Signals Inbox·September 2, 2026·AI Agents

Is Cognition really worth $47B today?

Cognition can plausibly be worth $47 billion today, but the price is still ahead of the business: revenue growth has kept pace with the valuation jump, while margins, compute spending and direct competition from Claude Code and Codex leave very little room for an ordinary slowdown.

Q1Did Cognition really jump from $26B to $47B in three months?

Yes, Cognition is currently on track to go from a $26 billion valuation to about $47 billion in roughly three months, although the new financing is still being finalized.

Bloomberg reports that Cognition is raising around $1 billion at the new valuation, with the final amount potentially going higher. The striking detail is demand: investors reportedly expressed nearly $10 billion of interest in the round. The investors in this latest financing have not yet been disclosed, and the terms can still move before closing.

The speed is easier to see when we zoom out. Founders Fund valued Cognition at $2 billion in spring 2024, only a month after an earlier financing at roughly $350 million. Cognition reached about $4 billion the following year, $10.2 billion later in 2025, then $26 billion after raising more than $1 billion in May. The latest proposed price would take it to $47 billion.

Cognition was founded in 2023 and launched Devin publicly in 2024. A company that was worth $2 billion shortly after launching its main product could therefore be worth almost $50 billion barely two and a half years later. AI has produced several extraordinary valuation jumps, but this is still an unusually compressed trajectory.

Cognition valuation trajectory

Q2Is Cognition actually making enough revenue for a $47B valuation?

Cognition is already producing serious revenue, but today’s $47 billion price still represents roughly 52 times its latest reported annualized revenue.

Bloomberg now puts Cognition above $900 million of annualized revenue. The Information independently reported roughly $75 million of monthly revenue, which annualizes to about $900 million. Having two well-sourced reports converge around the same figure makes it more credible, but we should still be precise about what it means.

This is run-rate revenue rather than twelve months of audited revenue. Cognition is annualizing its current pace during a period of extreme growth. A company doing $75 million in one recent month can legitimately say it is running at $900 million annually, but that does not mean it collected $900 million over the previous year.

We do have a harder reference point. When Cognition announced its previous financing, the company itself disclosed $492 million of run-rate revenue. That gives us a company-confirmed baseline only a few months before the newer figure appeared.

The current revenue number is credible enough to use for valuation work, with one important discount: we have much more confidence in the direction and order of magnitude than we would in audited public-company revenue.

Q3Did Cognition’s valuation rise faster than its business?

Surprisingly, no: Cognition’s latest valuation jump has almost perfectly matched its reported revenue growth.

At the previous financing, Cognition was valued at $26 billion while reporting $492 million of annualized revenue. That worked out to about 52.8 times revenue.

The new proposed valuation is 80.8% higher. Meanwhile, annualized revenue has reportedly risen about 83% from the company-confirmed May level. Using the latest numbers gives us a multiple of roughly 52.2 times.

So investors are paying roughly the same amount for each dollar of current revenue as they were before. The headline valuation looks dramatically more aggressive because $47 billion is dramatically larger than $26 billion, but the underlying pricing has barely changed.

That does not suddenly make 52 times revenue cheap. The useful part is that Cognition has so far grown fast enough to keep up with one of the fastest private-market valuation increases in software.

Q4Did Cognition buy most of its growth with Windsurf?

No, although buying Windsurf gave Cognition a very meaningful head start and makes the raw growth numbers look better than purely organic growth would.

Before the acquisition, Devin went from $1 million of ARR in September 2024 to $73 million by June 2025. Cognition then acquired Windsurf, which had $82 million of ARR, more than 350 enterprise customers and hundreds of thousands of daily active users.

A rough combined starting point was therefore around $155 million. Cognition also said customer overlap between the two businesses was below 5%, which made the deal particularly attractive: Windsurf brought a largely separate customer base rather than simply duplicating Devin accounts.

The combination began growing almost immediately. Cognition said combined enterprise ARR increased more than 30% during the first seven weeks after the acquisition. The company also gained something harder to quantify than revenue: Windsurf had an established IDE product and enterprise sales motion, while Devin was built around delegating complete engineering tasks to autonomous agents.

So we should absolutely adjust our interpretation of Cognition’s growth for the acquisition. Still, going from a roughly $155 million combined base to the level the business has reached these days requires far more than simply adding Windsurf’s existing revenue.

