General Intuition started as a side effect of a refusal. OpenAI reportedly offered $500 million for Medal, a platform where gamers upload and share short clips of their gameplay. Medal's founder, Pim de Witte, turned it down. Instead, he spun out a separate AI research lab in October 2025 using Medal's data as the foundation. That lab, General Intuition, just closed a $320 million round at a $2.3 billion valuation, with Jeff Bezos, Eric Schmidt, and Khosla Ventures among the backers.
The central idea is simple to state and hard to execute. Current AI systems that read and write text are good at language. They are poor at understanding physical space: how objects move, where walls are, what happens when you walk around a corner. Teaching that to an AI requires different data. General Intuition's argument is that video games already contain it.
Medal's platform generates roughly 2 billion video clips per year from 10 million monthly active users playing across tens of thousands of games. The crucial detail is not the footage itself. Every clip contains a record of exactly which buttons the player pressed and when. The AI learns that pressing a button in a given situation produces a specific outcome. Repeat that across billions of examples and the model starts to understand physical cause and effect, not just visual patterns.
This matters because the gap between a capable AI in simulation and a capable AI in the real world is where most robotics projects quietly die. Collecting real-world training data physically is slow and expensive. The average cost of one hour of high-quality teleoperation data, captured and packaged for AI training, was still around $136 per hour as recently as late 2025. General Intuition's claim is that gaming footage is a scalable shortcut past that cost.
Demos support the idea in limited form. The company showed a four-legged robot navigating an unfamiliar office after just eight minutes of real-world data collected outdoors. The same underlying model played a game continuously for 100 hours. Whether that transfers to the reliability and consistency that industrial or commercial deployment requires is a genuinely open question. Nobody has fully solved it at scale.
The broader context makes this bet worth watching. Goldman Sachs projects the humanoid robot market will reach $38 billion by 2035. Manufacturers across logistics, construction, and heavy industry are facing labor shortages that robots are being asked to fill. The bottleneck is not hardware anymore. It is training data and the software intelligence to use it. General Intuition is positioning itself as a provider of both, not as a robot builder.
The company's stated plan is to sell access to its AI model through an API, letting other companies build products on top of it: smarter game characters, warehouse robots, search-and-rescue drones, testing environments that mirror factory floors. It explicitly does not want to build end products itself.
There is also a workforce angle worth noting. General Intuition launched a platform called Nerve, which lets gamers earn money by doing data labeling tasks using their existing gaming setups, with a path toward eventually operating robots remotely. The user base it is drawing from, younger gamers, is the group most directly exposed to automation-driven job displacement. The company is trying to turn that group into a paid contributor to the AI supply chain rather than just a casualty of it.
The core risk is straightforward. A 2.3 billion dollar valuation for a company with a handful of customers rests on the assumption that game-trained AI will actually hold up in unpredictable physical environments. Games, even complex ones, follow rules. The real world does not. That gap has humbled many well-funded robotics projects before. General Intuition's data advantage is real. Whether it is sufficient is what the next few years will answer.