Dead Startups' Data Is Becoming AI Training Fuel

Keelan Miskell netted roughly $25,000 selling his defunct crypto startup's code base, plus a likely $15,000 more for its Google, Notion, Figma and Slack workspace data. The buyers were AI training operations, and the deal was brokered by SimpleClosure, a startup wind-down firm that now runs a marketplace for the digital remains of dead companies.

"At least all that stuff we were doing for two and a half years is worth something to someone somewhere," Miskell told Wired.

His sale is a small transaction in a market that has scaled fast. SimpleClosure's Asset Hub platform has handled more than 150 deals since the firm added data sales to its services in April, and the company expects to return over $10 million to shuttered startups by the end of 2026. CEO Dori Yona says these transactions, once rare and "somewhat taboo," have become routine — 47% of SimpleClosure deals now include code or data sales.

The stakes are straightforward. AI labs have hit the limits of public and synthetic training data, and the operational exhaust of real companies — Slack threads, email archives, internal code — has become a substitute commodity. Whether that substitution produces measurably better models is, notably, a question no buyer has answered publicly.

Why are AI labs buying dead companies' archives?

Estimates from Carta and AngelList show startup shutdowns rising between 26% and 56% annually in recent years. An end-of-life ecosystem has formed around that failure rate: coaches, "death doulas," and firms handling the legal, tax and operational headaches of closing down. Selling the data remains is the newest revenue line.

Buyers include foundation AI labs, vendors that build simulated workspaces for AI training, and mature tech companies training their own models. Surge AI, for one, buys data to build what it calls "reinforcement-learning gyms" — simulated environments where AI agents practice tasks such as debugging a complex IT architecture or drafting appellate briefs.

Nick Heiner, who leads Surge AI's work on these environments, says he looks for archives reflecting a range of situations, from disagreements to tricky judgment calls. One company's Slack history, if it captures conversations with variety and depth, can be worth more than another's Slack, email and calendars combined, Heiner says. Real archives also avoid common synthetic-data errors, such as references to dates that don't exist.

The performance evidence, however, is thin. No company has released tests comparing models trained on real-world data against the alternatives. Pricing reflects that vacuum: Heiner says buyers quote wildly different figures because valuations rest on "research theses" — estimates of how many useful tasks can be extracted from, say, 100,000 lines of code. Takeaway: the market is trading on hypotheses, not benchmarks.

How big do these deals get?

The Miskell-scale sales coexist with eight-figure transactions. This past August, as part of Spirit Airlines' bankruptcy, Google outbid the AI data company Mercor to buy the carrier's decades of operational data for $10 million. The trove included 100 million emails, 500 million Teams messages and 176,000 employee records.

SpaceX may be next. Bloomberg reported in September that SpaceX's AI unit is weighing purchases of data from struggling startups. SimpleClosure's rival Sunset also runs a rapidly growing data resale business. Founder and CEO Brendan Mahony says Sunset values a dying client's assets by head count, years in operation, volume of data, and how the team used its tools.

"This operational exhaust data is like the wood shavings that everyone thought had no value, but it actually has a ton," Mahony says.

Is any of this legal?

Legally, the terrain is unsettled. Spirit's flight attendants' union objected to the Google sale on privacy grounds. Greg Ewing, an attorney who advises AI startups, calls such data "legally radioactive."

Suppliers employ data scientists to scrub personally identifiable information and exclude HR records, and buyers say they scrub too. Ewing points to what still lurks in emails and calendars after that filtering: doctor appointments, school pickup arrangements, date-night plans. He describes a "Wild West of AI training" that is only starting to give way to proper licensing agreements and vetted use.

The consent question is sharper still. Ifeoma Ajunwa, a law professor at Emory University and author of The Quantified Worker, argues the sales represent a form of "captured capital." Legal precedent permits employer monitoring to help workers do their jobs; once the firm is out of business, that justification disappears.

"Employers should have express consent before selling workers' data," Ajunwa says. She adds that no federal law currently bans such sales. Takeaway: the scrubbing promises are procedural, and the legal guardrails do not yet exist.

Who actually gets the money?

The proceeds from these windfall sales typically flow to senior creditors and investors — not to the employees whose messages and documents make up the archive. That distribution pattern is the core of Ajunwa's objection, and it is unlikely to change without legislation or litigation forcing the issue.

Founders, too, operate with limited information. They have little to no visibility into who is buying the data or what the buyers plan to do with it, and they have little assurance the offer on the table is fair. There is no public price index for operational data; the numbers come from buyers' internal models.

What happens next?

Data prices will likely stabilize in the coming year, unless an AI bubble pops first. The source material suggests both outcomes remain live: demand is rising because labs have exhausted cheaper alternatives, but the buyers' own valuation methods are guesses, and the legal challenges — like the Spirit flight attendants' objection — are just beginning to work through courts.

Either way, the undertakers are positioned. SimpleClosure, Sunset and their peers will keep appraising what is left, and the market will keep deciding what a dead company's Slack history is worth. What nobody has shown yet is the number that would justify the prices: how much better an AI agent trained on real corporate exhaust actually performs.