Q5Can Cognition really keep growing this fast now?

Cognition can probably keep growing very quickly for a while, but the next leg is much harder because the company is already operating at a huge scale.

The Information says Cognition has more than tripled its annualized revenue since the beginning of the year. Cognition had already reported enterprise usage growing more than tenfold over the first part of the year, and CEO Scott Wu said enterprise usage had been rising around 50% month over month for six consecutive months.

Those usage figures are company-reported, so we give them less weight than cash revenue. Still, usage and revenue are moving in the same direction, and the trajectory has lasted long enough to look more substantial than a temporary launch spike.

Now the base effect becomes brutal. Management reportedly expects annualized revenue to exceed $1.5 billion by year-end and reach $4 billion to $5 billion next year. Getting to the first target requires another large jump. Reaching the second would require Cognition to add several billion dollars of annualized revenue after already becoming one of the biggest AI application companies.

We have enough evidence to say Cognition has already achieved hypergrowth. Whether a business this large can continue compounding at anything close to that pace remains the main financial question.

Q6Is Cognition more expensive than Cursor, Lovable and Replit?

Yes, Cognition currently commands a very large premium over the other breakout AI coding companies.

Cursor is the hardest comparison to ignore. Before SpaceX completed its $60 billion acquisition, Forbes reported that Cursor had passed $4 billion of annualized revenue, up from $2 billion in February and $3 billion in late April. That puts the transaction at roughly 15 times run-rate revenue.

Lovable recently raised $400 million at a $13.3 billion valuation after telling TechCrunch it had reached $500 million of annualized revenue. That works out to about 27 times revenue. Lovable expected the run rate to approach $600 million shortly afterward, which would move the multiple closer to 22 times if achieved.

Replit was valued at $9 billion earlier this year. TechCrunch reported just last week that its current run rate is tracking toward $1 billion annually. Using that level would put Replit near 9 times revenue.

The businesses are different enough that we should not rank them mechanically. Cursor is now inside SpaceX, Lovable serves many nontechnical builders, Replit increasingly offers an entire application-building stack, and Cognition leans heavily into enterprise engineering. Even with those differences, the gap is hard to dismiss. Investors are currently paying far more for each dollar of Cognition revenue than for revenue at the closest successful private peers.

AI coding peers by valuation and revenue

Q7How crazy is Cognition’s valuation next to public software companies?

Cognition currently trades at roughly five to seven times the sales multiple of some of the best-known public enterprise software companies.

GitLab trades around 7.8 times trailing sales these days. Atlassian is around 7.9 times. ServiceNow, one of the public market’s premium enterprise software franchises, sits around 10.6 times.

Cognition deserves a higher number. GitLab’s latest quarterly revenue grew 21% year over year, Atlassian grew quarterly revenue 28%, and ServiceNow grew 24%. Cognition has recently been growing several times faster.

What becomes difficult is deciding how large that premium should be. Paying twice ServiceNow’s multiple for a company growing dramatically faster would be easy to understand. Paying around five times as much asks us to assume that Cognition will preserve an unusually large growth gap for years while also improving the economics of serving every customer.

Public companies are imperfect comparisons because they have already gone through much of their growth curve. They still give us a useful reality check. A valuation above 50 times current revenue sits far outside normal premium software pricing today.

Public software valuation comparison

Q8Are Cognition’s margins good enough for a $47B company?

No, Cognition’s current margins are much too low for us to treat its revenue like normal high-margin SaaS revenue.

The Information recently reported that Cognition’s enterprise gross margin is close to 50%. That means roughly half of every enterprise revenue dollar is being consumed before sales, R&D and other operating expenses.

Traditional software looks very different. GitLab reported an 84% GAAP gross margin in its latest quarterly results. ServiceNow reported a 73.5% GAAP subscription gross margin. Cognition has much heavier inference and computing costs every time Devin performs work.

At a 50% margin, Cognition’s current annualized revenue base produces roughly half as much gross profit as an equivalent amount of high-margin SaaS revenue would. So the 52-times sales multiple actually understates how demanding the valuation is.

Cognition is working directly on this problem. The company has started training specialized coding models, and its Devin Fusion system routes jobs among different models to improve price-performance. Cognition says Fusion can reach near-frontier performance at dramatically lower model cost on its own evaluations. SWE-1.7, another internal model, also reached results close to several leading frontier systems on Cognition’s benchmarks.

