AI labs are buying data from dead startups to train their models
The liquidation of failing startups is quietly becoming a lucrative pipeline for AI training, as data brokers resell discarded archives to labs starved for real-world datasets. This emergent market raises ethical and legal questions about the commodification of corporate data, the opacity of transactions, and who ultimately benefits from the windfalls—while offering a glimpse into the AI industry’s growing reliance on unconventional sources of information.
Who are the key players profiting from the resale of defunct startups' data?
Specialized brokers like SimpleClosure and Sunset act as intermediaries, connecting dying startups with AI labs, foundation model developers, and tech firms. These brokers estimate asset value based on factors such as team size, operational data volume, and tool usage, with deals ranging from a few thousand dollars for small archives to multimillion-dollar transactions like Google’s $10 million purchase of Spirit Airlines’ data during bankruptcy.
What kind of data is being sold, and why is it valuable for AI training?
The data includes code repositories, workspace communications (emails, Slack chats, documents), and operational records that capture real-world workflows, disagreements, and decision-making processes. AI labs value this operational exhaust because it provides authentic, context-rich datasets that synthetic or public data often lack, such as nuanced human interactions or task-specific problem-solving—though the actual performance gains from such data remain unproven.
What legal and ethical concerns surround these transactions?
Critics argue that selling employee data—even from defunct companies—without explicit consent is ethically dubious and legally risky, as such data may contain sensitive personal information (e.g., medical appointments, family arrangements). Legal experts describe the practice as legally radioactive, yet there are no federal laws prohibiting it. Additionally, proceeds typically flow to investors and creditors, not workers, raising questions about equity in the distribution of these windfalls.
How does the valuation of this data work, and what uncertainties does it introduce?
Valuations are speculative, often based on rough estimates of how many useful tasks or training scenarios can be extracted from a dataset (e.g., 100,000 lines of code). Prices vary widely because there is no standardized method for assessing value, and no company has yet published comparative tests showing the superiority of real-world data over synthetic alternatives. This opacity could lead to unfair deals for sellers and unpredictable costs for buyers.
Ce que ça pourrait changer
This trend could normalize the treatment of corporate data as a fungible asset, even in dissolution, potentially eroding privacy norms in the AI industry. It may also deepen the divide between those who profit from data extraction (investors, AI labs) and those whose labor produced it (employees), while accelerating the AI sector’s dependence on unconventional data sources that bypass traditional curation or consent mechanisms. Without clearer regulations or ethical frameworks, the practice risks reinforcing inequities in the tech ecosystem.
