AI labs are buying data from dead startups to train their models

When your startup fails, usually the only thing left to do is call the lawyer, file the paperwork, and try to salvage your reputation. But the AI boom has introduced a strange new line item in the liquidation process: selling your digital corpse to the highest bidder.

We are seeing a new trend where AI labs are scouring the wreckage of the last tech cycle to buy up data from defunct companies. Founders who thought their user logs, code bases, and internal chats were worthless are finding out that, in the eyes of a hungry Large Language Model (LLM), everything is fuel. Even a failed crypto startup with a few thousand users can net a founder five figures just for handing over the keys to the database.

The Scrap Metal Phase of AI

For the last decade, we told builders that data was the new oil. It turns out that was wrong. Data is more like scrap metal. Individually, a rusted pipe isn't worth much, but if you have enough of it, you can melt it down and build something new. AI labs have reached a point of desperation where they have exhausted the high-quality, public internet. They have scraped Reddit, Wikipedia, and every news site they could find. Now, they are going deeper.

They want the stuff that was never meant to be public. They want the internal logic of proprietary code bases. They want the raw, messy interactions of users in a niche financial app. This isn't about quality; it is about volume and diversity. If you are training a model to understand how humans build software or how they trade assets, the failed experiments of 2021 are just as useful as the successes of 2024.

Founder Perspective: The Ethical Gray Zone

As a founder, I get the appeal. If you are staring down a complete loss, and an AI lab offers you twenty-five grand for your archives, it feels like a win. It covers some of the debt. It gives you a tiny bit of runway for the next thing. But we have to look at the optics and the ethics here. Most of these startups promised their users privacy. They promised that their data was being used to provide a specific service, not to train a global intelligence that belongs to a multi-billion dollar corporation.

The reality is that most Terms of Service are written so broadly that this is technically legal. But legal and right aren't always the same neighborhood. When you sell your data to an AI lab, you are effectively selling your users' behavior without their consent or knowledge. In the crypto world, where we talk a big game about sovereignty and ownership, this feels particularly hypocritical. We spent years telling users to trust us because we were decentralized, only to sell their footprints to the most centralized entities on the planet.

Why Labs Want the Garbage

You might wonder why a top-tier lab would want the data from a failed crypto startup. It comes down to edge cases. AI models are getting very good at general knowledge, but they still struggle with specific domains. Crypto is a goldmine of specific domain knowledge. It involves complex cryptography, unique smart contract logic, and a very specific dialect of financial jargon.

By buying up the code bases of these defunct firms, labs are essentially buying a textbook on how *not* to build things, as well as how to build them. They are looking for the 'long tail' of data that doesn't exist on GitHub. They want the stuff that lived in private repositories. It helps the model understand the nuances of debugging and the architectural choices that developers make in the real world, rather than just the polished examples found in tutorials.

The Builder’s Takeaway

If you are currently building, this should change how you think about your data architecture. If your data has value even if your product fails, you need to be intentional about how you store it and what your privacy policy actually says. Don't just copy-paste a generic TOS from a generator. Actually think about whether you want your users' data to be a liquid asset in a fire sale.

For those currently winding down, don't be surprised if the vultures start circling. They aren't looking for your intellectual property to relaunch your product; they are looking for the raw bits to feed the beast. It is a grim sort of recycling, but it is the reality of the current market. We are in a cycle where the value of a company's 'ghost' might actually be higher than the value of its living product.

The Honest Truth

We are entering an era where nothing truly disappears. Every line of code you write and every user interaction you log is being eyed as training material. As builders, we have to decide if we are okay with that. AI labs are desperate for growth, and they will pay for it in the short term. But for the ecosystem, this creates a weird incentive where failing fast is just another way to generate a data set for a competitor.

It is a cynical end to the 'build in public' era. Now, we are 'dying in private' and selling the remains to the highest bidder. It might be a good exit strategy for a founder in a pinch, but it is a concerning precedent for the future of digital privacy and the value of human-generated work. If everything we do is just fodder for the next model, what are we actually building?

Read the original at MIT Technology Review →