AI Labs Buy Failed Startups’ Data to Train Models
AI labs are buying failed startups’ code, emails and chats as training data, creating new payouts while raising privacy, consent and pricing concerns.
Summary
When Keelan Miskell closed his crypto startup with anonymous users in 2026 after two and a half years, he netted roughly $25,000 for its code and expects about $15,000 for Google, Notion, Figma and Slack archives. Carta and AngelList estimate shutdowns have risen 26% to 56% annually. Dori Yona’s SimpleClosure added data sales in April through Asset Hub, selling code, emails, chats and documents to foundation AI labs, simulated workspace vendors and established model builders. After more than 150 deals, it expects to return over $10 million by year end 2026; 47% of deals now include code or data. Rival Sunset, led by Brendan Mahony, values assets by head count, operating years, volume and tool use.
Demand appeared in 2025 as labs reached the limits of public and synthetic datasets. Surge AI buys archives for reinforcement learning gyms where agents practice debugging complex IT systems and drafting appellate briefs; Nick Heiner seeks deep records spanning disagreements and judgment calls. Real data avoids synthetic mistakes such as nonexistent dates, but no company has published comparative performance tests. Prices vary because buyers estimate useful tasks from inputs such as 100,000 code lines.
In August 2026, Google outbid Mercor in Spirit Airlines’ bankruptcy, paying $10 million for decades of data including 100 million emails, 500 million Teams messages and 176,000 employee records. In September, SpaceX’s AI unit was weighing similar purchases. Spirit flight attendants objected over privacy. Attorney Greg Ewing calls such archives legally radioactive: suppliers and buyers scrub identifiers and exclude HR records, but emails and calendars can expose medical, school and romantic plans. Emory professor Ifeoma Ajunwa, author of The Quantified Worker, says proceeds usually reach senior creditors and investors, not employees; she urges express consent, but no federal law bans sales. Founders often cannot see buyers, uses or fair value. Prices may stabilize in 2027 unless an AI bubble bursts.
Positives
- Roughly $40,000 could be recovered from Keelan Miskell’s code and workspace archives after his crypto startup closed.
- More than 150 Asset Hub deals could return over $10 million to failed startups by the end of 2026.
- 47% of SimpleClosure transactions now include code or data, turning previously discarded digital assets into recoverable value.
- Real workplace archives avoid synthetic data errors such as references to nonexistent dates.
- Google’s $10 million Spirit Airlines purchase demonstrates significant demand for large, decades-spanning operational datasets.
Risks & concerns
- No company has published comparative tests showing that real workplace archives improve AI performance over synthetic or public data.
- 100 million emails, 500 million Teams messages and 176,000 employee records in the Spirit sale triggered union privacy objections.
- Senior creditors and investors usually receive sale proceeds, while employees whose communications created the data typically receive nothing.
- No federal law bans these sales, despite concerns that emails and calendars expose medical, family and romantic details.
- Founders have little visibility into buyers, intended uses or fair pricing, while valuations vary widely and an AI bubble could disrupt demand.