AI labs buying startup archives: who, why, how much
The short answer
AI labs buy startup archives to train agents on real workplace records. SimpleClosure processed nearly 100 deals in a year at $10,000 to $100,000 per company, and Mercor made offers up to $300,000. All figures are company claims or press reports, not guaranteed payouts.
So who is actually writing the checks? Mostly data vendors and agent builders. The press names few labs outright.
Why do labs want a dead startup's records?
Gizmodo reported that buyers build "RL gyms." These are simulated workplaces where an AI agent practices tasks such as planning a birthday celebration for a coworker.
Futurism reported that the data is valued for its realistic workplace noise. That noise makes agent tests look like a real company.
Fortune wrote that AI companies call the record of steps in a workflow a trajectory. My page on what AI labs do with company data explains the uses Epoch lists. Why labs want real data covers the synthetic alternative.
Who is buying?
Mercor approached Warmly, an AI agent startup, about eight days after Warmly announced its acquisition by HubSpot in late June 2026.
Warmly's founder declined (BigGo). BigGo says Mercor targets Slack logs, GitHub records, Asana content, Google Drive files and meeting transcripts.
Bloomberg, relayed by Yahoo Finance, reported that SpaceX's AI unit was weighing purchases. SpaceX did not respond to a request for comment.
Gizmodo reported that Anthropic leadership discussed spending up to $1 billion on RL gyms. I found no archive purchase by Anthropic named.
How do founders license their data?
There are three routes:
- a shutdown service such as SimpleClosure's Asset Hub (launch post)
- a listing board such as BuyMyDeadStartup
- approaching a vendor directly
My page on licensing startup data compares the routes. It also lists the buyers that take a company still operating. The shutdown checklist puts the data decision before you close accounts, and I'd keep it there.
What do critics say?
Marc Rotenberg of the Center for AI and Digital Policy told Gizmodo the privacy issues are "quite substantial." Futurism reported that privacy experts question anonymization claims. They say employee data stays identifiable. See re-identification risk.
What is not shown?
Epoch's review of the market prices a website clone near $20,000, a clone of a product like Slack near $300,000 and a task at $200 to $2,000.
It doesn't say real company data fills those environments. So the link from a startup archive to a training run is a stated plan from buyers. I'd keep that gap in mind with every number above.
Talk to a corporate or privacy lawyer in your state before you sign. The rights check and risk map list where deals break.
Our only referral relationship is with micro1, an AI data company. It may pay us a fee if a company we introduce goes ahead with it. You pay nothing extra. How we are paid
Vaultfish helps you run the process. We are not lawyers and this is not legal advice. If you need counsel who knows this area, we can point you to one. Ask for a referral
