AI Labs Turn to Defunct Startup Archives for Real‑World Training Data

A new source of data for machine‑learning models

Artificial‑intelligence research groups are increasingly buying the digital remnants of failed startups. The practice, highlighted in MIT Technology Review’s The Download newsletter on 28 October 2026, shows how companies that specialize in winding down businesses are repurposing codebases, version‑control histories, and internal workspaces as a commodity for AI training.

How the market emerged

The trend gained visibility when Keelan Miskell, founder of a crypto‑focused startup that was shutting down, discovered that his company’s archived repositories and collaboration tools could be sold for “tens of thousands of dollars.” Miskell’s experience is now part of a nascent market where firms that assist in the closure of startups package and sell these digital assets to AI labs.

Demand for such data has risen as large‑scale model developers exhaust the utility of publicly available datasets and synthetic data generators. Real‑world records—such as production‑grade code, bug‑tracking logs, and internal documentation—offer richer, more varied examples that can improve model robustness, especially for tasks like code generation, software‑engineering assistance, and automated debugging.

Legal and ethical questions

While the transactions appear legal—data owners retain the rights to sell assets they created—the practice raises several concerns:

• Privacy and consent – Even if a startup is defunct, its data may contain personal information about former employees, contractors, or users. The extent to which such data is anonymized before sale is unclear.
• Intellectual‑property rights – Some code may incorporate third‑party libraries or licensed components. Transferring ownership without clear clearance could expose buyers to infringement claims.
• Bias and fairness – Training models on data from a narrow set of failed ventures could embed specific industry or cultural biases, affecting downstream applications.
• Transparency – AI developers often disclose dataset sources in research papers, but the provenance of purchased startup archives is typically opaque, complicating reproducibility.

Practical implications for stakeholders

For startup founders and investors

• Potential revenue stream – Entrepreneurs winding down a company can monetize their digital legacy, offsetting closure costs.
• Due‑diligence – Before selling, founders should audit data for personal or third‑party content and consider legal counsel to mitigate liability.

For AI research labs and commercial developers

• Data diversification – Acquiring real‑world code histories can fill gaps left by public repositories like GitHub, especially for niche languages or proprietary frameworks.
• Compliance overhead – Labs must implement rigorous vetting processes to ensure purchased data complies with privacy regulations (e.g., GDPR, CCPA) and respects IP rights.
• Cost considerations – While the reported sales are in the “tens of thousands” range, larger aggregators may command higher prices, influencing budgeting for data acquisition.

What remains unknown

The newsletter does not provide quantitative estimates of market size, the number of firms involved, or which AI labs are active buyers. It also does not detail the contractual terms that govern data transfer, such as warranties, indemnities, or restrictions on downstream use. These gaps leave open questions about the scalability and sustainability of the model.

How readers can verify the story

The full discussion appears in MIT Technology Review’s article titled The Download: dead startups fueling AI and interspecies communication (published 28 Oct 2026). Readers can access the original piece at:

https://www.technologyreview.com/2026/10/28/1147767/the-download-dead-startups-fueling-ai-interspecies-communication/

The article provides the primary source for the example of Keelan Miskell and the broader observations about the emerging data‑sale ecosystem.

--- This report is based solely on the verified information provided by MIT Technology Review and does not contain any invented quotes or unverified details.