By Stuart Kerr, Technology Correspondent, LiveAIWire

A two-year-old startup with no robots of its own has just raised 320 million dollars at a 2.3 billion dollar valuation, backed by Jeff Bezos, Eric Schmidt, and Khosla Ventures, on a bet that the same data powering Fortnite and Call of Duty can teach machines to move through the physical world. The company, General Intuition, spun out of the gaming clip platform Medal in 2025 and has already pushed its total disclosed funding to 454 million dollars in under a year.

Its pitch is that robotics is stuck exactly where language AI was before GPT-3, a field of one-off, hand-built systems, and that a single foundation model trained differently could unlock it the same way large language models unlocked text.

The company’s chief executive, Pim de Witte, says the industry has been solving the wrong problem. Most robotics labs collect enormous real-world datasets, recording sensors and cameras for months to teach a single robot a single task in a single environment. General Intuition instead trained its model on hundreds of millions of hours of gameplay footage, footage that already contains something most video used to train AI lacks: a precise record of exactly which button a human pressed, and when, to produce each on-screen action.

Why Button Presses Might Matter More Than Video

That distinction, action data rather than passive video, is the core of General Intuition’s argument. Competing approaches try to infer what a person intended to do by watching video alone. De Witte contends that guessing at intent from pixels is not enough to build a model that understands causality, or the difference between an action the model itself is taking and something simply happening in its environment. A controller input removes the guesswork. It shows the model exactly what changed and why, millions of times over, across tens of thousands of different games and virtual physics.

The company says this produces a kind of transferable intuition about space, time, and consequence that does not need to be relearned from scratch for every new robot body or every new room. Vinod Khosla, whose firm led the new funding round, has backed that thesis publicly, arguing that action data of this kind is what physical AI has been missing.

What This Means for You

If General Intuition’s bet pays off, the practical effect is that the robots reaching warehouses, hospitals, and eventually homes could get competent far faster and far cheaper than the current model of training allows, because the expensive part, teaching a machine to understand physical cause and effect, would already be solved before a manufacturer ever builds a specific robot. That would matter to anyone who has watched humanoid robot demonstrations for the last several years and wondered why so few of them ever leave the lab. It would also matter to your job, in delivery, warehousing, cleaning, and eldercare, on a faster timeline than most forecasts currently assume.

It is also a reason for caution rather than certainty. A model that plays a video game convincingly is being asked to do something categorically different from safely sharing a kitchen with a toddler, and the gap between an impressive demo and a dependable product has swallowed plenty of robotics startups before this one.

The Eight-Minute Robot That Surprised Its Own Creators

General Intuition has already shown its model doing both jobs, the same underlying network that plays a video game for hours has also been used to control a four-legged robot, fine-tuned on just eight minutes of real-world robotics data. According to de Witte, speaking on TechCrunch’s Equity podcast, the robot was able to navigate an office using only its front camera, with no other sensors, while people walked past and objects were moved around it. “The reality is, you only need a few minutes,” he said of the real-world data required once the base model already understands movement.

That claim, if it holds up outside a single office demo, is the entire investment thesis in one sentence. Today, teaching a warehouse robot a new task can mean weeks of real-world data collection for every new task, every new layout, and every new piece of hardware. A foundation model that only needs minutes of fine-tuning would collapse that cost structure, which is exactly why investors who do not normally chase robotics wrote checks this large this fast.

Robotics Already Has Its Foundation Model Race

General Intuition is not alone in trying to become the operating layer underneath every robot rather than a robot maker itself. Physical Intelligence and Skild AI, both backed by major venture money, are pursuing versions of the same strategy, building general-purpose models that other companies license and adapt to their own hardware. NVIDIA has taken a parallel path with its Isaac GR00T models, paired with its own simulation tools, aimed specifically at humanoid robots. What separates General Intuition’s pitch is its data source: rather than expensive real-world capture or synthetic simulation, it uses recorded gameplay with embedded controls, a resource that already exists in the billions of hours and costs almost nothing to collect at scale.

Academic researchers studying robot foundation models sound more cautious than the venture money suggests investors are. A 2026 survey of the field flagged that the hardest unsolved problem is not collecting enough data but generalisation itself, whether a model performs reliably across environments and robot bodies it has never seen, rather than only the conditions it was shown. General Intuition’s eight-minute office demo is one controlled example, not proof the approach holds up on a factory floor or in a family kitchen.

Why the Comparison to ChatGPT Is Doing a Lot of Work

Calling this robotics’ ChatGPT moment is a useful shorthand and also a claim that has not yet been earned. Large language models had years of internet text, a relatively forgiving failure mode, and billions of users willing to tolerate mistakes while the technology improved in public. A household robot that misjudges a staircase or a kitchen knife does not get the same grace period. LiveAIWire’s own reporting on the pace of home robots reaching consumers has found that the gap between a viral demo and a dependable product has consistently been the part every robotics company underestimates, no matter how capable the underlying model looks on stage.

The size of General Intuition’s raise also lands inside a broader pattern of AI valuations moving well ahead of proven revenue, the same dynamic LiveAIWire examined in OpenAI’s own valuation, where investor enthusiasm for platform potential is pricing in a future that has not arrived yet. General Intuition has not sold a single commercial robot deployment. It has sold a thesis, backed by a demo, to some of the most experienced technology investors alive. Whether that thesis survives contact with a real customer’s warehouse, rather than an office with a single camera, is the question the next eighteen months should answer.

Meta’s own embodied AI ambitions, detailed in LiveAIWire’s look inside Meta’s AI labs, suggest the race to own the foundation layer of physical AI is only getting more crowded from here, not less.

About the Author

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.