We give those benchmark claims limited weight because Cognition designed the evaluations itself. The economic direction is more important: the company clearly understands that paying frontier-model prices forever would leave it with a weaker business than the valuation demands.

Q9Is Cognition burning too much cash?

Cognition’s current cash burn is aggressive, but most of the problem appears to come from a deliberate bet on owning more of its AI infrastructure and models.

The Information reports that Cognition could burn around $800 million of cash this year, including roughly $200 million in the second quarter. The company has leased large clusters of Nvidia servers that both run Devin and support development of Cognition’s own models, with the infrastructure reportedly costing hundreds of millions of dollars annually.

That is a dramatic reversal from Cognition’s early history. In 2025, the company said its cumulative net burn since founding was below $20 million. Cognition went from being remarkably capital-efficient to spending at a scale normally associated with AI model companies.

The twist is that, according to the same reporting, Cognition would be close to free-cash-flow breakeven if we excluded spending on developing its proprietary coding models. The core commercial operation looks much healthier than the $800 million headline suggests.

Investors still have to count the model spending. Cognition is funding it because relying entirely on Anthropic, OpenAI and other model providers could hurt both margins and strategic independence. If proprietary models eventually lower serving costs substantially, the spending could look smart. If Cognition has to keep spending hundreds of millions just to remain competitive, the valuation becomes much harder to defend.

Q10Are enterprises actually paying real money for Devin?

Yes, Cognition has already moved well beyond small AI experiments and into meaningful enterprise contracts.

The Information reports that around 50 companies are on pace to spend more than $1 million a year with Cognition. That is much more useful than a long list of logos because seven-figure annual spending usually means a product has moved into actual operating budgets.

Cognition publicly names customers including Citi, Mercedes-Benz, Goldman Sachs, Dell, Santander, BNY and major U.S. government organizations. Its government business already reaches the U.S. Army, U.S. Navy and NASA’s Jet Propulsion Laboratory. Cognition has also built versions of its products for regulated and defense environments.

Goldman Sachs is particularly interesting right now. Business Insider recently reported that all 12,000 Goldman developers have access to advanced AI engineering tools including Devin and Claude. Goldman’s problem has already moved beyond simply giving engineers access to agents: the bank is now building internal instruction sets so those agents understand Goldman’s engineering standards and institutional knowledge.

There is some concentration risk behind the enterprise story. The Information reports that one unnamed customer represents about 10% of Cognition’s business. Cognition therefore has real enterprise depth, but the revenue base is not yet diversified enough for us to ignore individual large accounts.

Q11Does Devin actually save companies enough money to justify those contracts?

Devin has produced enough concrete enterprise results for us to believe customers can get very large returns on the right engineering work.

Mercedes-Benz provides one of the cleanest examples. During an initial four-week pilot, Devin analyzed more than 200,000 lines of COBOL and cut the estimated modernization timeline from eight months to eight days. Mercedes subsequently expanded the partnership across global engineering teams.

Cognition has also highlighted work at Itaú, where Devin automatically remediated 70% of security vulnerabilities identified through static analysis, increased test coverage and helped document hundreds of thousands of repositories. These figures come from Cognition, so we should treat them as vendor case studies rather than independent experiments.

More interestingly, Cognition has put money behind the claim. The company recently introduced an enterprise productivity guarantee: if Devin produces less estimated engineering value than the customer paid for, Cognition will provide usage credits, with commitments reaching as high as $10 million.

That guarantee has obvious measurement limitations because Cognition itself estimates the equivalent engineering work. Yet a vendor burning expensive inference on every task has little incentive to promise huge amounts of free usage if customers repeatedly fail to see value.

The current evidence is strongest around boring, expensive engineering work: migrations, security remediation, testing, maintenance and large legacy codebases. Those jobs may be less exciting than asking Devin to build an entire company from scratch, but enterprises already spend enormous amounts of money on them.

Q12Can Claude Code and Codex crush Cognition?

Yes, Anthropic and OpenAI are probably Cognition’s biggest threat today because developer behavior is moving incredibly fast toward their own coding products.

JetBrains recently surveyed more than 15,000 professional developers worldwide. Claude Code was being used at work by 39% of developers, up from 18% near the beginning of the year. OpenAI Codex jumped from 3% to 16% over roughly the same period. Cursor fell from 18% to 12%.

