AI Labs Buy Defunct Startups' Data to Train Models

What happened

AI developers are increasingly acquiring the data assets of failed or shuttered startups to train their models, according to reporting by MIT Technology Review. The practice has created a nascent secondary market in which defunct companies' code bases, user records and other digital archives — once considered worthless by their founders — are being sold on for sums that can run into the tens of thousands of dollars.

The report centres on Keelan Miskell, founder of a crypto startup he began winding down this year. Miskell initially assumed the firm's data held little value, given its small size and largely anonymous user base. He has since sold the company's code base for roughly $25,000, with a further payment still pending.

The emergence of this market suggests AI labs are widening their search for training data beyond the large, well-known public datasets and partnerships that have dominated headlines, turning instead to the long tail of smaller, defunct ventures sitting on code, logs and user interactions accumulated before they folded.

Why it matters

This points to intensifying competition among AI developers for differentiated training data as the supply of readily available public text and code becomes increasingly exhausted or contested. A market for failed startups' digital assets suggests labs are willing to pay real money for even modest, niche datasets if they offer signal that generic web-scraped data does not — such as domain-specific code, interaction patterns, or structured records from a working product.

For founders and investors, it also introduces a new, previously overlooked line item in startup wind-downs: data as a recoverable asset, not a liability to be quietly deleted. That has implications for how companies handle data governance, user consent and deletion obligations when they shut down — questions that sit squarely at the intersection of AI development and trust.

By the numbers

• $25,000 — amount Keelan Miskell has received so far for his startup's code base
• A further, unspecified sum — still pending from the same sale
• 2026 — the year Miskell began winding down the crypto startup in question

The Renascence take

The obvious story here is a new revenue stream for failed founders. The more interesting one is what this market reveals about consent and expectation. Users who signed up for a small crypto startup almost certainly never imagined their activity would end up as training material for a large language model years later.

Data created inside one experience is increasingly being repurposed to shape entirely different ones, often without the original users ever being asked. That is a behavioural trust gap, not just a legal one: people consent to a product's stated purpose, not to whatever afterlife their data might have. Organisations building or winding down digital products should treat data provenance and deletion commitments as explicit, visible parts of the customer promise — not fine print to be reinterpreted once the company itself has stopped operating.