Cognition AI & Devin: The $26B AI Software Engineer Story

Welcome back to our GenAI Unicorns series. It’s been a while since the last episode, but we’re back – with a fascinating story of Cognition, the lab that launched Devin. In 2025, the company raised more than $400 million at a $10.2 billion post-money valuation. In May 2026, Cognition raised more than $1 billion at a $26 billion valuation and reported $492 million in run-rate revenue. By August, Bloomberg reported early talks for another round at $40 billion or more, with annualized revenue run rate approaching $1 billion. Their story fuses math-prodigy origins, a culture of extreme intensity, and one of the wildest acquisitions in the agent-lab wars. Along the way, they picked up developer-philosopher swyx, whose “Code AGI” thesis reframed Devin not as a product demo but as a trillion-dollar inevitability. And through him, they’ve more or less claimed the whole AI engineer movement. Smart.

We’re starting this deep dive with a quote buried in swyx’s reflection on why he joined Cognition AI. It’s the perfect lens to see a company that went from hacker house to a $10.2 billion valuation in less than two years – and then to $26 billion by May 2026. Now, it’s time to substantiate it. Below is the full map – origins, philosophy and culture, product, money, Windsurf drama, critics, risks, TAM, and why swyx is the narrative hinge that makes the whole thing click. Curious to learn more? Let’s go.

In today’s episode:

Cognition is what happens when childhood prodigies refuse to slow down as adults. The same intensity that once drove them to solve math problems under time pressure now drives them to build an AI workforce measured in compute units.

There’s a video of Scott Wu in a contest where, before you can even read the body of the problem, he already knows the answer. His voice sounds skeptical – he’s sure he’s right, but wonders why the answer isn’t obvious to everyone else.

Cognition AI was officially founded in November 2023 (according to the founders’ Linkedin pages) by Scott Wu (born 1997, 28 years old), Steven Hao (graduated MIT in 2014, so likely 28–29), and Walden Yan (finished high school in 2020, so about 22–23). Each brought an unusually decorated background in competitive programming:

From the start they felt two shifts coming. First, reinforcement learning would push beyond imitation models like the original ChatGPT. Instead of echoing the internet, high-compute RL could try, fail, get feedback, and improve. Second, products would move from text completion to true agents – systems that can reason, plan, and act over multiple steps.

Code was the obvious domain. Not only were they good at it, but code itself provides its own feedback loop: run it, check it, learn.

When they started, it wasn’t really a company – more like a project, a hackathon. Around Thanksgiving 2023, they rented an Airbnb, pulled in friends, and hacked on ideas. The first build targeted competitive programming problems, using agentic loops to improve test performance.

The form kept shifting. At first, each built their own agent – DevSteven, DevWalden, and so on – which only delivered finished code. The breakthrough came around Christmas 2023, when a prototype fixed a blocked server on its own. That was a revelation: an agent could be a partner – spot errors, recover, and finish the job.

Devin – all the Devs merging into one – was born.

In March 2024, Cognition emerged from stealth. A demo video showed Devin planning, browsing documentation, writing and testing code, and finally opening a GitHub pull request.

The claim: Devin had just completed a real task without hand-holding. The punchline: Devin is “the first AI software engineer” – lifting the phrase swyx had planted in his June 30, 2023 Latent Space post, The Rise of the AI Engineer.

The launch went viral, with millions of views and heated debate across developer networks: was it really the “first AI engineer”, are software developers cooked?!

We explored this question directly with Eli Hooten, CTO at Codecov

On April 6, swyx (again) jumped in: “ignore everyone who hasn’t used it.” In a long X thread with follow-ups, he described how he’d shipped Swift code to the App Store, ported React to Svelte, and wrangled Elixir games with Devin as his “five-engineer team.” His verdict: slow, pricey, bad at design, clumsy with Git – but still the best coding agent he had seen.

A week later, Gergely Orosz, author of the popular Substack Pragmatic Engineer, amplified a YouTube breakdown by the channel Internet of Bugs: Devin is all about staged demo, phantom files, trivial shell commands, work that could have been done by reading the README. Verdict: hype dressed as substance.

The split was instant – AI believers calling Devin a glimpse of the future, hardcore devs calling it smoke and mirrors. Cognition’s team jumped in to clarify:

So from week one, the question was never whether Devin got attention. It was whether it could survive a code review from the very humans it promised to replace. And: was it as much a prodigy as its founders – and could it actually compete at their level? To answer these questions, we need to explain why it matters that the founders of Cognition are prodigies.

