General Intuition Raises $320M to Train AI on Gameplay Data
General Intuition, a New York AI lab that trains models on video game footage, said Thursday it raised $320 million in Series A funding at a $2.3 billion post-money valuation, one of the largest early rounds yet for a startup betting that gameplay is the fastest path to artificial intelligence that can act in the physical world.
Khosla Ventures led the round, with participation from General Catalyst, Bezos Expeditions, Eric Schmidt, Innovation Endeavors, Hedosophia, former Formula One champion Nico Rosberg, and individual researchers at Google DeepMind and MIT. The financing brings the company's total disclosed funding to $454 million, following a $134 million seed it raised when it launched last October. According to Axios, General Intuition is already in talks for a Series B.
The company was spun out of Medal, a platform built by founder and chief executive Pim de Witte that lets gamers upload and share clips. Those hundreds of millions of hours of recorded play gave General Intuition its starting dataset. But de Witte argues the footage itself is not the breakthrough. The clips carry embedded action labels, exact records of which buttons a player pressed and when, and that pairing of pixels with inputs is what most rivals lack. Competitors, he says, are trying to infer actions from video alone, which he considers insufficient for teaching a model cause and effect.
From game controller to quadruped robot
General Intuition's goal is a single model that can move fluidly from playing a game, to navigating a simulated world, to controlling a real machine. In a demonstration for TechCrunch, the company showed an AI agent that had been playing a Fortnite-style game for 100 hours straight, then revealed that the same underlying model was driving a large four-legged robot exploring the office. The team said it took only eight minutes of real-world data, gathered on the street rather than in the office, to fine-tune the model for the quadruped.
The training environment, which the company calls "the gym" internally, is a world model that generates a playable scene frame by frame instead of rendering it with a traditional game engine. From millions of hours of gameplay, the system appears to have learned basic physical logic, that walls block movement, that ladders are for climbing, and that shadows shift as light moves. General Intuition does not plan to sell the world model. It plans to sell the agent that learns inside it.
Vinod Khosla, whose firm led the round, framed the bet in terms of a coming leap in capability. "The quantum leap is the emergence of intuition in the AI," he said, comparing it to the moment reasoning emerged in large language models. He added that the company's proprietary access to human action data through Medal is what convinced him this was a generational company rather than an acquisition target.
What it means for founders and operators
The round lands in the middle of a broader shift in where AI money is flowing. A Goldman Sachs report shared with Axios argues that the next phase of AI investment is moving out of pure software and into the physical economy, the world of robots, machines, and embodied agents. General Intuition's pitch fits that thesis directly. If gameplay really can substitute for the slow, expensive real-world data collection that robotics has traditionally required, it would lower the cost of building physical AI products and widen the field of who can compete.
De Witte is explicit that he wants to be a platform, not a product company. "We're gonna make it 10 times easier for the next person to build a self-driving car company," he said, comparing his ambitions to the model-provider role that companies like Anthropic and OpenAI play for software developers. For operators, that framing matters. A foundation model for action and world simulation could become infrastructure that startups in gaming, logistics, manufacturing, and robotics build on top of, much as a generation of software companies built on top of cloud and language models.
There is a clear caution worth keeping in view. Impressive demos are not the same as production reliability, and nobody has yet proven that a model trained largely on games can hold up in the messy physical world at scale. That simulation-to-reality transfer remains an open research question, and General Intuition's own investors acknowledge it. Founders watching this space should treat the technology as promising rather than proven, and weigh how much of a roadmap to stake on capabilities that are still being demonstrated.
The capital itself is going mostly toward compute. General Intuition has a deal with CoreWeave and plans to use the bulk of the round to pre-train the next version of its model, with a slice set aside to open its API more broadly by the end of summer. Access, in other words, is the gate. Until that API is widely available, outside builders cannot test whether the platform vision holds.
An unusual stance on values
The company is also drawing lines that stand out in the current climate. De Witte, who is Dutch and spent three years in humanitarian work including with Doctors Without Borders, says no agents will be built to harm people, ruling out lethal autonomy while leaving room for uses like search and rescue. He has staffed the company partly around that posture, hiring a chief of staff who publicly left Palantir over its immigration enforcement work. General Intuition has also launched Nerve, a marketplace that pays gamers for tasks ranging from data labeling to robot teleoperation, an attempt to give the generation most exposed to AI displacement a stake in the systems being built.
What did General Intuition raise and at what valuation?
The company raised $320 million in a Series A round at a $2.3 billion post-money valuation. The financing was led by Khosla Ventures and brings its total disclosed funding to $454 million. Reports indicate it is already in discussions for a Series B.
How does training AI on video games actually work?
General Intuition uses gameplay clips that include action labels, meaning the data records both what happened on screen and which buttons the player pressed at each moment. That pairing lets the model learn cause and effect, which the company argues is harder to achieve from video footage alone.
Can a model trained on games really control a robot?
In demonstrations, the same model that played a shooter-style game also drove a four-legged robot after only eight minutes of real-world fine-tuning. Whether that transfer holds reliably at scale, across many environments and machines, is still unproven and remains an open research question.
Why are investors like Bezos and Khosla interested?
Backers see proprietary action data, gathered through the founder's gaming platform Medal, as a defensible advantage that is hard for competitors to replicate. They are betting that this data can produce a foundation model for physical action and world simulation that other companies build on top of.
What does this mean for the broader AI market?
It reflects a shift of AI investment toward the physical economy, including robotics and embodied agents, rather than software alone. If gameplay can replace slow, costly real-world data collection, it could lower the barrier for startups building physical AI products.