Cognition at $48B: What Devin's Disclosed Run Rate Actually Supports

Bloomberg reported on 12 August 2026 that Cognition was in early talks at a valuation of at least $40 billion. The round that actually closed on 8 September priced the Devin maker at $48 billion. Set against the almost $900 million run rate Cognition itself disclosed that day, the Cognition valuation sits at about 53 times revenue, roughly double the multiple SpaceX agreed to pay for Cursor.

From $40B talks to a $48B close: the Cognition valuation in 27 days

The headline that circulated in August was a negotiating marker. The number that matters is the one Cognition published four weeks later.

What Bloomberg reported on 12 August

The reporter was Bloomberg's Rebecca Torrence. Per her report, Cognition AI was "in early talks with investors for a new funding round that may boost the coding startup's valuation by more than 50% to at least $40 billion". The story cited people familiar with the matter. It also said the annualised revenue run rate was "now approaching $1 billion".

That second figure was not a company disclosure. It came from the same unnamed sources, so it sits a tier below anything Cognition published itself. Coverage the same day from VKTR repeated it without an additional source.

What closed on 8 September

Cognition announced a round of over $2 billion at a $48 billion valuation. Andreessen Horowitz and Accel led as new investors, alongside existing backers Founders Fund, General Catalyst and Avenir. The company stated that run-rate revenue had grown "from $492M to almost $900M" since May.

So the round came in 20% above the leaked floor, at double the round size reported in August. My reading is that the August leak functioned the way most leaks do in a competitive round. It set a public anchor that bidders then had to clear.

Seventeen days after the close, on 25 September, Cognition posted that it had "crossed $1B in annualized revenue run rate". On that figure, the $48 billion price works out at 48 times.

Every revenue figure Cognition has put on the record

A valuation multiple is only as good as its denominator. Cognition has published five revenue statements in 13 months, and they are not all the same kind of number.

Only rows marked disclosed come from Cognition's own announcements. The Bloomberg row is included because it was the basis of the $40 billion headline, not because it carries the same weight.

The 2025 baseline is a Devin figure, not a company figure

The September 2025 announcement gave a precise number for Devin alone: $73 million of ARR in June 2025. It then said the Windsurf acquisition "more than doubled" ARR. That implies a combined figure above $146 million, but Cognition never printed the combined dollar amount.

This matters for anyone building a growth curve. The jump from $73 million to $492 million is partly organic and partly an acquired editor business. The public record does not let you split the two.

What has never been disclosed

None of the four Cognition posts defines run rate. None says whether it annualises a month, a week or a day, or whether it includes usage billed in arrears. None gives a gross margin, a customer count, or net revenue retention.

The September 2025 post did offer one cost figure: "total net burn under $20M across the company's entire history". No cost figure has been published since. That is not evidence of anything bad. It is simply the edge of what can be said from the record, and the standard for what a useful disclosure looks like is set out in the ranking of AI revenue disclosure metrics.

Cognition valuation multiples, dated

Here is every pairing of a valuation and a revenue figure that the record supports. Each multiple is a simple division. Where the denominator is reported rather than disclosed, the row says so.

The talks row uses $1B and $900M as the bounds of "approaching $1 billion". The May post does not state whether $26B is pre or post money; reporting on the round described it as $26B post-money.

Between 27 May and 8 September, the valuation rose 85% and the disclosed run rate rose about 83%. Over 104 days, the price per dollar of run rate barely moved. That is the single most useful fact in this whole analysis.

A rising valuation on a flat multiple is a very different signal from a rising multiple on flat revenue. The first says investors re-priced the business as it grew. The second says they re-priced the story. Cognition's 2026 rounds look like the first case.

The year-on-year picture runs the same way. The valuation rose about 4.7 times between September 2025 and September 2026. The run rate, measured from the implied 2025 floor, rose by somewhat less than 6.2 times. If anything, the multiple has compressed.

Cognition valuation against Cursor, Replit and Lovable

A multiple means little alone. The obvious comparison set is the other AI coding companies that priced a transaction in 2026. None of them is public, so every figure below is either a company statement or press reporting, and the evidence quality varies a lot.

Cursor's range uses "over $2 billion by early 2026" and "approximately $2.6 billion in annual B2B revenue", both carried in Techzine's June 2026 deal report. Lovable's figure is Dealroom's "on track for a revenue run rate near $600 million by the end of August 2026", which was a target when the round was announced.

Cursor's price is the cleanest benchmark, and it is still noisy

SpaceX's agreement to acquire Anysphere was reported on 16 June 2026 as an all-stock deal at $60 billion. It was scheduled to close in the third quarter. Techzine's report put Cursor at about $2.6 billion of annual B2B revenue, without naming a source for that figure. The same report carried SiliconAngle's "over $2 billion" for early 2026.

That gives a band of 23 to 30 times. Two cautions apply. The price was set in SpaceX stock, not cash, so it moves with a share price that has its own story. And the option behind it was struck in April, when Cursor's revenue was lower. The longer history of that company's growth and moat question is in the analysis of Cursor's $2B ARR and valuation.

