Dead Startups Are Becoming an AI Training Data Source
A growing market lets founders sell their company's digital leftovers to AI labs. The money rarely reaches employees, and the legal ground is shaky.
Key points
- SimpleClosure's Asset Hub platform has completed over 150 deals and expects to return more than $10 million to shutting-down startups by the end of 2026.
- Google paid $10 million for Spirit Airlines' data archive in August 2026, outbidding AI data company Mercor for 100 million emails and 500 million Teams messages.
- Shutdown rates among startups rose between 26% and 56% annually in recent years, according to data from Carta and AngelList.
- Proceeds from data sales typically flow to senior creditors and investors, not the employees whose messages and files are being sold.
When Keelan Miskell shut down his crypto startup this year, he assumed its digital records were worthless. He was wrong. He netted roughly $25,000 for his code and expects around $15,000 more for his team's files from Google Workspace, Notion, Figma, and Slack.
Miskell's experience, reported by MIT Technology Review, is becoming less unusual by the month.
Where is all this data coming from?
Startup closures are the supply side of this new market. Carta and AngelList, two platforms that track startup financing, recorded shutdown rates rising between 26% and 56% per year in recent years. That wave of closures is generating enormous archives of code, emails, internal chat logs, and documents, and AI labs want them.
The broker in Miskell's case was SimpleClosure, a company that helps founders manage the legal and financial headache of winding down a business. SimpleClosure added data sales to its services in April 2026. Its Asset Hub platform now handles deals between shutting-down startups and buyers that include AI foundation labs, companies building private AI models, and vendors who construct practice environments where AI systems learn new tasks.
The value is not just in volume. Nick Heiner, who leads data-environment work at a company called Surge AI, told MIT Technology Review that a single rich Slack archive capturing genuine disagreements and tricky decisions can be worth more than another company's combined emails, messages, and calendar data. Real workplace conversations, he argues, teach AI systems things that synthetic, computer-generated training data cannot.
Industry rival Sunset, another startup wind-down service, uses a formula based on head count, years in operation, and how actively a team used its tools. CEO Brendan Mahony compares the data to particle board: wood shavings nobody thought had value, now in demand.
What does the August Google deal tell us?
The biggest deal on record so far gives a sense of the scale. In August 2026, as part of Spirit Airlines' bankruptcy, Google outbid the AI data firm Mercor to acquire the airline's data archive, which included 100 million emails, 500 million Microsoft Teams messages, and 176,000 employee records. The price was $10 million.
Bloomberg reported in September 2026 that SpaceX's AI unit is also considering buying data from struggling startups, a deal class that AI2Day has been tracking alongside a broader look at AI asset competition.
| Deal or platform | Data type | Scale | Price |
|---|---|---|---|
| Spirit Airlines (Aug 2026) | Emails, Teams messages, HR records | 100M emails, 500M messages | $10 million |
| SimpleClosure Asset Hub (by end 2026) | Code, workspace data | 150+ deals | $10M+ returned to startups |
| Miskell crypto startup (2026) | Code, Notion/Figma/Slack files | Single company | ~$40,000 |
Should employees be worried?
Yes, in a practical sense. Lawyers and academics are flagging real problems. Attorney Greg Ewing, who advises AI startups, describes the current situation as a "Wild West of AI training." Buried inside the emails being sold are doctor appointments, school pickup times, and personal plans that employees never expected to become AI training material.
Ifeoma Ajunwa, a law professor at Emory University and author of "The Quantified Worker," calls the proceeds "captured capital." Under current law, money from these sales goes to senior creditors and investors, not the employees whose daily working lives generated the data. There is no federal law in the United States that bans such sales or requires employee consent.
Suppliers and buyers both say they strip out personally identifying information before any transfer, but the scrubbing is not independently verified.
The honest summary: this market is real, growing fast, and running well ahead of any rules designed to govern it. For employees at a startup that closes, the practical advice right now is simple. Assume your work communications could be sold, because the law does not stop that from happening.
Common questions
Who actually gets the money from selling a dead startup's data?
Typically the company's senior creditors and investors, not its former employees. That is the current legal default in the United States, and no federal law requires it to be otherwise.
Is this data actually useful for training AI?
The buyers believe so, though no company has yet published a study comparing AI systems trained on real workplace data against those trained on synthetic alternatives. Prices vary widely because valuations are still based on guesswork about how much usable training material a given archive contains.
What stops buyers from misusing the data?
Very little beyond contractual promises at this stage. Both sellers and buyers say they remove personal details, but the process is not standardised or independently audited.
