Cognition Hits $1B ARR Run Rate After Closing at $48B

Cognition, the maker of the autonomous coding agent Devin, is on track to reach $1 billion in annualized revenue based on its performance this month, Bloomberg reported on September 25, 2026. The milestone lands just over two weeks after Cognition closed its Series E on September 8, 2026, raising more than $2 billion at a $48 billion valuation, nearly double the $26 billion mark it set roughly four months earlier. Together the two data points complete the picture this blog covered in mid-August, when Cognition was reported to be in early talks for a round at "at least $40 billion." That round has since closed at a higher number than the early talk figure, and the revenue behind it has kept pace almost exactly.

~$1BAnnualized run rateBloomberg-reported trajectory, Sept 25, 2026

$48BSeries E valuationclosed Sept 8, 2026

$2B+Series E raisedled by a16z and Accel

~53xRevenue multipleroughly flat, May vs. September

$492M to $900MRun-rate growthSeries D to Series E close, ~4 months

What actually crossed $1 billion, and what didn't

The word doing the most work in this story is "run rate." An annualized run rate takes a recent, shorter period of revenue and multiplies it out as if that pace held for a full year. It is a projection built from a snapshot, not a count of dollars already collected and recognized. Cognition's own reported figures make the distinction concrete: the company told investors its run-rate revenue topped $900 million as of its Series E close on September 8, up from $492 million at its Series D close in May. Bloomberg's September 25 reporting describes the company as "on track to generate annualized revenue of $1 billion" based on that month's performance specifically, which is a further projection layered on top of the September figure, not a confirmation that $1 billion has already been billed and collected over a trailing twelve months.

Run-rate revenue roughly doubled in about four months

Cognition's reported annualized run-rate revenue, in millions of dollars, at three 2026 checkpoints. The September 25 figure is a Bloomberg-reported trajectory based on that month's performance, not a closed annual total.

Annualized run-rate, not recognized annual revenue. A run rate projects a recent short period forward across a full year, so it moves with whatever drove that period and can overstate a steady annual pace.

None of that makes the number meaningless. Growing a run rate from $492 million to roughly $900 million in about fifteen weeks, then trending toward $1 billion within the following two and a half weeks, is a genuinely fast growth curve by any standard, private or public. But a run rate is sensitive to whatever drove the measurement period, a large enterprise contract signed that month, a pricing change, a burst of usage, in a way that a full year of recognized revenue is not. Reporting on Cognition's business has consistently used run-rate language rather than audited annual figures, and this piece follows that same framing rather than treating "$1 billion" as a settled, backward-looking total.

A valuation that doubled without the multiple moving

The more analytically interesting detail sits in the ratio between the two headline numbers rather than in either one alone. At the May Series D, a $26 billion valuation against $492 million in run-rate revenue works out to roughly 53 times revenue. At the September Series E, a $48 billion valuation against a run rate near $900 million works out to almost exactly the same multiple, again around 53 times. The valuation nearly doubled in under four months, and the revenue multiple barely moved, because the revenue nearly doubled too.

The valuation nearly doubled in under four months, and the revenue multiple barely moved, because the revenue did too.
On Cognition's flat ~53x multiple across two rounds

That is a meaningfully different story from a valuation re-rating on sentiment alone. When a multiple expands, investors are paying more for the same dollar of revenue, a bet that the company's growth, margins, or competitive position have improved independent of the top line. When a multiple holds flat through a step-up this large, the market is instead pricing the new round almost entirely off realized growth, extending the same yardstick rather than a more generous one. Commentary on the round has placed that ~53x multiple between comparable software benchmarks, above Datadog's roughly 40x at scale and below the roughly 100x some analysts have attached to Snowflake near its 2020 IPO peak, framing Cognition's new investors as pricing a category leader without yet awarding it category-defining scarcity value.