Those are huge movements within a few months. They show how little permanent loyalty exists in AI coding right now. When a new model or agent gets clearly better, developers can move quickly.

The broader adoption numbers make the competitive prize obvious. JetBrains found that 90% of professional developers were now using AI coding agents at work at least weekly, with 68% using them daily. Anthropic and OpenAI can therefore attack Cognition inside a market that is already mainstream rather than waiting for enterprise adoption to arrive.

A recent Cursor episode makes the supplier risk even more visible. OpenAI is ending its model-supply agreement with Cursor following the company’s acquisition by SpaceX. Cursor only routes a minority of its traffic through OpenAI, but the episode shows how much strategic power model suppliers retain over applications built on top of them.

Cognition has anticipated this. Devin can use different underlying models, and Cognition is developing its own specialized models as another option. That reduces dependence, although it does not remove the threat of Anthropic or OpenAI selling competing coding products directly to Cognition’s customers.

Q13What can Cognition do that competitors cannot easily copy?

Cognition’s best defense currently comes from enterprise workflow, distribution and orchestration rather than from one magical piece of technology that competitors cannot reproduce.

The Windsurf acquisition gave Cognition two surfaces inside engineering teams. Devin handles delegated asynchronous work, while Devin Desktop, built from Windsurf, gives developers an interactive environment when they want direct control. Owning both lets Cognition move between human-led coding and agent-led coding without asking customers to leave its ecosystem.

The enterprise plumbing is harder to copy. Large customers have to connect agents to repositories, internal tools, security systems, review processes, permissions and company-specific engineering rules. Goldman Sachs is now explicitly teaching its agents internal standards. The more context and workflow a company builds around Cognition, the more annoying switching becomes.

Government work adds certifications and deployment requirements that a new coding startup cannot reproduce overnight. Cognition already offers products for highly regulated environments and has deployments inside military and civilian agencies.

None of those advantages guarantees dominance. Claude Code can become deeply embedded in workflows. OpenAI can build enterprise orchestration. GitHub, Microsoft, Google and JetBrains already control enormous distribution surfaces.

Cognition’s position is defensible enough to support a serious company. We do not yet see evidence of a moat so strong that a 50-times-revenue valuation should feel safe.

Q14Is the AI coding market big enough for a $47B Cognition?

Yes, AI coding is growing fast enough to produce companies worth tens of billions, and fresh spending data makes this one of the stronger parts of Cognition’s valuation case.

Gartner currently estimates the enterprise AI coding-agent market at roughly $9.8 billion to $11 billion annualized. Menlo Ventures estimated that enterprise spending on generative-AI coding tools jumped from about $550 million in 2024 to $4 billion in 2025. That is more than sevenfold growth in a year.

The latest adoption research points in the same direction. McKinsey found that 31% of respondents at companies with more than $1 billion of annual revenue are already scaling software coding agents. Nearly a third of respondents said their companies had decided against buying at least one software product or feature because agentic coding tools allowed them to build it internally.

The market is broader than developer-tool budgets. Coding agents can also attack outsourced engineering work, migration projects, IT modernization, consulting expenses and even pieces of existing SaaS spending.

There is already enough money moving into the category to support several large winners. Cognition’s challenge is that the market is attracting Anthropic, OpenAI, Google, Microsoft, Cursor, Replit, Lovable and dozens of smaller companies at the same time. A huge market solves the demand question much more convincingly than it solves the winner question.

Q15How much revenue does Cognition need before $47B looks normal?

Cognition probably needs to reach roughly $3 billion to $5 billion of annual revenue before today’s valuation starts looking normal rather than exceptional.

At 30 times revenue, a $47 billion valuation requires about $1.6 billion. That would still be a very expensive software multiple. At 20 times, Cognition needs roughly $2.35 billion. At 15 times, it needs slightly more than $3.1 billion.

The really interesting line is 10 times. Cognition would need about $4.7 billion of annual revenue for the current valuation to equal a premium public-software multiple.

As seen above, management reportedly expects $4 billion to $5 billion of annualized revenue next year. That forecast therefore lands almost exactly where the current valuation becomes much easier to explain mathematically.

We should not mistake a management forecast for achieved revenue. It does, however, give us a very clear benchmark. If Cognition gets close to that target while maintaining strong growth, investors who paid $47 billion may look far less aggressive than they do today. If revenue stalls well below it, the valuation has very little protection.