Replit and Lovable sit lower, on weaker figures

Replit raised $400 million at $9 billion in March 2026, a tripling in six months according to Trending Topics. The same page's table carries about $265 million of ARR at the end of 2025, with no traceable origin. That gives roughly 34 times, and I would not lean on it.

Lovable raised $400 million at $13.3 billion in August 2026, co-led by Menlo Ventures. Dealroom described its run rate as "on track" to reach near $600 million by the end of that month. A multiple on a target is about 22 times, and it is a softer number than anything Cognition published.

Why investors pay Cognition a premium over Cursor

A buyer paying roughly twice Cursor's multiple is making a specific claim. I see three parts to it, and only one is well supported by the public record.

The first is growth speed. Cognition's run rate went from $492 million to almost $900 million in about three and a half months. Growth at that pace, sustained, justifies a high multiple on today's revenue because today's revenue is a small fraction of next year's.

The second is the product shape. Devin is sold as an agent that completes tasks, and Cognition charts its usage in compute units rather than seats. Revenue that scales with work done has a higher ceiling than revenue that scales with headcount. That is the bet behind most of the agent pricing debate, and it is the part I find most persuasive.

The third is enterprise reach. Cognition names NVIDIA, GE Aerospace, Citi, Mercedes-Benz, Goldman Sachs and Palantir across its announcements. It disclosed under 5% overlap in enterprise customers between Devin and Windsurf at the time of the merger. Logo lists are not revenue, but they do suggest the run rate is not one or two whales.

What none of this establishes is margin. Usage-priced agents carry heavy inference costs, and a dollar of usage revenue can be worth far less than a dollar of seat revenue once the model bill is paid. That gap is the main reason I would discount the premium, and the cost side is laid out in the analysis of rising inference costs and AI margins.

What run rate the $48B Cognition valuation requires

Turn the question around. If you believe Cognition should eventually trade at a peer's multiple, how much run rate does the current price need? The arithmetic is simple and needs no forecast.

Doubling a run rate is something Cognition has done twice in the past 13 months. So the $48 billion price is not absurd on the record. It is a bet that the 2026 growth rate continues for roughly another two quarters, and that the revenue added along the way carries margins comparable to Cursor's.

My own view is that the first half of that bet is reasonable and the second half is unknowable from outside. A valuation that depends on an undisclosed number is a valuation you cannot check, and I would weight it accordingly.

Where this analysis is weakest

Every number above is a ratio of two figures that their owners chose to publish. Here is where that breaks down.

Run rate is not a standard measure

Cognition has never said what period it annualises. For a usage-billed product growing this fast, annualising the latest week rather than the latest month can move the figure materially. Cursor and Lovable figures carry the same problem, and the press reports often switch between ARR, run rate and annual revenue in the same paragraph.

This is the main reason a precise multiple is false precision. Treat every number in the multiples table as plus or minus several turns, and read the ranking rather than the decimals.

The peer figures are much weaker than Cognition's

Cognition's row is the best-sourced in the comparison, which creates an odd effect. The company that disclosed the most looks the most expensive. If Cursor's true run rate in June was above $2.6 billion, its multiple was lower still. If Replit's ARR was below $265 million, its multiple was higher. Neither can be settled from public sources.

A revenue multiple ignores what each dollar costs

The strongest objection to this whole post is that revenue multiples compare businesses with possibly very different gross margins. An agent product that runs long autonomous sessions may spend more on inference per dollar of revenue than an editor with autocomplete. If so, Cognition's premium is even harder to defend. If Cognition has better unit economics than its peers, the premium may be cheap. The record does not say which.

A fair critic would also note that a private-round valuation is not a market price. Preference terms and investor protections can make a headline valuation overstate common-equity value, a mechanism covered in the breakdown of down-round and cap table terms. Cognition's posts do not disclose its terms, so this post cannot adjust for them.

What the Cognition valuation means if you buy Devin

Most readers are not choosing whether to own Cognition shares. They are deciding whether to standardise an engineering team on its product. The valuation tells you three practical things.

First, your vendor is well funded. Over $3 billion raised across the May and September rounds means near-term continuity risk is low, even if cash burn has risen since the 2025 disclosure.

Second, the valuation is built on usage growth, and usage growth is your bill. A vendor priced at roughly 50 times revenue has every incentive to grow consumption per customer. Budget for usage-priced agents the way you would budget cloud, with alerts and ceilings, not the way you budget seats.

Third, the premium over Cursor is partly a bet that agents displace editors. If you run both, the decision about which to standardise on is a product decision, not a financial one. The valuations should not make it for you, and the hands-on comparison is in the verdict on Claude Code, Cursor and Copilot.

For the same reasoning applied to the model labs rather than the tools built on them, see the run-rate multiple analysis of OpenAI's valuation.

Frequently asked questions

How to use these numbers this week

If you are benchmarking an AI company's valuation, rebuild its multiples table yourself with a status column. Mark each revenue figure as disclosed, reported or implied. You will usually find the headline multiple rests on the weakest row.

If you are buying Devin or any usage-priced coding agent, ask the vendor one question before renewal: how is your run rate calculated, and what share of my usage would count toward it? A vendor priced at 50 times revenue will answer carefully. Whether it answers at all tells you how much of your future bill is already in its valuation.

Related reading