AI labs are buying failed startups’ data, but the value is unproven

A new market for the leftovers

Asset Hub helps founders sell source code and data from work services, including email, chat and documents. Its buyers include foundation-model labs, providers of simulated work environments for training models, and large technology companies developing their own models.

SimpleClosure expects the sales to return more than $10 million to startups by the end of 2026. CEO Dori Yona says that deals which were once rare and “somewhat taboo” are becoming routine: 47% of SimpleClosure’s deals now include code or data.

Sunset, a competitor, is also seeing rapid growth in data sales. Its founder and CEO, Brendan Mahoney, says the company values assets using factors such as team size, operating history, data volume and how employees used their tools. He compares the data to wood shavings: people assumed there was no value in operational traces, but there is.

Demand surged last year, as AI labs ran into limits on open and synthetic datasets. Surge AI, for example, buys data to build “reinforcement-learning gyms” where customers’ AI agents practice tasks such as debugging complex IT architectures or preparing legal appeals. Nick Heiner, who leads the company’s work on these environments, looks for archives containing varied situations, from disputes to difficult decisions that require judgment.

A company’s Slack history can be more valuable than another company’s Slack, email and calendar data, Heiner says, if the first archive contains richer and more varied conversations.

Value without a benchmark

Real records can help avoid familiar flaws in synthetic environments, such as references to dates that never existed. But no company has published tests comparing models trained on real company data with models trained on alternatives. Prices reflect that uncertainty: buyers offer widely varying sums, Heiner says, based on “research hypotheses” about how many useful tasks might be extracted from, say, 100,000 lines of code.

One deal shows how large the market can get. In August, during Spirit Airlines’ bankruptcy, Google outbid AI-data company Mercor for decades of airline archives. The $10 million purchase included 100 million emails, 500 million Teams messages and 176,000 employee records. Bloomberg reported in September that SpaceX’s AI division was considering buying data from struggling startups.

The archive’s scale is striking, but the missing benchmark matters more. My read is that buyers are paying for the possibility of useful training data before anyone has shown how reliably that possibility translates into better models.

The people inside the archive

The legal risks are less hypothetical. The Spirit Airlines flight attendants’ union opposed the sale over privacy concerns. Greg Ewing, a lawyer who advises AI startups, called the data “legally radioactive.”

Vendors hire data specialists to remove personal information and exclude personnel records, while buyers say they also clean the archives. But emails and calendars can still contain intimate details: medical appointments, plans for who will pick up a child from school, or a date. Ewing says the “Wild West of AI training” is only beginning to give way to proper licensing agreements and checks on how data is used.

The financial gains may also bypass the people whose records make the data valuable. Proceeds typically go to senior creditors and investors, not employees. Ifeoma Ajunwa, a law professor at Emory University and author of The Quantified Worker, describes the practice as a form of “captured capital.” Legal precedent can justify monitoring workers when it helps them do their jobs; once a company closes, that rationale disappears. Ajunwa argues that employers should get explicit employee consent before selling their data, but no federal law prohibits such deals.

I think the central tension is not whether abandoned company data has value, but who gets to decide what happens to it. Founders often know little about who is buying or what they plan to do, and there is no guarantee the price is fair. Prices may stabilize next year—unless the AI bubble bursts first. Either way, the new technology funeral homes will keep pricing what remains.