New York-based AI startup Deeptune has raised $43 mn in a Series A round led by Andreessen Horowitz with participation from 776, Abstract Ventures, and Inspired Capital, and angels including Noam Brown (Research, OpenAI), Brendan Foody (CEO, Mercor), and Yash Patil (CEO, Applied Compute), as it works on one of the harder problems in artificial intelligence development: how to train agents for practical, real-world work once easy access to strong public data starts drying up.
Large language models still handle information well, though they often fall short when asked to carry out multi-step work inside digital systems. That gap between knowing and doing is becoming more expensive.
Chief executive officer Tim Lupo said the fresh capital will help the company scale its effort to address what the industry increasingly describes as data exhaustion.
Deeptune is trying to close it by building what it describes as training gyms for AI. These are high-fidelity virtual environments designed to mirror professional digital workspaces, including those used by software engineers and customer support teams.
Inside those simulated settings, AI agents can learn by doing rather than only by reading or predicting text.
The company’s model relies on reinforcement learning. Agents work through tasks by trial and error inside the simulation, guided by a reward system that pushes them toward better ways of completing workflows. Over time, that process generates a large stream of proprietary training signals.
Lupo compared the environments to flight simulators for AI, making the point that practical training matters if the goal is genuine competence. In that framing, data collection stops being a scraping problem and starts looking more like an engineering and compute problem.
“Our team is a small, focused team of engineers and operators from Anthropic, Scale AI, Palantir, Modal, Glean, Retool, and Hebbia. If you want to work on the hardest (and weirdest) problems of our time, we encourage you to reach out,” Lupo said.
Deeptune’s ambition is to let AI agents learn any professional task that can be modelled in a simulated environment.
The funding round also drew backing from 776, Abstract Ventures, and Inspired Capital, alongside angel investors with links to organisations including OpenAI.
The investor lineup reflects growing belief that training infrastructure, not only models, will shape the next leg of AI development.
That belief ties into a broader market shift. The reinforcement learning market is expected to expand from $11.6 bn in 2025 to more than $90 bn by 2034, and Deeptune is positioning itself in the middle of that curve.
According to Beinsure analysts, the company is selling more than tooling. It is selling a different way to think about how AI gets experience.
Andreessen Horowitz partner Marco Mascorro said Deeptune’s technology has already produced clear gains for AI agents on industry benchmarks. In his view, the last decade of AI was shaped by better datasets.
With the new capital, Deeptune plans to expand its team of 20 engineers and researchers and speed up product development and deployment.
The company said it has already built hundreds of virtual training gyms for some of the world’s largest AI labs, which suggests the demand for simulation-led training is already there, not off in the distance.
The company’s New York base is also part of the strategy. Lupo said the city gives Deeptune access to a broad and deep talent pool, which matters when the company is trying to hire engineers and researchers willing to work on frontier AI problems. As the business grows, that hiring edge could matter as much as the funding itself.
