Data Marketplaces & Intermediaries

Mercor is buying its way into RL environments

Source: RL List · Sep 23, 2026

Mercor, one of the largest players in the AI-data market, is acquiring its way from labeling into reinforcement-learning environments. It bought Sepal AI in February 2026 and Deeptune in July — the latter a computer-use and enterprise-environment platform that came with a New York team — and in September published a guide to training a 397-billion-parameter knowledge-work agent on the open-source SkyRL framework. It is reportedly raising around $500 million at a roughly $20 billion valuation, though that figure is unconfirmed.

The strategic read is clear: the data leaders no longer see their moat as labeling throughput. They are buying the ability to build interactive environments — sandboxes with expert rubrics where agents can act and be scored — because that is where the next generation of training value sits. Owning the environment, and the expert judgment encoded in it, is the asset.

For a business built on structured, high-context human data, this is both validation and a signal of where demand is heading. The most valuable inputs are increasingly realistic, interactive, expertly-graded records of how people actually perform tasks — which is exactly what a rigorous interview, captured and structured, represents.

Key Points

• Mercor acquired Sepal AI (Feb 2026) and Deeptune (July 2026, a computer-use / enterprise-environment platform plus a NYC team)
• It published a September guide to training a 397B-parameter knowledge-work agent on the open-source SkyRL framework
• Reportedly raising ~$500M at a ~$20B valuation (unconfirmed)
• Signals that owning RL-environment capability — not just labeling labor — is where the data leaders are investing