In his show A Cheeky Pint, John Collison (Stripe’s co-founder) asked Scott Wu whether the return of very young founders is a biomarker of industry takeoff. PCs had Dell. Social had Zuckerberg. That’s an interesting point. But what stands out about Wu, Yan, and parts of their cohort – including Alexander Wang from Scale AI and Demi Guo from Pika, whom Scott mentions in that interview – is how many came through math and programming competitions. It’s not simply the return of young founders; it’s a wave of people trained unusually early to think under pressure, measure performance, and recover from mistakes.

For much of the 20th century, prodigies collided with institutions that dulled their edge. Genius mathematician Srinivasa Ramanujan dazzled Cambridge but strained under academic formalism. Alexander Grothendieck revolutionized algebraic geometry yet abandoned the academy, alienated by its politics. Even those who flourished, like John von Neumann, did so largely because wartime America bent its research ecosystem to accommodate brilliance.

The mismatch was structural: hierarchies prized seniority, committees slowed radical ideas, feedback loops rewarded tenure instead of outcomes. For minds trained under clocks and scoreboards, it was suffocating.

Startups flipped the environment. AI craziness accelerated it. Finally, these child prodigies found a world that ran at their pace – fast, unforgiving, and measurable. Flat ownership gave agency from day one. Risk-seeking cultures rewarded speed. Feedback came from launches and ARR, not citations. Suddenly, speed, stamina, and precision – once liabilities – became the perfect fit.

That’s why Wu, Hao, and Yan could turn Olympiad reflexes into a $10.2 billion company by September 2025, less than two years after Cognition was founded – and then into a $26 billion company by May 2026. In another era, they might have been absorbed into Bell Labs or IBM, publishing memos that never saw daylight. In this one, they can be independent, launching Cognition – compressing cycles, chasing measurable outcomes, building visible wins.

Youth may make an extreme startup schedule easier to tolerate, but it does not create unlimited time. The more important point is that Cognition turned intensity into an explicit operating principle – and then asked the Windsurf team to join it →

Contest training can make long stretches of deep concentration feel familiar. Inside that kind of flow state, the emotional world narrows around a problem and hours can feel like minutes. To outsiders this can look obsessive; to the person inside it, the experience may feel calming or exhilarating.

Prodigy-founders shape the culture of the company in a distinct way. Contest programming demands brilliance, but it also requires high endurance under pressure. Competitors face five-hour windows, three problems, and one unyielding clock. The game is about precision, planning, and error recovery – skills that translate directly into the way Cognition builds software. The founders are intensely competitive, and they cannot imagine anyone around them not being the same. In a way, it’s their blinders.

Supporters praised this clarity. Critics called it unsustainable. Either way, it revealed Cognition’s ethos: make intensity explicit.

Here we arrive at the editorial heart of the Cognition story. Devin is not just an agent but a prosthetic for the overachieving brain.

For prodigies, the constraint has always been physical. One brain, two hands, limited hours. Olympiad competitors can only solve so many problems in five hours, no matter their skill. Startups can compress cycles, but humans still fatigue.

Devin changes that equation. It doesn’t replace ingenuity – it multiplies it, giving overachievers additional brain muscle with digital hands. One engineer can now orchestrate many Devins in parallel, turning individual brilliance into organizational output. Where once one prodigy wrote one solution, now fleets of agents can explore alternatives, attempt migrations, or grind through repetitive tasks in bulk.

This is what makes Cognition different from ordinary startups. It is not simply a company employing prodigies; it is a company that has encoded prodigy habits into its product. Devin carries forward the contest mentality: precise planning, measurable outcomes, relentless retries. It lets others participate in that rhythm, whether or not they trained for years in Olympiad halls.

In this way, Cognition is both a company and a new institutional form: a place where prodigy culture does not clash with structure but becomes the structure itself.

Devin is less chatbot than teammate. Each instance runs in its own cloud dev box with a Linux shell, code editor, browser, and toolchain.

Give it a task – through Slack, Linear, Jira, or the web interface – and it will sketch a plan, execute in its sandbox (installing dependencies, editing files, running tests, retrying after errors), then deliver results as a pull request with commits tied back to the ticket.

