Artificial intelligence startup Deeptune Inc. said today it’s going to solve the industry’s “data exhaustion” crisis after raising $43 million in an early-stage funding.

Andreessen Horowitz led the Series A round, which saw participation from 776, Abstract Ventures and Inspired Capital, plus high-profile angel investors such as OpenAI Group PBC researcher Noam Brown and Mercor.io Corp. Chief Executive Brendan Foody.

New York-based Deeptune is building what its CEO Tim Lupo (pictured) calls “training gyms” for AI agents. In an interview with Fortune, which first reported the funding, he explained that the company creates high-fidelity reinforcement learning environments that simulate the complex digital workspaces of professionals such as DevOps engineers, accountants and customer support representatives. The idea is that these virtual gyms can be used to train agents to automate many of the complex, multistep tasks that are performed by these professionals, using software tools such as Salesforce and Slack.

Lupo likens these environments to “flight simulators” for AI, and argues that they’re going to be essential for teaching agentic systems to step up to the plate. He compared AI agents to human pilots, and stressed that no one learned to fly an airplane by reading books alone. What they need is hands-on training in a realistic scenario where they can practice.

The round comes at a time when many in the AI industry, including Andreessen Horowitz founder Marc Andreessen, are warning that the supply of freely available internet data for training agentic systems is close to running out. Simply put, everything that’s available online to be scraped has already been scraped, and it’s still not enough to train AI agents that can be trusted to just get on with things. This is the issue Deeptune says it’s trying to address, by transforming data collection from a labor-intensive task into an engineering and compute problem.

Rather than just scraping text from websites, AI models can run what Lupo terms “rollouts” within Deeptune’s simulated environments, where they can practice the tasks they’re assigned to do. The agents will receive virtual “rewards” when they get it right, providing an incentive to learn the most effective way to go about completing various tasks. According to Lupo, the platform generates high-quality training signals through trial-and-error, and it does it at massive scale – more than enough to overcome the data exhaustion problem.

Andreessen Horowitz partner Marco Mascorro said in a blog post that Deeptune’s approach has led to a marked improvement in the performance of so-called “computer-use” agents on industry benchmarks. He said the company’s approach to training has led to significant gains, allowing advanced models to operate computers through desktop and command-line interfaces just as well as humans can.

“While there is still significant room for improvement in computer use benchmarks, SOTA models have made rapid progress on computer use over the past year,” Mascorro wrote. “Opus 4.6 scores 72.7% on OSWorld, surpassing the human baseline of 72.4%, and GPT-5.4 reaches 75%.”

Deeptune’s investors are confident in the company’s growth, pointing to studies that estimate the global reinforcement learning market will grow from $11.6 billion in 2025 to more than $90 billion by 2034. The startup intends to grab a big slice of that pie. It claims to have already built hundreds of its virtual training gyms for some of the world’s leading AI labs, demonstrating that almost any professional task performed by humans can be distilled into a simulation and then mastered by AI models.

“What makes Deeptune particularly compelling is that they are not just building tools, they are helping define a new paradigm for how AI models are trained,” Mascorro said. “If the last decade of AI progress was driven by better datasets, the next decade will be mostly driven by better environments.”

Lupo told Fortune he’ll use the funding to expand Deeptune’s team of 20 engineers and researchers. He said he’s not worried about the company’s ability to attract talent. He made a deliberate choice to base the company in New York, because the location gives it an edge in terms of talent recruitment. “If you want to be in New York and you want to work on frontier AI or AGI, Deeptune is one of only a couple places you could join, and probably the only early-stage place you could join,” he said.

Photo: Deeptune