General Intuition is in talks to raise around $300 million at a valuation of just over $2 billion, with Jeff Bezos and Eric Schmidt joining the round alongside existing backers Khosla Ventures and General Catalyst. The New York-based startup is building foundation models that teach AI agents how to navigate space and time, and the new round arrives eight months after it spun out of Medal with a $134 million seed. That pace puts General Intuition's paper valuation at roughly 15x its seed mark inside a single fiscal year.
The pitch hinges on data. General Intuition trains world models on Medal's archive of 2 billion gameplay videos per year, generated by 10 million monthly active users uploading first-person clips. The company argues that interactive gameplay footage — where a human is actively steering through a 3D environment — is a stronger substrate for spatial-temporal reasoning than the passive video datasets most rivals scrape from the open web.
Pim de Witte, who co-founded Medal, runs the new company. His co-founders — Eloi Alonso, Adam Jelley, and Vincent Micheli — come from world modeling and simulation research. The team's stated goal is not to sell world models as a product but to use them as a training ground for embodied agents, which would then become the commercial output.
Key facts
- 01General Intuition is in talks to raise around $300M at a valuation just over $2B, roughly 15x its seed round eight months earlier.
- 02The seed round closed at $134M when the startup spun out of Medal, the gameplay clip-sharing platform.
- 03New backers include Jeff Bezos and Eric Schmidt, alongside existing investors Khosla Ventures and General Catalyst.
- 04The company trains world models on Medal's dataset of 2 billion gameplay videos per year from 10 million monthly active users.
- 05OpenAI previously tried to acquire Medal, and other frontier labs have approached General Intuition since the spinout.
That positioning separates General Intuition from a crowded field. OpenAI previously tried to acquire Medal outright for the same dataset, and other frontier labs have approached the spinout since launch. Runway, Decart, and World Labs have each shipped world models in recent months, and Google's Genie 3 has begun pulling Google Maps data to ground simulations in real geography. Most of those efforts target gaming and robotics training as the revenue path.
General Intuition is taking the agent-first route instead. Rather than license a simulator to a robotics company, it intends to ship trained agents that can act inside arbitrary 3D environments — game worlds first, physical-world robotics later. The Medal dataset, which competitors cannot replicate without a comparable consumer product feeding it, is the moat.
The fresh capital will go primarily to compute. Sources told TechCrunch the company plans to release a new product by the end of summer or early fall, which implies a large training run is already underway. World model training is GPU-intensive even by frontier-AI standards, since the models have to learn physics, occlusion, and temporal continuity simultaneously.
The investor list is notable. Bezos and Schmidt have both been steady late-stage backers of frontier AI startups, with Bezos in particular writing checks into Anthropic and Physical Intelligence. Their presence on a Series A-stage round at a $2 billion mark suggests they are pricing the dataset advantage aggressively rather than waiting for revenue traction.
There are reasons to be cautious. General Intuition has not disclosed a product, customer roster, or revenue figure, and the world model category as a whole has yet to produce a commercial breakout. Genie 3, Runway's offerings, and Decart's real-time generation tools remain demos or developer previews rather than line-item revenue contributors. A $2 billion valuation on a pre-product startup assumes the agent thesis pays off before the next funding window closes.
If the bet works, the consequence for the broader AI market is that the gating resource for embodied agents shifts from model architecture to interactive training data — and the companies sitting on consumer-scale gameplay footage become the new bottleneck. That is bad news for labs trying to train world models on YouTube and good news for anyone who happens to own a clip platform with 10 million monthly users. OpenAI's earlier interest in Medal now looks less like opportunism and more like a strategic read that the rest of the field is only beginning to price in.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
