Cognition AI

IWhat it does

Cognition builds software that does software engineering work. Its main product is Devin, introduced in March 2024 as an AI software engineer. You give Devin a task, such as fixing a bug, upgrading old code or reviewing a change, and it plans the work, writes the code, runs tests and opens a pull request (a proposed code change for a human to approve)⁵². Devin became generally available in December 2024⁵.

In July 2025 Cognition signed an agreement to buy Windsurf, an AI-powered code editor, including its brand, product, team and customers. At that time Cognition said Windsurf had $82 million of annual recurring revenue, more than 350 enterprise customers and hundreds of thousands of daily users⁴. The price was not disclosed⁴. Google had just paid $2.4 billion to hire Windsurf’s chief executive and senior engineers and license its technology¹³. In June 2026 Cognition renamed Windsurf as Devin Desktop²²⁵.

Since then Cognition has widened Devin into a family of agents. The Series E announcement lists Devin Auto-Triage (a first look at incidents in running software), Devin Security Swarm (finds and ranks vulnerabilities) and Devin Automations (starts work from events in Slack, GitHub, Linear and similar tools)¹. In late October 2026 it announced that Corridor, a small security start-up, is joining to extend Devin Security. No price was given⁶. It also builds its own models, called SWE-1.5, SWE-1.6, SWE-1.7 and SWE-2, alongside models from other labs⁵¹⁴.

Cognition says its long-term aim is what it calls self-driving software, where agents take on work without being asked¹⁴. It also says Devin writes most of its own code: 89% of the code its engineers commit, per the Series D post². That is a company claim I cannot check.

IIHow it makes money

Cognition sells access to its agents. Per Sacra’s profile of the public price list, Devin has a free tier, a Pro plan at $20 a month, a Max plan at $200 a month, a Teams plan from $80 a month plus $40 a month per full developer seat, and negotiated enterprise contracts. Use beyond a plan’s allowance is billed at API-style rates¹⁹. I could not open Cognition’s own pricing page, so treat those prices as second-hand.

The unit of usage is the Agent Compute Unit, or ACU, a measure of how much machine work an agent uses. Cognition’s Series E post charts monthly ACU use rising steeply between August 2025 and August 2026 but gives no figures¹. That matters because every ACU costs Cognition real compute, and when a task uses a third-party model Cognition pays that lab as well.

Revenue is large companies first. The Series D post names Citi, Mercedes-Benz, Goldman Sachs, Elevance, Dell, Santander, the US Army and Navy, and Itaú, plus start-ups and the consulting firms Infosys and Cognizant, which resell and use Devin²⁵. The Series E post names Nvidia, GE Aerospace, Citi, Mercedes-Benz and Modal¹. The company said enterprise usage grew more than tenfold from the start of 2026 to May².

On costs, the best figures are press-reported and I mark them that way. The Information, as relayed by TechCrunch, says Cognition leases an Nvidia server cluster costing hundreds of millions of dollars a year and that total 2026 cash burn could reach $800 million¹². A second-hand report says about $200 million was burned in the second quarter and that gross margin in the enterprise business is near 50%, excluding the cost of training its own models¹⁷. The company has not disclosed a consolidated gross margin¹⁷. Whether that margin improves is something I cannot judge from public data.

Money figures, with who reported them

IIIThe market and the competition

AI coding is one of the first AI uses where businesses are paying large sums. Anthropic said in February 2026 that Claude Code, its coding product, had a run-rate above $2.5 billion, inside a company run-rate of $14 billion⁷. That is a model lab selling a coding agent directly, and it is the clearest sign of how crowded this area is. Cognition buys models from Anthropic and OpenAI, so some of its largest suppliers also compete with it¹³.

The nearest pure competitor was Cursor. It was reported to have more than $2 billion of annualized revenue when it was in funding talks in April 2026, and about $4 billion by the time of the sale, per Forbes via TheStreet¹²¹⁶. SpaceX exercised an option to buy Cursor’s parent, Anysphere, for $60 billion in an all-stock deal, and the deal closed in August 2026 per TheStreet and Techzine¹⁴¹⁶. TechCrunch reports investors said Cursor sold largely because it was short of compute¹². That is a warning for any company whose product depends on buying large amounts of compute.

