Sources: Mercor and other firms gathering data for AI labs are driving demand to buy or license internal datasets from startups shutting down or being acquired
Eight days after AI agent startup Warmly in late June agreed to be acquired by HubSpot, CEO Maximus Greenwald found an unusual email in his inbox.
Context & Ripple Effects
Mercor had already tested the boundaries of AI-training supply by asking professionals to sell prior work materials, including material that might belong to former employers, in its outreach for professionals' work materials. Separately, AI labs were reported to be acquiring workplace archives from defunct startups to create simulated work environments for agents.
The new demand reaches into the deal and shutdown process itself: internal startup records are being treated as a licensable training asset. That matters for Mercor because foundation-model makers accounted for roughly 91% of its first-half revenue in Mercor's reported H1 revenue mix.
First-order effects
- Startups that are shutting down or being acquired gain a new potential buyer or licensee for internal datasets, while Mercor and similar firms gain another channel for sourcing agent-training material.
- HubSpot's acquisition of Warmly places Warmly's internal data alongside the assets being solicited, making control of those records relevant to the transaction rather than solely to product integration.
Second-order effects
- Mercor's rival micro1 and other data suppliers face pressure to secure comparable sources of proprietary work data as labs seek material beyond contractor-produced datasets.
- Foundation-model makers can procure more realistic workplace data through intermediaries such as Mercor, increasing the value of firms that can package and license fragmented startup records.
Third-order effects
- If acquisition and shutdown data becomes a repeatable supply channel, startup internal records may become a distinct asset class in AI-data procurement, with intermediaries rather than individual startups organizing access for labs.
- The pattern shifts competition in agent training toward exclusive access to operational data, reinforcing the leverage of well-funded data brokers serving concentrated foundation-model buyers.
The trend: AI labs are moving from broadly sourced training material toward procuring proprietary operational data that can train agents in workplace-like environments.