and as TechCrunch notes in its coverage, is one month's top line multiplied by twelve. Almost $900 million annualized means roughly $75 million billed in a single month. It is not a year of revenue collected, nor a twelve-month contract signed: it is a thirty-day snapshot, extrapolated. In a per-seat subscription business that snapshot is fairly stable. In a compute-consumption business, where the customer decides every week how many agents to run, it is far more fragile. The same metric that makes growth look like a straight line can turn it into a staircase down at exactly the same speed.

There is a second data point that frames the picture, and almost nobody is connecting it. TechCrunch recalls that Cursor was negotiating at $50 billion in April before SpaceX bought it for $60 billion, with more than $2 billion in run-rate revenue at the time, and that part of the reason for that sale was access to compute. In other words: the company with triple the revenue ended up selling, and one of the reasons was that it could not secure the infrastructure to keep growing. Cognition today is worth less in absolute terms and trades at a higher price per dollar of revenue. Put those two facts together and the message is that in this market the price is not set by billings. It is set by who has locked in the electricity and the chips to serve the demand they already created.

What this means if you run a technology company

Cognition's own sentence matters more than its valuation. The company writes that engineers should operate more like architects and delegate execution to swarms of agents.

If that holds even halfway, what gets repriced is not the software. It is the unit of sale.

A development firm that bills by the hour is, at bottom, selling execution. And execution is precisely the part these products are pushing toward zero marginal cost. You do not have to believe the agent writes the whole system to see the problem: it is enough that it writes the repeatable part, which is the part that bills the most hours and requires the least judgment. The margin on that line compresses first, and it compresses quietly, because the client never asks for a discount. They simply start asking for fewer hours.

What does not compress is the part Cognition calls architecture and that in practice is something else: deciding what to build, what not to build, where to draw the boundary between systems, what can be automated without leaving debt nobody knows how to pay. That is still scarce, and for now still human. The difference is that it used to be billed wrapped inside the execution hours, and now you have to know how to charge for it on its own.

At Indrox we have spent months moving in that direction, and not out of foresight: because the projects where the team showed up with judgment rather than hands are the only ones where the price was never argued. The operational conclusion is an uncomfortable one for a services company: if your proposal can be described as a number of people times a number of weeks, you are already competing against a swarm of agents that does not sleep.

My read

I do not think this $48 billion is a bubble, and I do not think it is a validation either. It is something else: it is an extrapolation taken seriously. Investors did not pay more per dollar of revenue than they did in May, which suggests they are disciplined on price. But they paid on top of a denominator that is, by construction, one month of billings stretched to twelve, in a market where the customer can switch off consumption on a Monday.

You do not settle that by staring at the valuation. You settle it at the next round. If the multiple stays flat again when revenue doubles once more, the market is discovering a real price. If it expands, that is enthusiasm. And if revenue stops doubling, neither question will matter.

In the meantime, the part I can act on today is not in the round. It is in the sentence. The engineer as architect is not a product promise: it is a pricing warning for all of us who sell custom software.