A startup that spends two and a half years failing quietly may turn out to have one last product to sell: its own paper trail. Keelan Miskell, winding down a crypto company this year, assumed his archives were worthless. The users were anonymous and the firm was only a few years old. He has since collected roughly $25,000 for the code base, plus a pending payment, likely around $15,000, for the workspace data sitting in Google, Notion, Figma and Slack. His verdict: at least the work was worth something to someone somewhere.
That someone is increasingly an AI lab. A small industry has grown up to broker the transaction, and it is, depending on your mood, either a sensible recycling program or a slightly grim estate sale.
The undertakers have a product line
Startup deaths are a growth market. Estimates from Carta and AngelList show shutdowns rising between 26% and 56% a year in recent years. Around that decline has formed an end-of-life ecosystem of coaches, "death doulas," and firms that deal with the legal, tax and operational mess of going belly-up.
The newest service on the menu is selling the remains to AI companies.
SimpleClosure added data sales in April through a platform called Asset Hub. Founders can sell their code and workspace data, meaning emails, chats and documents. Buyers include:
- Foundation AI labs
- Vendors that build simulated workspaces for AI training
- Mature tech companies building their own models
Asset Hub has handled more than 150 deals and expects to return over $10 million to startups by the end of 2026. CEO Dori Yona says these sales used to be rare and "somewhat taboo." Now 47% of SimpleClosure's deals include a code or data sale.
Sunset, a rival, also has a fast-growing resale business. Founder and CEO Brendan Mahony says it values a client's assets by head count, years in operation, volume of data, and how the team actually used its tools. His preferred analogy is particle board: "operational exhaust data" is the wood shavings everyone assumed were trash.
Why a dead company's Slack is suddenly valuable
Demand appeared abruptly last year, when AI training labs ran into the limits of public and synthetic datasets. Public data has been mined, and synthetic data has a tendency to invent things.
Surge AI is one buyer. It uses data to build what it calls "reinforcement-learning gyms," environments where clients' AI agents practice jobs such as debugging a complex IT architecture or drafting appellate briefs. Nick Heiner, who leads that work, wants archives that show a range of situations, from disagreements to tricky judgment calls. Variety and depth matter more than volume. In his telling, one company's Slack history, if the conversations are rich enough, can be worth more than another company's Slack, email and calendars combined.
Real archives also avoid some of the classic blunders of synthetic worlds, such as references to dates that do not exist. A fake office, it turns out, is hard to fake.
Nobody knows what it is worth
Here is the awkward part. How much real-world data actually improves a model remains unclear, and no company has published tests comparing training on real data against the alternatives.
Pricing behaves accordingly. Heiner says buyers quote wildly different numbers, because each valuation rests on a "research thesis" about how many useful tasks you can get out of, say, 100,000 lines of code. A thesis is a polite word for an educated guess, and the guesses are being paid out in cash.
The guesses can get large. This August, in Spirit Airlines' bankruptcy, Google outbid the AI data company Mercor for the carrier's decades of data. The haul included 100 million emails, 500 million Teams messages and 176,000 employee records. The price was $10 million. A September report also said SpaceX's AI unit is weighing purchases of data from struggling startups.
The legal footing is soft
The Spirit sale drew an objection from the airline's flight attendants' union on privacy grounds. Greg Ewing, an attorney who advises AI startups, has called this kind of data "legally radioactive."
Sellers say they take precautions. Suppliers employ data scientists to scrub personally identifiable information and to exclude HR records, and buyers say they scrub as well. Ewing's concern is what survives that process. Emails and calendars hold doctor appointments, school pickup arrangements and date-night plans, the sort of detail no one files under "business information." He describes a "Wild West of AI training" that is only beginning to give way to proper licensing agreements and vetted use.
Who gets paid
Not the people who wrote the messages. Proceeds typically go to senior creditors and investors, not employees.
Ifeoma Ajunwa, a law professor at Emory University and author of The Quantified Worker, calls this a form of "captured capital." Her argument is that the law tolerates workplace monitoring because it helps people do their jobs. Once the company is gone, that rationale goes with it. She says employers should get express consent before selling workers' data, and also notes that no federal law bans such sales.
What founders are signing up for
For a failing company, a sale can recover some value from an otherwise dead asset. Miskell's experience shows the upside is real, if modest.
The limits are also real:
- Little visibility. Founders often have almost no idea who is buying the data or what they plan to do with it.
- No fairness check. There is little assurance that an offer is a good one, which is unsurprising in a market where buyers cannot agree on a price.
- Uncertain future. The expectation is that prices will settle over the coming year, unless an AI bubble pops first.
Either way, the report that prompted this piece closes on the obvious point: the undertakers will be there to appraise what is left. Startups were told to leave a legacy. Few expected it to be a searchable archive of their arguments about the logo.