Marc Andreessen spent 2011 arguing that software would eat the world. On Tuesday he backed a narrower and stranger claim: software is now eating itself. Andreessen Horowitz co-led a Series E that values Cognition, maker of the Devin coding agent, at $48 billion , four months after investors priced the company at $26 billion .

The round exceeded $2 billion; new investors a16z and Accel led it, with Founders Fund, General Catalyst and Avenir returning alongside a syndicate that runs past thirty names and includes Nvidia itself . Cognition says run-rate revenue climbed from $492 million in May to almost $900 million at announcement. The company has not disclosed how it computes that run rate , a detail investors paying roughly 53 times it should want. The Information projects Cognition will reach $4 billion to $5 billion in annualized revenue by year end, which would require the business to quadruple in four months.

The venture logic can be somewhat unusual since a16z also backed Cursor, which SpaceX bought in an all-stock deal worth $60 billion that closed August 14. Leading a round in Cursor’s closest rival weeks after that exit is a public statement that the firm sees AI coding as a multi-winner market rather than a race for one seat. Cursor’s own market share fell from 41% in June 2025 to about 26% by May.

Andreessen framed the bet in a post titled Investing in Cognition . He notes that technology’s share of market capitalization among the ten largest companies rose from 31.5% to 94.4% today over fifteen years, and that this happened while only one in 300 people on earth wrote code, every line by hand. His central data point is internal: Devin went from writing 13% of Cognition’s production code to more than 90% in one year. The figure is self-reported by a vendor about its own product, which is exactly the kind of number that should be discounted until customers replicate it.

Cognition says Mercedes-Benz cut an eight-month legacy modernization project to eight days , and that Itaú, Latin America’s largest bank, now remediates 70% of its security vulnerabilities automatically. Cognition also reports that Itaú completed COBOL migrations five to six times faster without adding headcount after other agents broke on the language’s 72-character column limit. RV Tech, the Rivian and Volkswagen joint venture building software for up to 30 million vehicles, uses Devin for ticket triage and test generation . It should be noted tht every one of these figures comes from Cognition or its investor.

Andreessen’s counter to the obvious objection is Milton Friedman: human wants are infinite, so demand for software is infinite, and cheaper production expands the market rather than shrinking it. He points out that compilers, open source and cloud each triggered predictions of the end of software engineering, and each grew the profession. The public markets have not fully bought the argument. Morgan Stanley analysts this year described a selloff exceeding $1 trillion in software stocks, driven by the fear that if AI lets a company halve its staff, it also halves the seats it pays for. Andreessen’s thesis and the SaaS repricing are two readings of the same event.

Tech Funding News reports Cognition’s cash burn could reach $800 million this year on compute alone, and that CEO Scott Wu told staff the company does not believe in work-life balance after cutting 30 jobs nine months ago. Cognition positions itself as an independent agent lab that routes work across every frontier model rather than tying customers to one provider. That independence is a hedge against the labs that supply its intelligence and also compete with it.

The product direction points at where the market is heading. Cognition’s new Auto-Triage, Security Swarm and Automations let Devin start work from events in Slack, GitHub and Linear without a human opening a chat. Competition has moved from code completion to the plumbing that connects an agent to incidents, tickets and repositories so work begins before anyone asks.

Actionable insights for VC are the following; Insist on run-rate methodology and net revenue retention before accepting a 50x multiple, because the gap between $900 million and a $5 billion forecast is the entire thesis. Treat seat-based software in the portfolio as a repricing risk. And look for the companies that own the integration surface, since Cognition’s round says that layer is where enterprise value is accumulating.

For founders, the lesson is to pick work with a measurable before and after. Legacy migration and vulnerability remediation won Cognition its logos because eight months to eight days is a number a CFO can sign. For enterprises, the same rule applies in reverse: start agents where the baseline is already known, and demand that the vendor’s outcome figures be reproduced on your own codebase before the contract expands.

Andreessen closed his 2011 essay with one line about knowing where he was putting his money. Fifteen years later the money is $2 billion and the counterparty is a two-year-old company that says it barely writes its own code. Whether that is the future of software or the peak of a cycle will be settled by customer numbers.