Summary: General Intuition announced on June 25, 2026 a $320 million round at a $2.3 billion valuation, bringing disclosed funding to $454 million. The Medal spinout trains world models on hundreds of millions of hours of gameplay carrying action labels, and fine-tuned a quadruped robot with only eight minutes of real-world data. Funds will largely buy CoreWeave compute, while the company rules out lethal autonomy.
General Intuition, a startup based in New York, announced on June 25, 2026 that it had closed a $320 million round at a $2.3 billion valuation. Khosla Ventures led the investment, with General Catalyst, Jeff Bezos, Eric Schmidt, F1 world champion Nico Rosberg, and researchers from Google DeepMind and MIT all taking part. Count the $134 million seed from October 2025 and disclosed funding has reached $454 million.
The company is built on a single wager: to teach AI about the world, games work better as material than the internet’s text. The startup grew out of Medal, a game-clip sharing platform created by CEO Pim de Witte, and its core asset is hundreds of millions of hours of Medal gameplay — every clip tagged with embedded action labels that record exactly which buttons the player pressed, and when. As de Witte puts it plainly: “It works on anything that you can control using a game controller or a keyboard mouse.”
The Deal At A Glance
Word of this round was circulating by mid-June: TechCrunch reported at the time that the company was negotiating to raise roughly $300 million at around a $2 billion valuation, and the confirmed numbers came in somewhat above those reports. Vinod Khosla compressed the thesis into a single line: “In world models, I think the quantum leap is the emergence of intuition in the AI.” One intriguing footnote: before the spinout, Medal rejected an acquisition approach from a major AI lab, and suitors have kept coming since. On that acquisition interest, Khosla sounded confident: “At this point, it would be a data acquisition, which is sort of uninteresting.”
Why Gameplay Footage Gives The Model An Edge
A typical world model watches video and has to guess what lies beyond the frame. General Intuition’s data arrives with the answers included: shown a car taking a corner, the model also knows which inputs the driver made at that moment. One model, then, handles three jobs at once. It plays games — an in-house demo agent kept playing for 100 hours straight, per CPO Kent Rollins. It serves as a generated world model used internally as a training environment, nicknamed “the gym.” And it controls physical robots. Eloi Alonso, Adam Jelley, and Vincent Micheli complete the founding team beside de Witte, and the whole bet rests on action labels — a data moat nobody else holds.
Out Of The Simulated Gym, Onto Real Hardware
The headline result is a quadruped robot fine-tuned with only eight minutes of real-world data. The reasoning: a robot first learns to walk, run, and carry objects in simulation, then touches up those skills in the physical world — dramatically cheaper than collecting real-world data starting from zero. De Witte is explicit that the company has no plan to enter the vehicle business: “We’re not gonna build a self-driving car company.” What it will do is clear the road for whoever comes next: “We’re gonna make it 10 times easier for the next person to build” one.
Where The Money Goes, And The Nerve Marketplace
The largest share of the new capital buys compute, via a deal with CoreWeave, to pre-train the next version of the model. The API is also on track to open broadly before summer ends, with current customers already spread across gaming, simulation, and robotics. A second product deserves attention: Nerve, a jobs marketplace where gamers get paid for labeling data and teleoperating robots. That effectively crowdsources the data supply chain to the player community and hands ordinary people a slice of the robotics economy. Nerve also cleans up a problem of the company’s own making: as the robot fleet expands, so does the demand for human demonstration data and teleoperation hours, and the marketplace turns that cost center into a gig network that pays.
The Competition, And The Lines The Company Won’t Cross
Capital is flooding into world models. Fei-Fei Li’s World Labs has already put out its first commercial product, Decart’s newest model can render hours of photorealistic driving, and Odyssey closed a round in mid-June at a $1.45 billion valuation. Most of these rivals are selling simulation or generation. General Intuition stands apart on its action-labeled dataset and its path into robotics: the same weights that predict the next frame also choose the next action. The company has set boundaries, too — de Witte has said, “We don’t want to be an escalatory part of the system,” ruling out lethal autonomy while accepting uses such as search-and-rescue. For a startup holding defense-adjacent technology and raising money from billionaires, stating that position in public is itself a way of defining itself.