The basic loop is still intact in 2026, but the product around it has widened. Interactive Planning, Devin Search, Devin Wiki, and MultiDevin established the original pattern. Devin 2.2 added faster startup, computer-use testing, and more self-verification. Devin Review and the Devin CLI brought the agent closer to pull requests and terminals; scheduling made recurring work possible; Devin Desktop became the command center for supervising many agents at once. Underneath it, Cognition kept updating its own coding models, reaching SWE-1.7 in July 2026.

Where it shines today

Teams report the clearest value on well-scoped tickets that are laborious for humans and easy to verify: version upgrades; dependency and API migrations; large-scale linting and layout fixes; flaky-test hunts; doc and type improvements; on-call triage that starts with reading logs and reproducing bugs. These jobs often make up a large fraction of engineering hours, and they respond well to planning + execution + human review.

Where it still needs shaping

Open-ended product work, ambiguous architecture changes, and cross-team dependencies still require careful scoping. A January 2025 Register review put Devin’s success rate on unscoped tasks at about 15%. That was an early test of a much earlier product, not a measurement of Devin in 2026, but its lesson still holds: an agent cannot rescue a task whose goal, constraints, and definition of done are unclear.

Cognition’s answer has been to tighten the loop between autonomy and review: approve the plan, show code-cited context, isolate the work, test the result, let Devin inspect its own output, and keep a human responsible for what gets merged. The agent can now run for longer and operate across more surfaces, but reliability still depends on how well the task is framed and how easy the result is to verify.

Cognition began with a flat $500-per-team monthly plan. In April 2025, it added a $20 pay-as-you-go entry using ACUs – Agent Compute Units – as the billing unit. At launch, TechCrunch estimated that $20 bought roughly nine ACUs, or a couple of hours of active Devin work, and warned that costs could climb quickly on large repositories. By April 2026, the ladder had changed again: Free, Pro at $20 per month, Max at $200 per month, Teams with an $80 monthly minimum and $40 per full seat, and custom Enterprise plans. Self-serve customers now use plan quota plus on-demand credits; enterprise contracts can still use ACUs. The packaging changed, but the logic did not: Cognition wants the bill to follow work performed, not merely seats assigned.

Cognition’s early customers were startups and tech-forward enterprises. By 2025, the roster included Ramp, Nubank, OpenSea, Lumos, and Curai Health. By 2026, Cognition was pointing to a broader enterprise group that included Citi, Mercedes-Benz, Goldman Sachs, Elevance Health, Dell, Santander, and work with the U.S. government. Those names show distribution; the operational results below are the more useful evidence – and, where they come from Cognition or a customer case study, should be read as company-reported.

Revenue tracked those wins. ARR grew from $1 million in September 2024 to $73 million in June 2025, while cumulative net burn remained below $20 million. By May 2026, Cognition reported $492 million in run-rate revenue and more than tenfold growth in enterprise usage since the start of the year. This is not the same metric as ARR, but it shows how quickly usage expanded after Devin, Windsurf, and enterprise distribution began reinforcing one another.

Cognition’s funding story shows how quickly coding agents became one of the hottest AI markets. The company raised $21M in March 2024 and $175M the following month at a $2B valuation. Private-market records later reported a roughly $145M Series B-1 in March 2025 at about $4B. In September 2025, Cognition raised $400M at a $10.2B post-money valuation; at the time, it said ARR had grown from $1M in September 2024 to $73M in June 2025 while cumulative net burn remained below $20M. Then, in May 2026, Cognition announced more than $1B in new funding at a $26B valuation, led by Lux Capital, General Catalyst, and 8VC.

The story may move again. On August 12, Bloomberg reported that Cognition was in early talks for another round at a valuation of at least $40B, with its annualized revenue run rate approaching $1B. Those talks were not a completed financing. As of September 1, 2026, $26B remains Cognition’s latest confirmed valuation; the $40B figure should be treated as a reported target, not a closed round.

These numbers change how you look at Cognition. It is no longer only a fast-growing AI startup with a viral product and ambitious founders. Its revenue trajectory suggests that enterprises are paying for coding agents at meaningful scale – and that investors expect agents to become a new layer of software labor.

The key question is what sits inside the run-rate figure. Cognition does not publicly break it down across Devin usage, Windsurf subscriptions, enterprise contracts, or ACU-based agent compute. That mix matters because the valuation depends not only on rapid growth, but on whether customers keep paying for durable, repeatable engineering output.

