Cognition is raising about a billion dollars at a $47 billion valuation. Every take I read called it the frothiest number of the year. Do the arithmetic and it is the opposite, which is why it is worth your time.
Bloomberg reported on September 1 that the maker of Devin is closing a round at roughly $47 billion, up from $26 billion in May. Investor interest reportedly reached nearly $10 billion against a $1 billion raise. Annualized revenue is now reported above $900 million, more than double the $492 million figure from late May. Bloomberg notes talks are ongoing and terms could still change, so treat the number as a strong signal rather than a closed fact.
The multiple did not move
In May, $26 billion against $492 million of annualized revenue was about 53 times. Today, $47 billion against $900 million is about 52 times. Revenue grew 83%. The valuation grew 81%.
That is not what a mania mark looks like. A mania mark is the multiple running while revenue stands still. This is the same price per dollar of revenue that a different set of investors paid four months ago, applied consistently across two rounds. Whether 52 times is the correct number is a real argument. But it is a disciplined one, and it moves the interesting question off the valuation entirely.
The question is whether the $900 million is durable. And that is not a venture question. It is the same supplier question I keep landing on.
Your supplier is your competitor, again
Cognition does not train frontier models. Devin sits on top of them. Every capability gain that makes Devin better is one the model providers can ship directly into their own agent products, and all of them now do. The moat has to be the harness, the enterprise integration and the workflow, not the intelligence. That is a real moat. It is a thinner one than 52 times revenue implies.
I wrote about this two weeks ago from the ugly end, when OpenAI terminated Cursor’s model access over a change of ownership. The lesson there was that four years of partnership and $3 billion in revenue bought no durable right to the inputs. Cognition is the same structure with a better mark on it.
Which brings up the part that actually explains $47 billion, and it is not the fundamentals. SpaceX closed its $60 billion acquisition of Anysphere, Cursor’s parent, in August — the largest startup exit on record. That removed the single most comparable independent asset from the market. If you are a large fund that has decided it needs coding-AI exposure and Cursor is no longer purchasable, your shortlist is very short. Ten billion dollars chasing a one billion dollar round is what a short list looks like from the inside. Some meaningful part of that 52 times is a scarcity premium.
Scarcity premiums do not survive new entrants. Moats do. They are not the same thing and the multiple cannot tell them apart.
What I think
For scale on how fast this has moved: $500 million at $9.8 billion in August 2025, $400 million at $10.2 billion a month later with Founders Fund leading, $1 billion at $26 billion in May 2026 led by Lux Capital, General Catalyst and 8VC, and now this. Thirteen months, roughly five times the mark, with revenue tracking it the entire way. Windsurf was acquired in July 2025 and folded in.
None of which matters to you unless you are writing a check. What matters is that a lot of teams are about to standardize on a coding agent because a $47 billion valuation reads as durability, and it does not mean that. Three questions I would put to any agent vendor in this category, Cognition included, before I made it load-bearing.
Who supplies the intelligence, and on what term? Not which models it supports today. Which contracts exist, how long they run, and what the change-of-control and termination provisions say. A vendor that treats this as commercially sensitive is telling you it is a risk they carry and you inherit.
What survives if the supplier becomes the competitor in earnest? Every frontier lab ships an agentic coding product now. Ask the vendor to describe, concretely, what their product does that the model provider’s own agent will not do in twelve months. If the answer is a list of integrations, that is a genuine answer and a defensible one. If the answer is about model quality, they are selling you somebody else’s asset.
What is the exit cost? The one nobody prices. If this vendor is acquired, repriced, or cut off from its models, how much of your engineering workflow is welded to it? That number is knowable today and gets much harder to compute after you have three hundred developers depending on it.
So stop pointing at $47 billion. Point at $900 million, and ask whether it is still there in twelve months once the model providers have shipped their own agents into the same enterprise accounts. If it is, this round will look cheap and the disciplined investors will have been right. If it is not, the problem was never the multiple. It was that everyone, including the careful people, was underwriting a revenue line that belonged to somebody else’s platform.
The market has gotten good at pricing AI coding revenue. It has not yet had to price what happens when the supplier decides it wants that revenue back.
If you are working through an AI initiative and want a second technical opinion, or you want to talk through any of this, connect with me on LinkedIn or see the production side at coreagenticgroup.com.
