AI labs acquiring training data from shuttered startups

By Splitfeed EditorialLast updated Oct 13, 2026, 16:50 UTC

On October 13, 2026, MIT Technology Review reported that AI laboratories are buying digital archives from defunct startups to use as training data for their models. The report centers on Keelan Miskell, who wound down his crypto startup in 2026 and sold its code base for roughly $25,000, with an additional sum still pending. Miskell said he had not expected the data to attract buyers, given that his users were anonymous and the company only a few years old. The transaction illustrates a broader pattern in which AI companies are treating the digital leftovers of failed businesses as a commodity with market value.

The practice matters because it surfaces questions about where AI training data comes from and who has the right to sell it. When a startup closes, its data assets, including code, user interactions, and communications, do not simply disappear. They can change hands through asset sales, and the original users whose activity generated that data may not know it has been sold, let alone used to train commercial AI systems. The story also points to a growing secondary market for training data, one that exploits moments of corporate failure to source material that might otherwise be unavailable. What remains genuinely in dispute is whether existing privacy frameworks, terms of service agreements, and data protection laws are adequate to govern these transactions, especially when user consent was never sought for this kind of downstream use.

None of the three camps have published reactions to this specific story yet. The Pro-AI camp would typically argue that purchasing archival data from willing sellers is a legitimate and efficient way to build better models, and that the transaction benefits both the seller and the broader goal of AI development. The Anti-AI camp would be expected to raise concerns about user privacy, arguing that anonymous users never consented to have their data repurposed as training material and that the absence of identifiers does not neutralize the ethical problem. The Middle Ground camp would likely call for clearer regulatory standards governing data asset sales at startup wind-downs, seeking a framework that allows beneficial data reuse while protecting individuals from unknowing participation.

The argument is unlikely to resolve without regulatory action or a high-profile legal challenge to one of these transactions. Developments worth watching include any guidance from data protection authorities on whether startup asset sales trigger consent or notification requirements, and whether any AI laboratory publicly discloses its participation in this emerging data market.