This is why the $26B valuation is both impressive and risky. On one hand, AI coding is one of the clearest places where agents can prove real economic value: code can be tested, pull requests can be reviewed, bugs can be reproduced, and productivity gains can be measured. On the other hand, autonomous software engineering still faces hard limits: long-horizon reliability, unclear task scoping, compute costs, developer trust, enterprise security, and competition from GitHub Copilot, Claude Code, Cursor, OpenAI Codex, Replit, and other agentic coding systems.

The valuation also reframes the Windsurf acquisition. The deal gave Cognition IDE distribution, developer mindshare, enterprise relationships, and a front-end surface for agentic coding. In 2026, that looks less like a product add-on and more like a platform strategy: Devin can run work in the background, while Windsurf gives developers a place to supervise, review, and coordinate it.

This is the core of the Cognition bet: software engineering may not be replaced by one agent, but it may be reorganized around fleets of coding agents. And Cognition is selling a new unit of software labor.

is hard to estimate but one thing is clear – it’s huge. According to Evans data, there was ~27 million of developers globally (2024), according to Slash data there is 47.2 million developers globally (quite a difference, huh). Average fully-loaded cost per developer: $100k–$300k per year depending on geography.

If we take the smaller headcount estimate (27M) and the low end of cost ($100k), that implies current global developer labor spend of $2.7T per year. If AI coding agents capture just 5–10% of that value in the near term, the serviceable market is $135–$270B per year. Near term.

On the higher end (47.2M developers at $300k), the same 5–10% capture yields $708B–$1.416T per year. Also, near term.

A seat-pricing cross-check – say $300–$600 per developer per month with 50–80% adoption – gives $49–$272B per year in direct revenue. That lines up with the lower end of the value-capture view.

Here you need to stop and imagine being Scott Wu: he doesn’t have to work through all the calculations – he simply sees the numbers outlined in front of him.

And this doesn’t even count the expansion of software demand: the universe of products that were never economical to build when developer labor was the bottleneck.

In July 2025, OpenAI came close to buying Windsurf, an agentic IDE startup, for $3 billion. The deal collapsed at the last minute. Within days, Google DeepMind hired Windsurf’s leadership and licensed parts of its technology for a reported $2.4 billion. Roughly 250 employees remained with the business Cognition would acquire.

Cognition moved fast. Over the weekend, Wu and his team struck a deal to acquire Windsurf’s product, brand, and customer contracts. By Monday morning, it was signed. In a market known for hard landings, Cognition’s offer was unusual: every employee participated financially, and vesting was accelerated, even for those still under the one-year cliff.

Two elements stood out. First, the structure – a rare gesture of respect in a week when many felt whiplash. Second, the combination itself: Devin’s autonomous engine joined to Windsurf’s IDE surface and go-to-market machine. Together, they offered enterprises a continuous workflow: plan in the IDE, hand subtasks to Devins, leave judgment calls to humans, and merge in one environment. Press accounts and Windsurf’s interim CEO later described the weekend as chaotic, emotional – and, in the end, a relief.

The strategic prize was clear. Windsurf brought distribution, a strong sales and customer-success team, and a reported $82 million in annual recurring revenue, doubling quarter over quarter. For Cognition – efficient but not yet scaled – it was transformative. In August 2025, the combined team shipped Wave 12. By July 2026, Cognition said it had grown from 44 to 350 people and surpassed $500 million in revenue run rate. More importantly, it began unifying the two sides around Devin Desktop: Windsurf as the place where developers work, and Devin as the agent layer that can run, review, and coordinate work in the background. The figures are company-reported, but the product direction is now much clearer than it was during the rescue weekend.

You might have noticed that Shawn “swyx” Wang appears quite a few times in this article. In September 2025, the same day Cognition announced the $400M round, he tweeted that he was joining the company. It felt like a long-planned move by Scott Wu. If you are competing at the edge, you cannot claim “the first AI software engineer” without also bringing in the person who defined the category of AI engineers.

Devin is not universally loved by software developers, and that is exactly where swyx matters. He has been shifting tides, persuading programmers to see themselves as AI engineers. He is loved in the community, trusted by practitioners, and influential in shaping narratives. These are all the things Cognition lacked.

With this “Code AGI”, swyx reframed Devin not as a product demo but as a trillion-dollar inevitability. Three roles he will play (as we see it) inside Cognition:

The intellectual frame – His phrase above reframes Devin+Windsurf from “ambitious IDE” to the shortest path toward AGI-scale returns. Code is verifiable and recursive; agent teams can dogfood their own tools; the loop is tight. Cognition had the product and some confusion with brands (Devin, Windsurf, Cognition) – he gave it the explanatory scaffolding.

