Mercor buys Deeptune for agent training
Mercor, at $2B ARR, acquired a16z-backed Deeptune on July 9 to pair its expert network with Deeptune's enterprise software simulators for AI agent training.
Mercor announced on July 9 that it will acquire Deeptune, a startup that builds software environments for reinforcement learning. Deeptune had raised $43 million in a Series A led by Andreessen Horowitz earlier in 2026. Financial terms of the acquisition were not disclosed.
Mercor runs a network of five million domain experts who build tasks and scoring rubrics for AI model training. Deeptune built the other side of that equation: realistic replicas of enterprise applications, including spreadsheets, CRM systems, and productivity tools, where models can practice tasks before touching production. The two businesses together cover the full RL training stack for enterprise workflows.
Mercor CEO Brendan Foody was a personal angel investor in Deeptune’s $43 million Series A, a detail reported by Fortune. Foody told Fortune that the angel investment was made with acquisition in mind. Mercor reported $2 billion in ARR in June 2026 and is now in talks for a $20 billion valuation, up from $10 billion last October.
What it means for you
The enterprise software simulator layer is an unglamorous but real constraint in agentic AI deployment. Agents that have never seen a real Salesforce environment in training will fail in ways that are hard to predict and expensive to debug. Deeptune’s approach of building full application replicas is the highest-fidelity option, and Mercor is betting that selling this capability as infrastructure to frontier labs is more valuable than building products on top.
This acquisition fits the pattern described in the guide on what happens when your AI lab buys your dev tool: the companies building training infrastructure are getting absorbed by whoever controls the model or the evaluation pipeline.