AI labs buy dead startups' code and Slack archives as training data

A new market for dead companies' data

When Keelan Miskell began shutting down his crypto startup this year, he assumed its digital remains were close to worthless. The users were anonymous and the company was only a few years old. Instead, he has collected about $25,000 for the code base and stands to receive roughly $15,000 more for workspace data spanning Google, Notion, Figma and Slack, according to MIT Technology Review.

The sale was handled by SimpleClosure, one of several firms that now cater to startups at the end of their lives. Carta and AngelList estimate that shutdowns have been climbing between 26% and 56% a year in recent years, and a whole end-of-life ecosystem, from coaches and "death doulas" to firms that sort out the legal and tax loose ends, has grown up around it. The newest service on the menu: selling what is left to AI labs.

SimpleClosure added data sales in April through its Asset Hub platform, which helps founders sell code and workspace data such as emails, chats and documents. Its buyers include foundation model labs, vendors that build simulated workspaces for AI training, and established tech companies training their own models. The platform has handled more than 150 deals and expects to return over $10 million to startups by the end of 2026. CEO Dori Yona says such transactions were once rare and "somewhat taboo" but have become routine; 47% of SimpleClosure deals now include a code or data sale.

Competitor Sunset runs a fast-growing resale business of its own. CEO Brendan Mahony tells MIT Technology Review that his firm values a dying client's assets by head count, years in operation, data volume and how the team used its tools, the operational exhaust that companies leave behind, which he compares to scrap material everyone assumed was worthless until it proved otherwise.

Why labs want real companies' archives

Demand appeared abruptly last year, as AI training labs ran into the limits of publicly available and synthetic datasets. Surge AI, for instance, buys archives to build "reinforcement-learning gyms" where client AI agents practice tasks like debugging a complex IT setup or drafting an appellate brief. Nick Heiner, who leads that work, says he looks for records reflecting a range of situations, including disagreements and difficult judgment calls, and that one company's Slack history can be worth more than another's Slack, email and calendar data combined.

Real archives also avoid common synthetic-data mistakes, such as references to dates that do not exist. Even so, no company has published tests comparing models trained on real-world data against the alternatives, and that uncertainty shows up in pricing. Heiner says buyers quote wildly different figures because valuations rest on "research theses" about how many useful tasks, say, 100,000 lines of code will yield.

The deals are getting bigger

In August, as part of Spirit Airlines' bankruptcy, Google outbid the AI data company Mercor and paid $10 million for decades of the carrier's data, including 100 million emails, 500 million Microsoft Teams messages and 176,000 employee records. Bloomberg reported in September that SpaceX's AI unit is considering buying data from struggling startups.

Legally radioactive

The Spirit sale drew an objection from the flight attendants' union on privacy grounds, and Greg Ewing, an attorney who advises AI startups, calls such archives "legally radioactive." Suppliers employ data scientists to strip out personally identifiable information and exclude HR records, and buyers say they scrub too, but Ewing points out that emails and calendars still hold details like doctor appointments, school pickups and date-night plans. He describes a "Wild West of AI training" that is only beginning to give way to proper licensing agreements.

There is also the question of who benefits. Proceeds typically flow to senior creditors and investors rather than employees. Ifeoma Ajunwa, a law professor at Emory University and author of The Quantified Worker, calls this "captured capital": monitoring workers was defensible while it helped them do their jobs, but that rationale disappears once a company is gone. She argues employers should obtain express consent before selling workers' data, while noting that no federal law in the US bans such sales.

Founders face their own caveats. They have little visibility into who is buying the data or what it will be used for, and little assurance that any offer is fair. Prices may stabilize in the coming year, MIT Technology Review suggests, unless the AI bubble pops first.

Why it matters

A new commodity market is forming in real time, priced by guesswork and governed by almost nothing. Most of the raw material is ordinary workplace communication produced by people who never agreed to have it sold and who will see none of the proceeds. As long as labs need training data that resembles real work, pressure will keep building on archives like these, and the regulatory gaps around privacy and consent are where the risk sits. It also changes how the AI data supply should be judged: scraping the public web and generating synthetic sets are no longer the only options, and the economics of startup failure now include a data resale.