AI labs are buying data from dead startups to train their models– www.technologyreview.com

SimpleClosure added data sales to its services in April. Its Asset Hub platform helps founders sell their code and workspace data—emails, chats, documents. Buyers include foundation AI labs, vendors that construct simulated workspaces for AI training, and mature tech companies building their own models. Asset Hub has handled over 150 deals and expects to return over $10 million to startups by the end of 2026. CEO Dori Yona has seen these transactions, once rare and “somewhat taboo,” become increasingly routine, with 47% of SimpleClosure deals now including code or data sales.

And it’s got company. Industry rival Sunset also has a rapidly growing data resale business. Founder and CEO Brendan Mahony says Sunset values a dying client’s assets by head count, years in operation, volume of data, and how the team used its tools. Particle board, he says, is a good analogy: “This operational exhaust data is like the wood shavings that everyone thought had no value, but it actually has a ton.”

Demand for defunct companies’ data emerged suddenly last year, as AI training labs hit the limits of public and synthetic datasets. Surge AI, for example, buys data to build “reinforcement-learning gyms” where clients’ AI agents practice tasks—think debugging a complex IT architecture or drafting appellate briefs. Nick Heiner, who leads the company’s work on these environments, says he seeks data archives reflecting a range of situations, from disagreements to tricky judgment calls. Heiner explains that one company’s Slack history, if it captures conversations with variety and depth, can be worth more than another’s Slack, email, and calendars combined.

Real archives avoid common errors in synthetic worlds, such as references to dates that don’t exist, but the extent of performance gains remains unclear. No company has yet released tests comparing results of real-world data training with the alternatives. Pricing reflects this uncertainty: Buyers quote wildly different numbers, Heiner says, because valuations are based on “research theses” estimating how many useful tasks come from, say, 100,000 lines of code.