Overview
Deeptune builds simulated software environments where AI agents can practice tasks. Its customers include frontier AI research organizations that need more than a static set of text examples. In March 2026, the company announced a $43 million Series A led by Andreessen Horowitz. Mercor acquired Deeptune in July 2026 to combine the environments with its network of human experts.[1][2][3]
Founding Story
Founder and CEO Tim Lupo worked on a problem facing teams that train AI agents: a model may know how a spreadsheet or business application works without being able to complete a sequence of tasks inside it. Deeptune set out to supply environments that resemble software used at work. The company said in March 2026 that it had begun this work about a year earlier; other profiles give an earlier founding year.[1][4]
Timeline
- 2025: Deeptune later described the start of its environment work as roughly one year before its March 2026 funding announcement.[1]
- March 2026: Deeptune announced its $43 million Series A, led by a16z. Mercor CEO Brendan Foody was among the individual investors.[1]
- July 2026: Mercor announced its agreement to acquire Deeptune. Fortune reported the deal as completed, and Deeptune's counsel confirmed its role in the acquisition.[2][3][5]
What They Built
Deeptune calls its products training gyms. Each is a simulated environment where an agent can interact with software, attempt tasks, and receive a signal about its performance. The company says it built hundreds of environments resembling applications such as Slack and Salesforce, with problems, datasets, and supporting infrastructure.[6]
The environments support reinforcement learning: a training process that uses feedback from actions to improve behavior. Realism and scoring matter together. An agent must encounter enough of a real workflow to learn useful behavior, and the evaluator must distinguish a correct result from a convincing but wrong one. a16z reported that Deeptune worked with leading labs and described improvements on computer-use benchmarks; the investor did not provide a controlled attribution of those gains to Deeptune alone.[4]
Market Position
Target Customers
Deeptune targeted frontier AI labs training and evaluating models. Mercor was already a customer, according to its CEO. That relationship gave the eventual acquirer direct experience with the product before the transaction.[2]
Market Size
The company's statement that it built hundreds of gyms describes its library, not the number of paying customers or its revenue. The public funding announcement does not disclose those figures.[6]
Competition
Labs can build their own environments or buy training and evaluation services. The difficulty for any supplier is keeping simulations, tasks, and scoring useful as both enterprise software and model capabilities change. Mercor's acquisition joined an environment supplier to an organization that also assembles human expertise for AI work.[2]
Business Model
Deeptune supplied environments and supporting infrastructure to research labs. It did not disclose pricing, margins, or revenue in its funding announcement. The $43 million round funded the business; it was not a measure of sales.[1]
Outcome and evolution
Mercor said Deeptune's team would join it and that the combination would connect software environments to a network of subject-matter experts. The buyer's announcement used prospective acquisition language; Fortune reported the acquisition as completed, and counsel described representing Deeptune in the deal. The terms were undisclosed.[2][3][5]
The company's site continued to describe its gyms after the announcement. The acquisition is a change of ownership and team context; it does not establish a shutdown of the environment product.[6]
Key Lessons
- Practice requires a world and a score. A realistic interface is useful only if the task outcome can also be evaluated.
- A customer can become a strategic buyer. Mercor used Deeptune's work before announcing the acquisition.
- Count the right unit. Hundreds of simulated environments describe product breadth, not customer count or revenue.