The $2.3B Series A, which the company says Khosla Ventures led, bets that action labels embedded in Medal's 17 million users' gaming uploads can train the AI that controls robots in the real world.
General Intuition, based in New York, raised a $320 million Series A at a $2.3 billion valuation this week, betting that action labels buried inside Medal's billions of gaming clips are the missing substrate for physical AI, the kind of artificial intelligence that controls robots and other machines in the real world.
The round brings General Intuition's total raised to $454 million, including a $134 million seed from October 2025, and was first announced on LinkedIn by founder and CEO Pim de Witte. The company stated that Khosla Ventures led the round, with General Catalyst, Hillspire (Eric Schmidt's family office), and Jeff Bezos participating; Robot Report reported General Catalyst as lead, creating a direct conflict between sources on this point with no third-source resolution currently available.
The mechanism behind the $2.3 billion price is Medal. De Witte also cofounded Medal, a gaming-clip-sharing platform whose roughly 17 million monthly users upload short videos of their play. Those uploads arrive already embedded with the button presses and timing that drove each in-game action, since players record their own sessions. General Intuition treats those labels as a proxy for how humans perceive and decide inside an environment, and uses them to pretrain what it calls "large action foundation models" alongside world models that generate fresh training environments on demand. The company's pitch, laid out in de Witte's interview on Latent Space and TechCrunch's coverage of the round, is that this human-action-labeled video corpus is the moat: a dataset competitors cannot assemble without running their own consumer gaming platform.
That framing recasts the category as a data-substrate problem rather than a model-architecture problem. Most competing bets in the same space lean on teleoperated robot demonstrations or large-scale synthetic simulation; General Intuition is leaning on consumer gameplay recorded by humans who were not told they were training a model.
The bet is also the obvious weakness. Game physics are not warehouse physics, residential-floor physics, or sidewalk physics. The action vocabulary available inside a first-person shooter or driving game is a thin slice of what a real robot body must execute. The company has not yet published independent benchmarks showing that models pretrained on Medal clips transfer to deployed robot tasks, and it has not named shipped customers. The $320 million is, in practice, a wager that the action-label layer generalizes, plus a bet that the 17 million monthly users and "billions of clips" figures, both drawn from Medal's own metrics and the company's announcement coverage, keep growing fast enough to stay ahead of competitors who could scrape public gameplay video without Medal's built-in action labels.
Stated uses of the capital are compute, the next pretraining run, and broader API availability, per TechFundingNews and MLQ.ai. Dutch regional press noted the round's significance for the Dutch founder's home ecosystem. Whether the action-label moat holds is now a falsifiable question. The next model release, and any peer-reviewed or third-party benchmark General Intuition permits, will be the first hard evidence that gamer button-presses scale into robot competence, or the first signal that the $2.3 billion was paid for a clever dataset story rather than a working one.