General Intuition Hits $6B Valuation Betting on Game Data for Robots

General Intuition is closing in on a $6 billion pre-money valuation just months after its last funding round valued the startup at $2.3 billion. The jump reflects growing investor confidence in a bet that video game footage, rather than warehouse or lab robotics data, can teach machines to move through the physical world.

The company's core dataset is unusual: hundreds of millions of hours of gameplay footage paired with action labels, essentially records of every button press made during play. That combination lets General Intuition train a foundation model designed to move generalized AI agents through space and time, an approach it argues generalizes better than data collected from narrow robotic tasks.

The new round brings in Valor Equity Partners, known for its early backing of SpaceX, alongside Point72 Ventures and Seven Seven Six. Existing backers Khosla Ventures and General Catalyst are returning to participate, a signal that the company's earlier supporters see enough progress to keep doubling down.

That earlier round, a $320 million raise at a $2.3 billion valuation, already marked a steep bet on the company's thesis. The new capital is earmarked for improving the general model with a specific push toward robotic embodiments, moving the technology closer to physical deployment rather than remaining a research exercise confined to simulated or gaming environments.

From game clips to robot brains

General Intuition's origin traces back to Medal, a video game clip-sharing platform where CEO Pim de Witte built up the trove of gameplay data that now underpins the startup. He spun the AI effort out as its own company in October 2025, betting that the platform's archive of player behavior could be repurposed to train embodied AI systems.

To handle the compute demands of training on that scale of video and action data, General Intuition has a partnership with CoreWeave for infrastructure. That relationship gives the startup access to GPU capacity without building out its own data centers, a common strategy among AI labs racing to scale foundation models quickly.

The wager underlying all of this is that skills learned from navigating virtual game worlds, judging distances, reacting to obstacles, sequencing actions, transfer meaningfully to robots operating in physical space. If that holds up as the company moves from gameplay data toward real robotic embodiments, it would offer a notably cheaper and more scalable alternative to the expensive teleoperation and sensor-rig data collection methods robotics labs currently rely on.