General Intuition Eyes $6B Valuation to Train Physical AI Models on Gameplay Data

Robotics and foundation model startup General Intuition is finalizing a funding round valuing the company at $6 billion pre-money, driven by investors including Valor Equity Partners, Point72 Ventures, and Seven Seven Six alongside returning backers Khosla Ventures and General Catalyst. The New York-based startup leverages hundreds of millions of hours of action-labeled gameplay recordings gathered from CEO Pim de Witte's clip-sharing platform, Medal, to train large action foundation models. By using rich synthetic interactions from video games, the company aims to teach generalized embodied AI agents how to navigate space and interact with environments prior to real-world deployment. The development addresses the most persistent hurdle in modern robotics engineering: the physical data bottleneck. While large language models scale relatively easily on scraped internet text and multimodal media, embodied AI has historically required specialized teleoperation rigs, custom simulation setups, or expensive robot test fleets operating inside physical facilities. If video game interactions can serve as a viable proxy for physical dynamics and spatial reasoning, engineering teams can drastically lower the cost per training hour and shrink the fine-tuning phase—with General Intuition claiming its gameplay-pretrained models require mere minutes of real-world footage to navigate physical spaces. This shift fits into the wider race toward general-purpose physical foundation models, where companies like Physical Intelligence, Skild AI, and NVIDIA are competing to build unified brain layers for diverse robotic form factors. Rather than designing specialized control loops for individual manipulation or locomotion tasks, the industry is converging on world models and large action models (LAMs) that transfer broad environmental representations across varied robotic hardware. Pre-training on massive simulated and virtual action datasets represents a strategic divergence from hardware-heavy data collection methodologies. In practice, robotics engineers and platform architects should evaluate whether virtual action representations transfer reliably to noisy physical sensors and uncalibrated actuators. While gameplay footage provides immense diversity in spatial reasoning and path planning, it cannot fully account for real-world physical anomalies such as frictional variance, sensor occlusion, or mechanical wear. Engineering teams adopting similar foundation models will likely need to construct robust Sim2Real validation pipelines, combining game-trained priors with rigorous hardware-in-the-loop testing and edge safety monitors.