Cognition Raises $1 Billion as Devin Revenue Run Rate Nears $500 Million

Cognition has raised more than $1 billion at a $26 billion valuation, with Lux Capital, General Catalyst, and 8VC leading the round. Scott Wu said the financing is meant to support aggressive growth, expand compute, add to the team, and keep the company independent.

The valuation discussion turned quickly to revenue. Ed Ludlow noted that Cognition’s revenue run rate had been only a few million dollars when Wu first appeared on Bloomberg Technology around the beginning of 2025 or 2024, and about $37 million a year earlier. Wu said it is now “getting close to 500 million” after roughly two years in business.

A Bloomberg graphic attributed to Cognition put the current revenue run rate at $492 million, and said Devin enterprise usage had grown 11x in six months and 75x in 12 months. The same graphic said Devin had been deployed across leading companies in every vertical.

Wu framed the financing as a response to demand from large organizations using AI systems for software work. His central claim was that “AI is doing real work at real companies everywhere,” and that the relevant market is not confined to software vendors. “Every company in 2026 is a software company,” he said.

The customer examples he gave were large institutions: Cognition works with all of the top five health insurers in the United States, with banks, and with the Treasury and NASA. In health care, Wu said customers are building more tooling for care providers, cutting down on price, and covering more people. In banking, he described customers using software delivery to give end users access to what they need.

Wu tied that demand to the size of the software labor market. He said there are about 30 million to 35 million software engineers in the world today, and Cognition wants to make all of them “ten times more efficient.” He added that the company believes there is “a lot more than ten times more software to build.”

Caroline Hyde pressed Wu on the crowded AI-coding market and on the threat from the large AI labs themselves, especially as labs and adjacent companies acquire or absorb startups, talent, assets, and products. Wu’s answer was that Cognition’s independence is not a weakness against the labs; it is the company’s strategic position.

He described Cognition as “fully independent and fully neutral,” and said that makes it better aligned with customers. Cognition works closely with OpenAI and Anthropic, but also with Google, xAI, and others, according to Wu. He said the company has deep relationships with those labs and collaborates with them on research.

The product implication is that Devin is not tied to a single model family. Wu called Devin a “compound system” that works with different models, allowing Cognition to choose the best model for a given use case. That is what he meant by describing Cognition as “the Switzerland” of the sector: “Devin is a compound system that works with all of these different models and because of that we’re able to be the Switzerland in the equation here.”

Ludlow sharpened the issue with a concrete competitive scenario: if SpaceX were to acquire Cursor, how would that change the field? He said engineering teams at Nvidia liked Cursor partly because they could swap underlying models depending on the coding objective, and suggested the same model-choice freedom may be one reason customers like Devin. He also raised the importance of data and of whom a company is beholden to.

Wu did not predict a specific outcome for Cursor or SpaceX. He said there are many strong teams working on code, including the labs themselves, and that the ecosystem is broad and vibrant enough to sustain different players in different positions. There will continue to be first-party products from the labs, he said. But he argued that a neutral provider has an important place because it can serve whichever model is best for each case, rather than being incentivized toward “a single series of models.”

Asked whether Cognition itself would use the new funding for mergers and acquisitions, Wu said he was sure more M&A would happen in the sector and that “different things” would come up. But he said Cognition’s focus is continuing to grow the business and serve customers.

Wu described Devin not as a narrow code-generation tool, but as an AI software engineer being adopted by teams to execute more aggressively on product roadmaps. When Ludlow asked whether that work was new code or work on old codebases, Wu said it was both. The majority of work, he said, is on existing codebases: building on them, adding features, and maintaining forward progress.

Cognition’s own internal use was the main example Wu offered. He said that more than 90% of the code Cognition writes is written by Devin, because the company uses Devin throughout the day to build Devin itself.

A Bloomberg chart attributed to Cognition showed the share of Cognition code committed by Devin rising sharply over several months: 13% in December 2023, 17% in January 2024, 33% in February, 76% in March, 83% in April, and 89% in May. The chart’s headline stated that “self-driving software development” was on the rise, and its subtitle said that at Cognition, roughly 90% of code is committed by Devin.

The on-screen product examples showed Devin inside project and task workflows, not only in a code editor. One Cognition-sourced interface displayed “Devin Spaces,” a project board called “COG-GTM,” work columns such as “TO DO,” “IN PROGRESS,” “IN REVIEW,” and “DONE,” and tickets including “Delete Fund Transfer,” “View Transaction History,” “Implement Biometric Login,” and “Two-Factor Authentication Setup.” Another pane showed a task to “Convert JS to TS in MetaMask Mobile,” with Devin assigned through automation and a message: “Hey, it’s Devin! I’m currently researching your task and I’ll be back soon with a proposal. Please check back in 5 minutes.”

Those examples support Wu’s broader claim that coding agents are being adopted within ordinary software-team workflows. The visible interface placed Devin around boards, tickets, assignment, automation, and task research; Wu’s spoken claim was that teams adopting Devin and coding agents “en masse” are able to do more and execute more aggressively on roadmaps.

The labor-market concern surfaced directly: if Devin and similar systems materially increase productivity, what happens to software engineers? Hyde noted that Cognition has continued hiring even with that productivity, and asked what AI disruption means for how many software engineers will be needed.

Wu’s answer was that the job will evolve, and some skill sets will change, but he expects “far more people” building software and products, not fewer. His evidence was historical: he said there are now about 30 million to 35 million software engineers globally, compared with under 1 million roughly 20 to 25 years ago at the start of the century. That increase, he argued, coincided with a massive rise in the amount of software produced.

The same logic underpins his view of AI coding agents. If software becomes more efficient to produce, Wu expects the world to make more software rather than simply require fewer people to make the same amount. Cognition’s opportunity, as he described it, depends on software demand expanding with productivity.

The near-term target is correspondingly aggressive. Asked by Ludlow for one metric defining the rest of the year, Wu said Cognition intends to cross $1 billion in revenue run rate and “keep going further even beyond that.”