AI Training Data Market: Labs Are Buying Code and Archives From Defunct Startups
When Keelan Miskell started winding down his crypto startup this year, he assumed its data was nearly worthless: anonymous users, a company only a few years old. He has since netted roughly $25,000 for the code base, with a further payment still pending. His experience points to a quiet secondary market where AI labs buy the leftovers of failed companies to feed their models.
What Miskell actually sold
Miskell's company had no obvious data asset. Its users were anonymous, and the business was young. Yet the firm's digital archives, including its code base, found a buyer willing to pay about $25,000, with more money still on the way.
The price is modest by AI-industry standards. What matters is the direction: a founder with a small, short-lived company discovered that what he thought was a liability to clean up was something labs would pay for.
Why labs want a dead startup's files
Public web text has been scraped heavily for years, and the easy supply is largely used up. Labs now care about material that was never published: working code, internal documents, and records of how real products were built and debugged. That kind of data shows how people actually work, which is hard to get from public pages.
A shuttered startup is also an unusually clean seller. There is one owner, no ongoing product to protect, and a founder who wants to close the books. That makes these deals faster to do than a negotiation with a large platform.
How this differs from the platform licensing deals
The best-known data deals so far have involved big content platforms. Reddit's licensing agreement with Google, reported at around $60 million a year, and Stack Overflow's partnership with OpenAI are examples. Those are negotiated between large companies with lawyers on both sides.
Deals with failed startups run at a different scale and with much less scrutiny. The seller is often an individual, the data can be a messy mix, and nobody may have checked what is inside. That is where the risk sits.
Where the legal and ethical problems come in
Three questions follow any such sale, and the answers are rarely simple:
- Consent. Did users agree to their data being used for model training? Terms of service written years ago rarely anticipated it.
- Personal data. Anonymous does not always mean unidentifiable. Logs, wallet addresses and support messages can be linked back to people.
- Third-party rights. Code bases often contain open-source components or contractor work whose licenses may not allow resale or training use.
There is also a precedent from insolvency cases. When 23andMe filed for bankruptcy in 2025, what would happen to its customers' genetic data became a public concern. Regulators have also at times required companies to delete models built on improperly obtained data. A buyer that skips due diligence could end up carrying that risk.
What founders and AI buyers should take from this
For founders, a shutdown now includes a decision about data: delete it, keep it, or sell it. Before selling, check user terms, strip personal information, and audit licenses in the code. Treat the transaction like selling any asset with strings attached.
For labs and companies using their models, provenance becomes a procurement question. If a model was trained partly on secondhand startup archives, the supplier should be able to say where the data came from and what rights came with it. I'd expect enterprise customers to start asking for this, though that is a prediction based on how they already handle open-source license compliance, not something that has been announced.
Takeaway
Miskell's $25,000 is small, but it shows that data from failed companies now has a market price. The open questions are who sets the rules for these sales and how buyers prove they had the right to buy. Expect those two issues to matter more than the dollar amounts as the market grows.
Comments
Post a Comment
Please read our comment policy before commenting.