Other rivals include GitHub Copilot, OpenAI’s Codex, and a long list of start-ups and open-source tools. I have not opened current figures for Copilot or Codex, so I do not quote any. TechCrunch’s reading is that investors see AI coding as not winner-take-all, which is the bet behind Cognition’s price¹². a16z, which backed Cursor and then led Cognition’s Series E, is also reported to be leading a round in another Cursor competitor¹²¹⁶.

Cognition’s own pitch is a different shape of product: agents that run in the cloud on their own for hours, rather than an assistant inside an editor. Its enterprise sales, bank and government clearances (it announced a FedRAMP High in-process status in July 2026) and system-integrator partners are aimed at large buyers who must review security and legal terms⁵. The bet is that these buyers value an agent with an audit trail more than a cheaper chat window. I cannot tell from public data how much of the $1 billion run-rate is Devin, how much is the Windsurf and Devin Desktop editor, and how much is a one-off or pilot.

IVThe moat and how it could be disrupted

I rate the moat contestable. That means the company has real advantages today, but I do not think they are durable. The honest starting point is that the core ingredient, a strong language model, is made by someone else for most of the work. Cognition has built its own models, SWE-1.7 and SWE-2, and says Fusion mode cuts cost by 60% against frontier models⁵, but those are company claims and I have not seen independent tests.

What Cognition does have is a head start in enterprise deployment: reference customers in banks, carmakers and government, partnerships with Infosys, Cognizant and AWS, and the Windsurf user base²⁵. Switching costs exist once an agent is wired into a company’s repositories, tickets and security rules. But these costs are modest compared with, say, a database. In my judgement teams can try a new agent quickly, so loyalty is weak.

The ways this breaks are concrete. One: the model labs ship better agents of their own and bundle them, as Claude Code already shows⁷. Two: price. If the cost per task falls quickly, the premium for a specialist wrapper shrinks. Three: margin. If Cognition must keep paying for compute and for third-party models, its press-reported 50% gross margin may not improve¹⁷. Four: a bigger-balance-sheet rival. Cursor now sits inside SpaceX¹⁴¹⁶. Five: customers decide to build agents in house.

One more point on the Windsurf deal. Cognition did not buy a full company in the ordinary sense. Google took the leadership and technology licence first, and Cognition got the product, users and remaining staff¹³⁴. The Windsurf name has since been dropped for Devin Desktop²². That limits how much brand strength was bought.

VThe numbers we can see

There are no audited accounts. The only company data are announcements. Cognition reported $492 million of run-rate revenue in May 2026, nearly $900 million at the Series E on 8 September, and over $1 billion on 25 September²¹³. The company does not define run-rate. TechCrunch notes it is usually one month of revenue times twelve¹². That flatters a business that is growing quickly or has usage-based spikes, because one strong month sets the number.

A few details complicate the growth line. TNW calculates $37 million of run-rate revenue in May 2025¹³, but Windsurf’s $82 million, bought two months later, is in the later numbers⁴. So the climb from $37 million to $1 billion is partly bought, not all grown. A second-hand report puts August at about $900 million and says the figure had more than tripled since the start of the year¹⁷. Sacra, which is an estimate and not company data, puts September at $1 billion¹⁹. These are consistent with each other but all trace back to company statements.

On valuation, TNW computed about 53 times run-rate for the Series D at $26 billion and $492 million¹³. TheStreet says the multiple stayed flat at about 53 times for the Series E¹⁶. On the over $1 billion figure the multiple is about 48 times. For comparison, Cursor’s $60 billion price was about 15 times a $4 billion run-rate, per Forbes via TheStreet¹⁶.

Sources disagree in a few places, and I show both sides. On the forecast, one second-hand report says internal goals are above $1.5 billion by the end of 2026 and $4 billion to $5 billion in 2027¹⁷¹⁸. TechCrunch, citing The Information, says $4 billion to $5 billion by the end of 2026¹². Both cannot be right, and neither is a company statement. On the Series D, Cognition and TNW call it $26 billion, while a headline linked by TechCrunch calls it $25 billion pre-money²¹³¹², which is consistent with $26 billion after new money of about $1 billion. For the Series E, Cognition says only that it raised over $2 billion at $48 billion, and one outlet reports the Wall Street Journal put the round at $2 billion¹¹⁵. Bloomberg was reported to have said the week before that the target was about $1 billion at $47 billion, so the round came in larger than first reported¹⁴.

