General Intuition Raises $320M to Train AI Agents Using Video Game Action Data

General Intuition, a New York based startup, has raised $320 million in a Series B round led by Khosla Ventures, bringing its valuation to $2.3 billion and total disclosed funding to $454 million. The company is betting that the action labels embedded in millions of hours of video game footage can teach AI agents spatial-temporal reasoning, a skill it argues is essential for real world robotics.

The funding was announced on June 25, 2026, and confirms prior reports from TechCrunch. Participants include General Catalyst, Jeff Bezos, Eric Schmidt, former Formula 1 champion Nico Rosberg, and researchers at Google DeepMind and MIT.

Gameplay as a training shortcut

General Intuitions approach differs from most competitors because it uses not only video footage but the exact button presses a player made. CEO Pim de Witte, who also founded the game clip sharing platform Medal, argues that most AI models try to infer actions from video alone, which he calls insufficient. The companys model, trained on hundreds of millions of hours of Medal uploads, learns how objects like walls, ladders, and shadows behave across time.

In a demonstration for TechCrunch, a quadruped robot powered by the same model navigated an office after only eight minutes of fine tuning on real world data collected outdoors. The model also powers an agent playing Fortnite for 100 consecutive hours, with the same neural network controlling both the game character and the physical robot.

• General Intuition was spun out of Medal in late 2024 and raised $134 million at launch.
• The company has a compute deal with CoreWeave and will use most of the new capital to scale pre training of its next model version.
• Khosla Ventures partner Vinod Khosla said the bet is that action data will produce a quantum leap in AI intuition, similar to how reasoning emerged in large language models.
• General Intuition plans to make its API broadly available by the end of summer 2026.
• The company sees its world model, which generates environments frame by frame, as an internal training gym rather than a product.

Challenges of scaling from virtual to physical

The approach has attracted prominent backers, but scaling a model trained primarily on gameplay to reliable operation in the physical world has not yet been proven at scale. Most comparable methods require large, expensive real world datasets, which General Intuitions strategy aims to bypass by using gameplay as a proxy. De Witte argues that the action labels embedded in game clips help the model distinguish the self from the environment, giving it a richer understanding of causality.

General Intuition faces competition from other startups and labs working on world models and embodied AI. The company plans to focus its capital on expanding compute capacity and pre training the next version of its model, with the API opening to broader access later this year.