AI labs have found a new source for training models: the digital archives of dying startups. According to a detailed review published by MIT Technology Review on October 13, 2026, code, email correspondence, work chats, and documents that nobody once needed have become a product that AI companies are paying for. SimpleClosure, a company that handles the formalities of shutting businesses down, added data sales to its services in April: more than 150 deals have been struck through its Asset Hub platform, and over $10 million is expected to be returned to startups by the end of the year. In August, Google offered $10 million for the data of bankrupt airline Spirit Airlines; Bloomberg reported that SpaceX's AI division is discussing buying data from struggling startups.

SimpleClosure: From Shutdown Services to Data Sales

SimpleClosure is a company that helps shutting-down startups with legal, tax, and operational matters. In April 2026, it added data sales to its services: the Asset Hub platform helps founders sell code and working data — emails, chats, and documents. Buyers include leading AI labs, vendors that build simulated work environments for AI training, and large technology companies developing their own models.

According to the company's CEO, Dori Yona, more than 150 deals have been closed through Asset Hub, and over $10 million is planned to be returned to startups by the end of 2026. Such deals, once rare and considered 'somewhat taboo', are becoming commonplace: 47% of SimpleClosure's deals now include a code or data sale.

A living example of such a deal is founder Keelan Miskell, who is shutting down his startup. He received about $25,000 for his codebase; another roughly $15,000 is expected for his archive of working data from Google, Notion, Figma, and Slack (not yet paid).

Competitors and the Biggest Deals

SimpleClosure is not alone in this market. Its competitor Sunset also runs a fast-growing data-resale business. According to the company's founder and CEO, Brendan Mahony, Sunset evaluates the assets of closing clients against four criteria: headcount, the company's years of operation, data volume, and how the team used its tools.

One of the largest deals took place in August 2026: as part of Spirit Airlines' bankruptcy, Google outbid the AI data company Mercor with a $10 million offer for the airline's data accumulated over decades. The haul includes 100 million emails, 500 million Teams messages, and 176,000 employee records; the goal is to improve the company's products and AI models.

According to a September 2026 report by Bloomberg, SpaceX's AI division (SpaceXAI) held informal internal discussions about buying customer and operational data from struggling or closed startups. The talks are at an early stage and may not lead to a deal.

The legal side of the market is murkier, though. The union of Spirit Airlines' former flight attendants objected to the Google deal on privacy grounds, and a U.S. bankruptcy court postponed approving the purchase.

Greg Ewing, a lawyer who advises AI startups, called such purchased corporate data 'legally radioactive':

'Such purchased corporate data is legally radioactive,' says Greg Ewing, a lawyer who advises AI startups.

He notes that emails and calendars hide confidential details — doctor's appointments, plans to pick up children from school, personal meetings. Suppliers hire data specialists to scrub identifiable information and exclude HR documents, and buyers say the scrubbing happens — but the risk never fully disappears.

Another question is who gets the proceeds. Typically, money from shutdown sales goes to large creditors and investors, not employees. According to Emory University law professor Ifeoma Ajunwa, this is a form of 'captured capital': 'Employers should obtain employees' explicit consent before selling their data,' but no federal law bans such sales.

Where the Demand Comes From

Demand for the data of shuttered companies emerged sharply last year — AI training labs hit the limits of open and synthetic data. For example, Surge AI buys such archives to build 'reinforcement-learning gyms': clients' AI agents practice tasks there like debugging a complex IT architecture or drafting an appeal complaint.

Nick Heiner, who leads this effort at the company, says he looks for archives that capture a wide range of situations — from disagreements to difficult decisions. In his view, the Slack history of a single company, rich in diverse and deep conversations, can be worth more than the combined value of another company's Slack, email, and calendars.

Still, it remains unclear how much of an edge real archives actually hold over synthetic environments: no company has published tests comparing real-data training results against alternatives. Prices are equally uncertain — buyers name wildly different sums based on 'research theses', for instance estimating how many useful tasks can be extracted from 100,000 lines of code.