How to read the run-rate. The company says run-rate, not revenue. Over the past 12 months the company has booked far less than $1 billion, because the business was much smaller early in the period. The May 2026 figure of $492 million was described as a multiple of about 13 times the year before, but that comparison includes the Windsurf purchase¹³⁴. A usage-based product also moves with each customer’s activity. If a few large customers run big migration projects in one month, annualising that month can overstate steady revenue. I do not know if that is the case here, and the company does not say.

Public software companies give a sense of scale as of 2 and 3 November 2026: Palantir at 57.4 times trailing sales, Datadog at 18.6 times, Atlassian at 7.2 times and GitLab at 7.0 times⁸⁹¹¹¹⁰. Cognition at about 50 times sits near Palantir, which is growing far faster than the rest of that group and has years of profit. Cognition has none that is public.

Valuation history

VIWho runs it and who owns it

Scott Wu is co-founder and chief executive¹². Steven Hao is co-founder and chief technology officer, and Walden Yan is a co-founder²⁰. A March 2024 report described Yan as chief product officer; I could not find a more recent statement of his title²⁶. Forbes describes the three as gold medalist coders who met on the competitive circuit²⁰, and Cognition says its founding team holds 10 gold medals from the International Olympiad in Informatics, a programming contest²¹. Russell Kaplan is president and Alex Stamos is chief security officer, as named in the Corridor post⁶. The founding date differs by source: The Next Web says 2023¹³ and the San Francisco Standard dates the founders’ first idea to late 2023²⁵, while Cognition says it started in January 2024³ and TechCrunch also says 2024¹².

Ownership is spread across many venture firms and funds. The Series E was led by Andreessen Horowitz and Accel, both new. Existing investors Founders Fund, General Catalyst and Avenir took part, with a long list of others including Benchmark, Bessemer, Kleiner Perkins, Greylock, Lightspeed, Altimeter, T. Rowe Price and DST¹. The Series D leads were Lux Capital, General Catalyst and 8VC². I cannot see the cap table, the preferred terms or how much the founders own.

Avenir said the Series E cheque was the largest it has written, per PYMNTS¹⁵. Avenir also cites a market that it sizes at $26 billion today with a path to $800 billion¹⁵. That is an investor’s view, in the same post that announced its stake, and I treat it as a sales pitch for the round, not as evidence. The sources do not give a figure for how much of the company the new investors bought, and an over-$2-billion round at $48 billion would mean roughly 4% sold, if the valuation is post-money (my arithmetic: 2 divided by 48). If it is pre-money the share is slightly lower.

One thing to note on incentives: Nvidia is a customer using Devin for chip design and, per TheStreet, an investor in the Series E, and a16z backed Cursor and profited from its sale before leading Cognition’s round¹¹⁶. These links are not wrong, but they are part of how a round of this size got done. Late-stage preferred shares usually carry rights that sit ahead of common stock, so a sale below the $48 billion price can pay the latest investors in full while employees’ shares get less. I have no access to the actual terms.

VIIValuation: is the last price supported?

Equity value in US$ billions. Not a share price: a private company has none. Latest mark: primary round (Series E), pre/post-money not stated.

What the last valuation assumes

At $48 billion, with my base discount rate, dilution and net cash, the price needs about $111 billion of enterprise value in 2030. At 7 times revenue that is roughly $16 billion of revenue, and at 10 times it is about $11 billion. It is also priced at about 48 to 53 times run-rate today³¹³. The company reports a run-rate that is not defined, and the costs are press-reported¹²³.

Bear, base and bull

• Bear. Run-rate of about $1B grows to $2.5B of revenue in 2030 as agents become a bundled feature and prices fall. Margins stay thin. Terminal multiple 5x revenue, below Atlassian at 7.2x and GitLab at 7.0x. Discount rate 25% (late-stage private, illiquid; my judgement). Dilution 15% by 2030. Net cash at the end of the period $1.5B after continued burn (my estimate). Result about $5.7B, shown as 6.
• Base. Revenue reaches $8B in 2030, about 68% a year from $1B, below the $4B to $5B for 2027 that one report attributes to internal goals¹⁷. Gross margin moves toward the 70s and the company is profitable by then. Terminal multiple 7x revenue, equal to Atlassian and GitLab, as growth slows. Discount rate 20%. Dilution 12%. Net cash $2.5B. Result about $25.4B, shown as 25.
• Bull. Revenue reaches $20B in 2030 as autonomous agents do a large share of software work and Cognition holds a top-three position. Terminal multiple 10x revenue, between GitLab and Datadog at 18.6x. Discount rate 18%. Dilution 10%. Net cash $3B. Result about $92.8B, shown as 93.