Where the revenue is coming from

Cognition's reported revenue is not a single product line. It spans two subscription businesses: Devin, the autonomous coding agent, and Windsurf, the AI-native coding IDE that Cognition acquired in 2025 and has since integrated, shipping a proprietary coding model (SWE-1.5) and embedding Devin directly inside the editor. Pricing runs on two tracks: a $20-a-month Core plan aimed at individual developers, and usage-based enterprise contracts billed through Agent Compute Units that scale with how much work an agent actually performs, rather than a flat per-seat license.

That structure matters for how durable the revenue is likely to be. A consumption model tracks usage more tightly than a seat-based SaaS license, in both directions, revenue rises when engineering teams lean on the agent harder, but it can also fall quickly if usage drops, unlike an annual seat contract that keeps billing through a slow quarter. Reported enterprise names give a sense of where that usage is concentrated: Goldman Sachs, Citigroup (reportedly running roughly 40,000 developers on Devin), Dell, Santander, Mercedes-Benz, Cisco, Palantir, Nvidia, NASA, Infosys, Nubank, and both the U.S. Army and U.S. Navy have all been reported as customers across Cognition's own announcements and subsequent coverage of the funding rounds. That is a genuinely broad base across finance, industrials, government, and technology, which reduces the risk that the run rate rests on one or two outsized accounts.

South Park, the small park and surrounding neighborhood in the SoMa district of San Francisco — South Park in San Francisco's SoMa district, the neighborhood where Cognition has reported keeping its headquarters. The photo shows the public park itself, not Cognition's office.

The durability question underneath the growth curve

Cognition is reportedly still operating at a loss, with dedicated Nvidia server clusters cited as its largest cost line running into the hundreds of millions of dollars a year, a normal position for a company scaling agent infrastructure this fast but a reminder that a fast-growing top line and a healthy business are not the same claim. The more structural question is how defensible that revenue is likely to stay. Two pressures sit on either side of Cognition's margin. On the cost side, foundation-model providers have continued cutting inference prices, Anthropic reportedly cut cache-read pricing by 75% earlier in 2026, which lowers Cognition's own compute bill over time but also lowers the bar for anyone else to build a comparable agent. On the product side, open-source coding-agent orchestration layers have matured enough that some observers now treat the "agent wrapper" around a model as a thinner, more commoditizable layer than the underlying model itself.

OpenAI

Anthropic

NVIDIA

Google

The frontier labs whose pricing and model releases set both the cost floor and the competitive ceiling for every coding agent built on top of them, Cognition included.

Set against that, Cognition's own bet, reflected in its acquisition of Windsurf and its shift to a metered enterprise product, is that distribution, integration depth, and workflow trust inside large engineering organizations compound in a way that a raw model capability gap does not. Coding is one of the few domains where agent output is directly checkable, code compiles or it doesn't, tests pass or they don't, which is part of why investors have been willing to underwrite category leaders in it at multiples like this one. Whether that checkability translates into switching costs sticky enough to hold a 53x multiple once growth naturally slows from a fifteen-week doubling pace is the open question the next few quarters of reporting will actually answer, not this one.

The practical takeaway

For engineering organizations evaluating tools like Devin, the more durable lesson from this round is less about Cognition specifically and more about the pace of change underneath every agent in this category. Model capability, pricing, and the competitive field are all moving quickly enough that the best agent for a given codebase this quarter may not be the best one next year, regardless of how a single company's valuation is trending. That is the same argument for keeping AI tooling model-agnostic that applies across the stack, and it is the premise behind workspaces like Metir AI, which are built to let a team change which model or agent sits underneath its workflow without rebuilding the workflow itself.

Cognition's two milestones this month, a $1 billion revenue trajectory and a $48 billion valuation with the multiple behind it barely moving, are consistent with each other and with the broader run of 2026 AI financing: real, fast-growing revenue commanding rich but not obviously irrational prices. Whether that revenue proves as durable as the multiple assumes is not yet something either number can answer on its own.

Sources:

Image credits

Hero and in-body figure: "South Park Facing NE" by Lexi Mattick, licensed under CC BY 4.0. Source: Wikimedia Commons. Reviewed before publication; shows the public park in San Francisco's South Park neighborhood, where Cognition has reported keeping its headquarters, not Cognition's office or any Cognition branding.