AI labs are buying data from dead startups to train their models
2026-10-17 · MIT Technology Review
AI Labs Are Buying Data from Dead Startups to Train Their Models
As AI training labs hit the limits of public and synthetic datasets, a new growth sector has emerged in tech: businesses catering to dying startups are now selling their data remains to AI labs.
Monetizing Defunct Startups' Archives
When Keelan Miskell wound down his crypto startup, he unexpectedly netted roughly $40,000 by selling his code base and workspace data from Google, Notion, Figma, and Slack. The sale was brokered by SimpleClosure, part of an end-of-life ecosystem that has emerged as startup shutdowns rise between 26% and 56% annually. These tech "undertakers" help handle the legal, tax, and operational headaches of going belly-up.
The Data Resale Ecosystem
SimpleClosure introduced data sales to its Asset Hub platform in April, helping founders sell code and workspace data like emails, chats, and documents.
- Buyers: Foundation AI labs, vendors building simulated workspaces for AI training, and mature tech companies developing their own models.
- Scale: SimpleClosure has handled over 150 deals, expects to return over $10 million to startups by the end of 2026, and now includes code or data sales in 47% of its transactions.
Industry rival Sunset is also experiencing rapid growth in data resale. Sunset values assets based on head count, years in operation, data volume, and tool usage. CEO Brendan Mahony compares this "operational exhaust data" to particle board—wood shavings previously thought worthless but actually highly valuable.
The Demand for Real-World Data
Demand for defunct companies' data surged last year when AI labs exhausted public and synthetic options. Companies like Surge AI buy data archives to build "reinforcement-learning gyms" where AI agents practice tasks like debugging or drafting legal briefs. Real archives avoid synthetic world errors, such as referencing non-existent dates. However, performance gains remain unclear, and no comparative tests have been released. Consequently, pricing is highly volatile, based on "research theses" estimating useful task yields.
Despite uncertainty, major deals are proceeding:
- Google's Spirit Airlines Purchase: In August, Google outbid Mercor to buy Spirit Airlines' decades of data during its bankruptcy for $10 million. The haul included 100 million emails, 500 million Teams messages, and 176,000 employee records.
- SpaceX's Interest: Bloomberg reported in September that SpaceX’s AI unit is considering buying data from struggling startups.
Legal and Privacy Risks
Legally, these transactions occupy a gray area. Spirit’s flight attendants’ union objected to the sale on privacy grounds, and attorney Greg Ewing labels such data "legally radioactive." While suppliers and buyers employ data scientists to scrub personally identifiable information and HR records, private details like doctor appointments and date-night plans still lurk in emails and calendars. Ewing describes a "Wild West of AI training" only just beginning to transition to proper licensing agreements.
Ethical Concerns: Who Profits?
The proceeds from these data windfalls typically flow to senior creditors and investors, not employees. Ifeoma Ajunwa, a law professor at Emory University, views this as a form of "captured capital." While legal precedent permits monitoring to facilitate work, that justification vanishes once a firm is out of business. Ajunwa argues employers should obtain express consent before selling workers’ data, though no federal law currently bans such sales.
Market Limitations and Future Outlook
While offering a chance for failing businesses to recover value, wind-down sales have limitations. Founders have little visibility into who buys the data or how it will be used, and fair pricing is hard to guarantee. Data prices may stabilize in the coming year unless an AI bubble pops first. Regardless, tech’s new undertakers will remain to appraise what’s left.