General Intuition Raises $320 Million to Train AI Agents on Video Game Data

A startup that trains AI agents on video game footage has raised $320 million in a Series A round led by Khosla Ventures, valuing General Intuition at $2.3 billion — less than a year after it emerged from stealth with a $134 million seed round.

The round, first reported by TechCrunch, also includes General Catalyst, Jeff Bezos, Eric Schmidt, Formula 1 driver Nico Rosberg, and researchers from Google DeepMind and MIT, bringing General Intuition's total disclosed funding to $454 million.

The Insight: Games Know What Happened, Not Just What It Looked Like

Most AI agent training relies on video prediction — models watch footage and try to infer what actions produced what outcomes. General Intuition's approach is different. The company built its training pipeline on top of Medal, a gameplay-capture platform with 17 million monthly active users, which records not just video but the explicit action labels behind every frame: the precise button presses, analog stick positions, and timing data that produced each movement.

That distinction matters enormously for learning causality. A model trained on video alone has to guess whether a character jumped because of user input or because it hit a spring. A model trained with Medal's action labels knows exactly who did what, and when. According to CEO Pim de Witte, this richer causal signal is what gives General Intuition's models a more grounded understanding of the "self vs. environment" distinction — a capability he argues is essential for agents that need to act reliably in unstructured settings.

One Model for Games and Robots

The company's technical ambition is a single foundation model that can control both virtual agents in game environments and physical robots in the real world. The spatial-temporal reasoning developed through gameplay — understanding how objects move through space, how sequences of actions create outcomes over time, how environments respond to inputs — transfers, the company argues, to physical manipulation tasks where robots face the same underlying challenges.

This positions General Intuition alongside a growing cluster of companies attacking the robot foundation model problem from different angles: Figure AI and Physical Intelligence are training on physical robot data; World Labs and others are building world models from internet video. General Intuition's bet is that gaming data, with its explicit action annotations and enormous scale, offers a faster and cheaper path to the same destination.

What the $320 Million Buys

The majority of the new capital is allocated to scaling compute capacity through a partnership with CoreWeave, the GPU cloud provider that has become a major infrastructure partner for AI labs. A significant portion will fund pre-training the company's next model version. The remainder is earmarked for expanding API access — General Intuition plans to open its models to external developers by the end of summer 2026.

The funding round's size reflects both the capital intensity of training large models and investor confidence that the gaming-data approach can compete with pipelines built on more traditional robotics or internet data. Whether a model trained on Medal gameplay can generalize to the full complexity of physical manipulation remains the central technical question — and the one the next model version is designed to answer.