AI labs buy data from dead startups to train models

When a startup fails, its servers usually go dark and its assets get liquidated. But in the AI boom, even dead companies can have valuable afterlives. According to reports, AI labs are buying data from failed startups to train their models, turning archived codebases, user interactions, support tickets, and internal documents into training material. One founder reportedly earned about $25,000 by selling the archives of a defunct crypto startup, including its codebase and user data. What was once a write-off became a modest exit. The demand reflects a broader shift in AI development. Public web scraping has become legally and ethically contested, and much of the internet is noisy, duplicated, or low quality. Proprietary datasets from real companies can be more structured and domain-specific: how customers actually behave, how software breaks, how teams communicate. For AI labs racing to build better agents and copilots, such data can offer an edge that public sources cannot. For struggling founders, the market offers a lifeline. Instead of watching assets disappear, they can sell data to recoup some costs or pay creditors. But the trend raises hard questions. Did users consent to having their data sold to an AI company? What happens to sensitive information in codebases or customer records? Bankruptcy proceedings may not be equipped to handle privacy obligations, and buyers may not inherit clear consent. There are also risks of re-identification, proprietary secrets, and unfair competition if data is repurposed beyond original terms. Regulators and privacy advocates may soon scrutinize these transactions. Companies should audit what data they hold, delete what they cannot legally sell, and disclose potential AI uses in privacy policies. Buyers need due diligence that goes beyond technical usefulness to include legal and ethical provenance. If the market matures without guardrails, it could erode trust in both startups and AI systems. If handled transparently, however, it might create a more orderly way to reuse valuable information, without treating users as an invisible asset on the balance sheet.