Revenue needed to support a $47B valuation

Q16What has to go right for Cognition to be worth $47B?

Cognition can grow into $47 billion, but the company now needs several unusually good things to happen at the same time.

Revenue has to keep compounding from an already massive base. A move into several billion dollars of annualized revenue would quickly compress the sales multiple and solve the most obvious valuation problem.

Margins also need to improve. These days, Cognition keeps only about half of enterprise revenue after direct costs, according to The Information. Training specialized models, routing workloads more intelligently and using cheaper models for easier tasks could move that number higher. Investors will eventually want proof that those savings are actually reaching Cognition rather than being handed straight back to customers through lower prices.

Enterprise usage then has to become sticky. Large customers such as Mercedes, Citi and Goldman Sachs cannot remain impressive pilot stories; Devin needs to become part of normal engineering operations and keep expanding inside those organizations.

Finally, Cognition has to stay useful even when Claude, Codex and future coding agents become much better. The company does not necessarily need the best underlying model. It needs customers to believe Cognition is the best place to manage AI engineering work across whichever models are best.

None of these assumptions looks absurd on its own. The difficulty comes from needing most of them to work together.

Q17What could make Cognition’s valuation fall apart?

Cognition does not need to fail for $47 billion to look too high; a slowdown from extraordinary growth to merely good growth could be enough.

Suppose Cognition eventually reaches $1.5 billion to $2 billion of revenue but growth then normalizes sharply. A 15-times revenue multiple would value that business at roughly $22.5 billion to $30 billion. Even those would be generous software multiples, yet both sit well below the current proposed valuation.

Weak margin improvement would create another problem. As pointed out above, Cognition’s enterprise gross margin is around 50% today. If autonomous coding continues consuming expensive inference and customers resist higher prices, billions of dollars of future revenue may be worth less than investors currently assume.

Competition could make both problems arrive together. JetBrains’ latest survey shows Claude Code and Codex gaining developer adoption extremely quickly. Stronger model-provider products could slow Cognition’s growth while simultaneously forcing it to cut prices.

Customer concentration adds another pressure point. One large customer reportedly contributes around a tenth of the business. Losing that customer would hurt much more than losing an ordinary enterprise account at a mature SaaS company.

The bad outcome is pretty easy to describe: Cognition keeps growing, but growth slows earlier than expected, margins stay compute-heavy, and Anthropic or OpenAI capture more of the economic value. Cognition could still be an excellent company in that scenario and be worth far less than $47 billion.

Q18So is Cognition really worth $47B today?

We think Cognition’s $47 billion valuation is aggressive but plausible, with the evidence currently leaning closer to “expensive” than “crazy.”

The case against the valuation is straightforward. Cognition trades at a much higher revenue multiple than Cursor, Lovable, Replit and premium public software companies. Its enterprise gross margin is still far below normal SaaS levels. Cash spending has exploded as the company builds model and compute infrastructure. Competition from Claude Code and Codex is getting stronger rather than weaker.

Yet Cognition has something many richly valued AI startups lack: the business is growing fast enough that the valuation has recently struggled to outrun it. The latest step up in price came alongside an almost identical increase in reported annualized revenue. Investors are also trying to put nearly ten times more money into the current round than Cognition originally planned to raise, according to Bloomberg. That does not prove the price is right, but it shows how scarce investors currently believe businesses growing at this speed have become.

The fresh market evidence is supportive too. AI coding-agent adoption is already mainstream among professional developers. Large companies are scaling the technology. Customers are putting seven figures into Cognition contracts. Mercedes has taken Devin from a pilot into a broader deployment, and Goldman is already working on how to teach agents its internal engineering knowledge.

We would still refuse to value Cognition as though the hard part were finished. The company has to turn today's hypergrowth into several billion dollars of annual revenue, improve a roughly 50% enterprise gross margin and remain valuable while the frontier model companies attack the same customers directly.

If Cognition does those things, $47 billion could look surprisingly reasonable within a relatively short period. If revenue growth normalizes before the company reaches several billion dollars, the valuation can fall a long way without Cognition ever becoming a bad business.

So yes, Cognition can genuinely be worth $47 billion. We just do not think the business has fully earned that number yet. Investors are paying today for a version of Cognition that still has to arrive.