According to The Information, citing sources, Mercor and similar firms gathering data for AI labs are driving demand to buy or license the internal datasets of startups shutting down or being acquired.

A concrete example

The example in the reporting captures the situation well: eight days after AI agent startup Warmly agreed in late June to be acquired by HubSpot, its chief executive Maximus Greenwald found an unusual email in his inbox.

The offer arrives, in other words, while the company is still winding down. That shows how quickly this demand moves.

Why this data is valuable

The easy data for training models has largely been used up. The web has been crawled, and most of what remains is either behind access controls or of low quality.

What is valuable now is data that exists nowhere else. A startup's internal dataset fits that description exactly: real user interactions, domain-specific labelled examples, outputs corrected by humans. None of that appears in any web crawl.

The data of a closing startup is also left without a claimant. When a company ceases operating, the data sits there as an asset and becomes sellable in the wind-down process.

The problems this raises

The legal side of this market has not settled, and several questions remain open:

  • Consent — did the users who generated the data agree to it being sold to an AI lab? Most privacy policies state that data may transfer if the company is transferred, but "transfer" and "sale to a third party" are not the same thing.
  • Scope — a user accepts that their data will be used to improve the product they are using; what does that acceptance cover once the product no longer exists?
  • Traceability — nobody tracks where the data goes. Once a wound-down company's dataset changes hands, the trail disappears.

What it means for startups

There is a founder's side to this as well. For a closing startup the dataset becomes an unexpected line of income — sometimes enough to cover final salaries.

That introduces a new calculation into the startup ecosystem: data accumulated while building a product can hold value even if the product fails. How that calculation gets explained to users has not yet been answered.

Why it has emerged now

This market has formed at the intersection of two trends. On one side, quality data for model training is becoming scarce; on the other, a large number of startups founded in 2021 and 2022 are reaching the point of closing or being acquired.

Put together, that produces a market with both supply and demand. Demand is urgent while supply is in liquidation with weak bargaining power — an equilibrium in which prices form in the buyer's favour.

What a user can do

As a user it is hard to see where you sit in this chain. A few practical points still apply:

  • If a service you use is shutting down, exercise your right to delete your data before it closes.
  • Check what the "transfer of business" clause in the privacy policy actually says.
  • Think twice before uploading sensitive data to small services of uncertain lifespan.

None of this offers full protection, but it is the one concrete lesson to draw from this story: a service's lifespan can be shorter than that of the data you gave it.