General Intuition has raised $320 million on a blunt bet: the button presses behind gaming clips may teach robots more about movement than cleaner, smaller robotics datasets ever could.
Pim de Witte had a $500 million reason to sell Medal to OpenAI in late 2024. He didn't take it. He looked at the clip-sharing platform he had built for gamers, saw the data sitting underneath it, and decided the footage was worth more as the foundation for a new AI company than as a line item inside somebody else's lab.
This week, TechCrunch reported that de Witte's General Intuition has raised $320 million in a round led by Khosla Ventures. General Catalyst, Jeff Bezos, Eric Schmidt, Nico Rosberg, and researchers from Google DeepMind and MIT also participated. The round values the company at $2.3 billion and brings disclosed funding to $454 million, after a $134 million launch raise in October 2025.
You can see why investors are paying attention. Medal has roughly 10 million monthly active users, and those users upload about 2 billion video clips a year. A gaming clip isn't just a video. It can carry the controller input, the button press, the mouse movement, the action that produced what you see on screen. That is the part robotics labs usually struggle to get at scale: not only what happened, but what a person did to make it happen.
The company's argument is sharp enough to deserve a hearing. If a model can watch a Fortnite player move through a cluttered fight and predict the next frame from the player's inputs, it may be learning something useful about space, timing, intention and consequence. A warehouse robot does not need to win a match. It needs to move through an awkward physical environment without smashing into a shelf, a cart, or a person.
Games are not the real world. That objection is fair, and frankly it is the first one you should make. A game engine has rules that are clean in ways a factory floor never is. Collision detection is precise. Objects behave according to code. A robot arm near a human worker has to deal with weight, friction, bad lighting, latency, and the kind of mess no designer placed there on purpose.
But scale changes the question. General Intuition is not claiming that a video game is a factory. It is betting that hundreds of millions of human decisions, attached to visible outcomes, can help a model build the spatial-temporal instincts robotics still lacks. That is why the CoreWeave deal matters. TechCrunch reported that most of the new money is going toward compute, with the next version of the model set to be trained on CoreWeave infrastructure and an API planned by the end of summer 2026.
The best counterargument is that this is synthetic data with better labeling. You're still training on worlds built by game studios, not the world a robot has to survive in. The answer is not to wave that away. The answer is to look at the labels. Medal's clips contain real human choices made under pressure, not trajectories generated by a simulator looking for clean paths. If you are training an agent to anticipate what a person might do next, that difference matters.
The investor list also tells you this is not a random AI fashion trade. Jeff Bezos has backed Physical Intelligence, the robotics software company that raised $400 million in 2024 at a valuation above $2 billion. Eric Schmidt has talked openly about robotics as one of the next giant technology markets. Google DeepMind has spent years pushing world models and agents that can reason about space. When people with that background write checks into a gaming-data robotics startup, you don't have to agree with them, but you should ask what they think the bottleneck is.
The wider market gives de Witte room to make the bet. Business Insider recently noted that venture funding for robotics climbed from $4 billion in 2019 to $26 billion in 2025. Figure AI, Physical Intelligence, Skild AI, Agility Robotics, you name it, the money is moving toward machines that can act in the physical world rather than merely answer questions in a text box. Jensen Huang has been calling embodied AI the next wave because Nvidia sells the chips that make those bets possible, but the phrase has stuck because the problem is real.
General Intuition still has to prove the hard part. A large dataset can get a company funded. It cannot by itself make a robot useful. The API expected this summer will be the first public test of whether the Medal thesis works outside a deck and a demo. If it does, de Witte's decision to walk away from OpenAI's reported $500 million offer will look obvious in hindsight. If it doesn't, it will be a very expensive reminder that the world is harder to model than a game.
That is the story here. Not that gaming data is magic, and not that robotics has found its final training source. The useful point is narrower and more interesting: one of the most valuable datasets for physical AI may have been hiding inside a consumer gaming product while the big labs were looking somewhere cleaner.
Also read: OpenAI is leaning toward a 2027 IPO as its CFO warns the company isn't ready for public markets • Patronus AI raises $50 million to build simulation environments that stress-test AI agents before they touch real systems • Europe bets its AI sovereignty on a Milan startup most people have never heard of
