General Intuition is raising $300 million to train AI agents on the video game data OpenAI tried to buy

General Intuition, a New York-based AI lab that trains agents to reason about space and time using billions of video game clips, is in talks to raise approximately $300 million at a valuation of just over $2 billion, according to TechCrunch. The round would come just eight months after the company launched with a $134 million seed, one of the largest on record.

Backers reportedly include Jeff Bezos, whose own physical AI venture Prometheus recently raised $12 billion, and Eric Schmidt, alongside existing investors Khosla Ventures and General Catalyst.

The Medal origin story

General Intuition exists because of a deal that fell apart. Late in 2024, OpenAI reportedly offered $500 million to acquire Medal, a platform where gamers upload and share short video clips of gameplay, because it wanted the data to train its models, according to The Information.

Medal’s founder, Pim de Witte, turned it down. Instead he spun out General Intuition in October 2025 with co-founders Eloi Alonso, Adam Jelley, and Vincent Micheli, three researchers who had previously published influential work on world modelling and diffusion-based simulation.

The company’s core asset is Medal’s dataset: roughly two billion video clips per year generated by more than 10 million monthly active users across thousands of games. Unlike YouTube or Twitch footage, which shows gameplay from a spectator’s perspective, Medal’s clips are first-person and interactive, capturing the spatial reasoning, timing, and decision-making that games demand.

Agents, not models

General Intuition’s pitch differs from most of its competitors in one important respect. Companies like Decart, which raised $300 million earlier this month, and Google’s Project Genie, which recently connected its world model to Street View’s 280 billion images, are building world models as products in themselves, selling simulation environments to developers and enterprises.

General Intuition builds world models to train agents, making the agents the product and the world model the training ground. The distinction matters commercially because it ties the company’s revenue to what its AI can do, not to how well it can render a virtual scene.

A crowded and well-funded field

The world model race has attracted enormous capital in 2026. Runway, valued at $5.3 billion, launched its first world model in December and has been adding $40 million in annual recurring revenue per quarter.

World Labs, founded by Stanford’s Fei-Fei Li, raised $1 billion at a $5.4 billion valuation in February. Odyssey AI raised $310 million with backing from AWS and AMD, and Yann LeCun’s AMI Labs raised over $1 billion at a $3.5 billion valuation.

The common thesis is that AI needs to understand physics, not just language, and that video data is the bridge. General Intuition’s founders argue that their data gives them an edge most competitors lack.

Alonso and Micheli developed DIAMOND, a diffusion-based world model that predicts future frames directly rather than compressing them into tokens. The Medal dataset provides the kind of dense, interactive footage that static video cannot match.

What comes next

The company reportedly plans to use the new capital to scale compute and release a product by late summer or early autumn. The broader question is whether the agent market OpenAI and others are chasing will materialise quickly enough to justify the valuations being attached to companies building the underlying reasoning.

OpenAI wanted Medal’s data badly enough to offer half a billion dollars, and the fact that General Intuition said no and is now worth four times that amount says something about how the market prices control of training data in an era when every major AI lab is hunting for it.