Comparables used

Where I come down

The latest valuation is about double my base of $25 billion and closer to it than to my bull case of $93 billion. To make it right you have to believe revenue keeps doubling for two more years, margins rise sharply from today’s reported level, and one of the few independent players holds off the model labs and a SpaceX-backed Cursor. The market may be right even so. Run-rate went from $492 million to over $1 billion in four months²³, investors paid 15 times for Cursor on a bigger revenue base, and a buyer such as a big cloud or chip company might pay more for strategic reasons¹⁶. My margin of safety price is $19 billion, about 40% of the round price. Investors bought preferred shares with protections that I cannot see, which may make the headline valuation overstate the common stock’s worth.

What I cannot judge

I cannot judge how much of the $1 billion run-rate is recurring, how it is defined, how much is one-off or pilot revenue, or what share goes to the Windsurf and Devin Desktop editor. I cannot see audited accounts, gross margin overall, cash balance, or the terms attached to the preferred shares. The burn and margin figures are press-reported and second-hand¹²¹⁷. I also cannot say if the Series E was pre- or post-money. My model is a rough guide with a wide range, and a different view of the 2030 revenue changes the answer more than anything else.

VIIIThe IPO

No IPO plan reported. I found no IPO filing or report of IPO preparation as of 3 November 2026. No S-1 or confidential filing was reported in the sources I opened. Scott Wu said on Bloomberg Television that the May raise lets Cognition stay independent, per TNW¹³. I did not check SEC EDGAR directly. Cognition was also not the subject of a reported takeover bid; Cursor’s sale to SpaceX is the closest event in the field¹⁴¹⁶.

What to watch

• A public or confidential S-1 filing, or reports of bank hiring
• Whether run-rate numbers are followed by audited or defined revenue figures
• Disclosure of gross margin and cash burn
• A tender offer or secondary sale that prices the shares below or above $48 billion
• New model-lab products that bundle coding agents, as Claude Code does⁷
• Any sale approach from a larger company

IXWhat could go wrong

What most likely destroys this business is not a collapse in demand for AI coding. It is a slow squeeze: model labs give away or bundle capable agents, prices fall, and a specialist company ends up with sales growth but low margins and large compute bills. The reported 50% gross margin and $800 million burn point at that risk¹⁷¹².

Several things have to go right together. As I read it, Cognition needs: (a) revenue to keep growing about 70% a year for four more years, (b) gross margin to move up from near 50% to the 70s, (c) customers to stay with a specialist when labs offer rival products, and (d) funding markets to stay open until it reaches break-even. If I give each a 55% to 65% chance and treat them as loosely linked, the combined odds are about one in five to one in eight. That is my estimate, not a measured figure, and the real odds depend on how closely the pieces move together.

Other risks: Cursor’s sale for compute reasons shows the cost of running out of supply¹². Cognition leases Nvidia hardware¹². Run-rate numbers can fall as quickly as they rise if customers move to another tool. An acquisitive streak (Windsurf, Dioxus, TierZero, The Interaction Company, Corridor) adds integration work and spending⁵⁶. And 90% of its own code is written by Devin per Cognition, which, if true, is a benefit, but also means a faulty agent update could break the company’s own product¹³.

Concentration and quality risks are also worth naming. The Series E post lists five named customers, and the 1B post names four others, yet neither gives a customer count or the share of revenue from the largest accounts¹³. If a few banks or manufacturers make up a big part of the run-rate, losing one would show up quickly. Agents also make mistakes: an AI that changes production code can introduce faults, and the large customers Cognition targets have strict rules on audit and liability. The company’s productivity guarantee and security programmes show it knows this⁵, but I have not seen results from independent testers.

Legal and supply risks are lower-profile. Cognition depends on a few chip and model suppliers, and it has said it is moving toward its own models to reduce that reliance¹²¹⁴. If its own models fall behind frontier ones, it must either pay up for outside models or accept weaker results. Neither outcome helps margins.

What an owner can really take out depends on exit. There is no dividend and none is likely, since the company burns cash. The only routes are an IPO or a sale, and at this price the number of buyers who could pay is small. The next best use of the same dollar might be a public company in the same field. Datadog trades near 19 times sales and Atlassian near 7 times, with audited public accounts⁹¹¹. That is not a recommendation, only a way to see how much extra one is asked to pay for Cognition’s growth.