Cognition is reportedly raising at $40B, three months after $26B

Cognition, which makes the Devin coding agent, is reportedly in talks to raise at a valuation of at least $40 billion, three months after raising $1 billion at $26 billion.

The reported basis is a $1 billion annualised revenue run rate, TechCrunch reported.

That works out to roughly 40 times revenue, which is a number worth holding still and looking at rather than scrolling past.

What the Cognition valuation multiple assumes

Mature software companies trade in single-digit to low-double-digit multiples of revenue. Fast-growing ones command more, and 40 times prices in years of continued growth at high margin.

Annualised run rate is one good month multiplied by twelve. It is not the same claim as a year of revenue, and the difference matters most when growth is this fast.

Margin is the specific question for an agent company, because every task run costs inference. Our piece on what a valuation hides covers why the headline number rarely answers it.

Coding agents are unusually expensive to run, too. Meta’s competing product fans work out to parallel sub-agents, and parallelism multiplies token spend per task.

The competitive picture got worse, not better

Between the May round and this one, the market for coding agents filled up.

Meta shipped its first coding agent on 5 August, taking on Anthropic and OpenAI directly, per CNBC. Google released an entry-level model aimed at coding and agent projects eight days later.

So a company whose product sits on top of models is being valued more highly while the model providers move into its category and price the underlying capability down.

That is the structural risk our piece on which layer keeps the money describes, and it applies to every application company built on somebody else’s model.

What would justify the Cognition valuation

Retention, mostly. If teams that adopt Devin keep it after the first renewal, the multiple starts to look like a growth premium rather than a bet.

Watch gross margin if it ever gets disclosed, since an agent business with thin margins is reselling inference with a workflow attached.

The comparison worth making is against the model providers themselves. If a lab bundles an equivalent agent into an existing subscription, the application layer has to justify a separate line item.

And watch the review bottleneck. CodeRabbit raised $143 million to help companies cope with AI-generated code on the same day, which suggests the output is arriving faster than organisations can absorb it.