Failed Startups Are Selling Their Data to AI Companies—But Privacy Risks Remain

SimpleClosure is adding data sales to its services, with the new offering scheduled to launch in April. Its Asset Hub platform helps founders sell company assets, including source code and workplace data such as emails, chats and documents.

Buyers include foundational AI labs, vendors building simulated workspaces for AI training and established technology companies developing their own models. Asset Hub has handled more than 150 transactions and expects to return more than $10 million to startups by the end of 2026.

SimpleClosure CEO Dori Jonah says transactions that were once rare and “slightly taboo” are becoming increasingly common. Forty-seven percent of SimpleClosure transactions now involve the sale of code or data.

SimpleClosure also has competition. Industry rival Sunset is rapidly expanding its data-resale business. Founder and CEO Brendan Mahoney said Sunset values the assets of companies that are shutting down based on factors including employee count, years in operation, data volume and how the team used its software.

Mahoney compares the data to particleboard. “This operational emissions data is like wood shavings that everyone thought had no value, but it actually has a ton of value,” he said.

Why AI companies want data from failed startups

Demand for data from defunct companies emerged suddenly last year as AI training labs began reaching the limits of public and synthetic datasets.

Surge AI, for example, buys data to build “reinforcement learning gyms” where clients’ AI agents practice tasks such as debugging complex IT architectures and drafting appeals. Nick Heiner, who leads the company’s efforts in these environments, said he wants archives that reflect a broad range of real-world situations, including disagreements and difficult decisions.

Heiner said one company’s Slack history may be more valuable than another company’s combined Slack, email and calendar data if it contains a greater diversity and depth of conversations.

Real company archives could improve AI training—but the value is uncertain

Real-world archives can help avoid common errors in synthetic data, such as references to nonexistent dates. However, it remains unclear how much performance improvement they provide. No company has yet released a test comparing AI training results based on real-world data with results based on synthetic or other alternatives.

That uncertainty is reflected in pricing. Heiner said buyers offer widely different amounts because valuations are based on “research papers” estimating how many useful tasks can be extracted from a given amount of data—for example, 100,000 lines of code.

Google paid $10 million for Spirit Airlines data

Uncertainty has not stopped larger deals. In August, Google outbid AI data company Melkor during Spirit Airlines’ bankruptcy proceedings. According to Reuters, the purchase included decades of airline data, 100 million emails, 500 million Teams messages and 176,000 employee records. The price was $10 million.

Bloomberg also reported in September that SpaceX’s AI division was considering buying data from struggling startups.

Privacy and legal risks complicate AI data sales

The legal status of these transactions can be murky. The Spirit flight attendants’ union opposed the sale. Greg Ewing, a lawyer who advises AI startups, also criticized the transaction on privacy grounds and described this type of data as “legally radioactive.”

Suppliers have hired data scientists to remove personally identifiable information and filter out human resources records, while buyers say they are taking similar steps. Ewing noted that emails and calendars can contain highly private information, including doctor’s appointments, transportation arrangements to and from school and dates.

He described the current market as the “Wild West of AI training,” though he said it is beginning to shift toward formal licensing agreements and more carefully vetted data use.

Who benefits when employee data is sold?

There is also the question of who benefits from these transactions. Proceeds typically go to senior creditors and investors rather than employees.

An Emory University law professor who writes about the quantified worker described the trend as a form of “captured capital.” Case law allows workplace monitoring to support employees’ work, but that justification becomes less clear when a company shuts down and sells the data.

“Employers should obtain explicit consent before selling worker data,” Ajungwa said, adding that no federal law currently prohibits such sales.

A growing market for distressed company data

Many observers see the market as an opportunity for distressed companies to recover value from their remaining assets. These sales also have clear limits. Founders often have little visibility into who is buying their data or how it will be used, and there is no guarantee that sellers are receiving a fair offer.

Still, unless the AI bubble bursts first, data prices are likely to stabilize next year. In either case, the technology industry’s new undertaker will be there to assess what remains.

Matthew King is a technology and environmental journalist based in Baltimore.