The developer bridge – Through Latent Space and the AI Engineer community, swyx connects to practitioners who adopt tools early. That audience trusts him to separate demo theater from durable capability. His threads turned Cognition’s private cadence into public playbooks. Even if he tries to keep it separate. It’s the reputation that is working.

The translator – Inside a company known for Olympiad-level intensity, he explains the market split in plain terms: model labs train frontier models; agent labs adapt them to domains, stitch workflows, and sell outcomes. Cognition belongs squarely in the latter camp.

Different players are racing toward the same ground from different angles:

But this isn’t winner-take-all terrain. GitHub Copilot, Cursor, Claude Code, Codex, and Devin can all perform agentic work now; the clean old distinction between “assistant” and “autonomous agent” has blurred. Most developers will blend an interactive coding surface with agents running off-screen. Cognition’s bet is that Devin + Windsurf becomes the most natural pairing for enterprise backlog compression, modernization, and work that can be delegated with a clear definition of done.

What is Cognition for its founders? At one level, it’s a puzzle that never ends – the same thrill as an Olympiad round, only played at company scale. At another, it’s a contest – with GitHub, Anthropic, Replit, Cursor, Warp, and the rest of the agentic stack circling the same territory. For Wu, Hao, and Yan, the attraction is obvious: a stage where speed, intensity, and precision aren’t eccentric traits but competitive edges. They hacked themselves into a unicorn in under two years because the company became their next scoreboard.

What should we watch for? Three things. First, survivability – how far Devin agents can run before stalling, and whether those guardrails truly bend cost curves. Second, culture – whether Olympiad-level stamina can sustain hundreds of employees, or if it burns them out. Third, distribution – Windsurf integration and swyx’s framing give Cognition reach, but developers are not easy converts. They need trust, not just awe.

What are the risks? Compute costs remain heavy, long-horizon tasks brittle, and cultural intensity may narrow recruiting. Competitors with broader distribution can catch up quickly. The magic is not only in building Devins but in making them useful at scale.

And what is the opportunity? Think of it as a math paradise: multiplication tables where every line ends in a trillion. The prize is the global software labor force – tens of millions of developers, with annual spend counted in the trillions. To someone like Scott Wu, the numbers line up as neatly as a contest scoreboard: inputs, outputs, capture rates. Even a single-digit percentage of that market would justify the intensity that defines Cognition.

The same agentic principles Devin applies to software are now being applied to scientific discovery – from drug discovery to theorem proving. See 12 AI co-scientists of 2026.

That mix – mathematical clarity, a culture of relentlessness, and a TAM so large it feels almost comic – is why Cognition matters. Whether they can cash it in is still an open problem. But for now, just as Wu once answered faster than anyone could finish reading the question, Cognition has written its solution on the scoreboard first.

Fascinating.

Cognition AI is an AI startup building autonomous software engineering agents. Its main product is Devin, an AI coding agent that can plan engineering tasks, inspect repositories, edit files, run commands, debug issues, and submit pull requests. Cognition became known for positioning Devin as the first AI software engineer.

Scott Wu is the CEO and co-founder of Cognition AI. He co-founded the company with Steven Hao and Walden Yan, and the founding team is known for its background in competitive programming, including International Olympiad in Informatics medals.

Cognition AI has not publicly disclosed full profitability. Reports point to fast revenue growth and strong enterprise adoption, but run-rate revenue is not the same as profit. Without official financial statements, the safest answer is that Cognition appears to be growing quickly, while its profitability remains unconfirmed.

No. Cognition AI is a private company. It has raised venture funding from private investors and is not listed on a public stock exchange.

Devin AI is Cognition’s autonomous AI software engineer. It can take a software task, create a plan, work inside a development environment, edit code, run tests, debug problems, and produce a pull request. Devin is designed for agentic software work, not just autocomplete.

Cognition AI is valued highly because investors see autonomous coding agents as a large software-labor market. If agents like Devin can reliably complete engineering work, they could capture part of the global developer productivity stack, from code maintenance and migrations to testing, debugging, and enterprise backlog execution.

The difference is now about workflow, not a clean split between assistants and agents. GitHub Copilot and Cursor can both run longer agentic tasks, while Devin is built around delegating scoped engineering work to agents that operate in their own environments. Cognition’s advantage depends on orchestration, enterprise controls, and the Devin + Windsurf workflow – not simply on being more autonomous.