Mercor announced on July 9 that it acquired Deeptune, the San Francisco startup building simulated office software environments for AI agent training. According to Startup Fortune, the deal came three months after Mercor CEO Brendan Foody personally invested in Deeptune’s $43 million Series A, which was led by Andreessen Horowitz.
What Deeptune Built
Deeptune created what it calls training gyms: reinforcement-learning environments that replicate the software knowledge workers already use. An accountant’s Excel workbook, a support rep’s Salesforce queue, a team’s Slack thread for resolving a customer issue. The point, according to Startup Fortune, is to let AI agents “click, type, make mistakes, and learn without touching a real company’s systems.”
This is infrastructure for proving agents can do actual work, not just pass benchmarks. Model labs and enterprise buyers increasingly need evidence that an agent can navigate messy business software before deploying it in production.
Why Mercor Wanted It
Mercor already operates a network of more than five million domain experts who write tasks, judge AI outputs, and build grading rubrics. Fortune reported, as cited by Startup Fortune, that Mercor’s work “touches every Magnificent Seven company except Tesla.” What Mercor lacked was the environment where those tasks could actually execute.
Deeptune fills that gap. Mercor now has both the human grader who knows what correct work looks like and the simulated workplace where an agent practices that work. That combination is more defensible than either piece alone, and it directly serves OpenAI, Anthropic, and other labs that need agent training infrastructure going beyond static prompt evaluation.
Deeptune’s team is relocating to Mercor’s New York office, according to Startup Fortune. Neither company disclosed the acquisition price.
The Governance Question
Foody told Fortune that his personal angel check in Deeptune’s March round was “in a lot of ways the main motivation” for the deal, Startup Fortune reported. A CEO who personally profits from a company his own startup subsequently acquires creates a related-party concern. The public reporting does not clarify whether Mercor’s board or outside investors approved Foody’s March investment before he wrote the check, or how any personal gain was handled.
The stakes are not trivial. Bloomberg reported this week that Mercor is in early talks to raise new funding at roughly a $20 billion valuation, double the $10 billion it reached after closing a $350 million Series C in October, as reported by Startup Fortune. Foody said on X that Mercor’s annualized revenue run rate hit $2 billion in June, up from $1 billion four months earlier.
Training Data Market Shift
The acquisition signals a broader shift in how AI agent capability is validated. The training-data market is moving from human labeling of model outputs toward full work simulations where agents must complete tasks end to end. Scale AI, Surge AI, and other human-data companies compete for the same budgets, but the spending is migrating toward infrastructure that can demonstrate whether an agent can finish a real job, not just generate a plausible answer.
Deeptune raised a large Series A in March and was absorbed by July. That timeline reflects a market where the missing piece, task simulation infrastructure, is considered more valuable inside a platform with distribution than as a standalone product.