General Intuition's Valuation Nearly Triples to $6B as World Models Become AI's Hottest Bet

Just eight weeks separated General Intuition’s $2.3 billion valuation from the $6 billion pre-money valuation it is now finalizing — one of the fastest re-ratings the AI industry has seen this year. According to TechCrunch, the New York-based world-model startup is in talks to raise fresh capital from new investors Valor Equity Partners, Point72 Ventures, and Alexis Ohanian’s Seven Seven Six, with existing backers Khosla Ventures and General Catalyst re-upping. A source close to the deal says the round is oversubscribed as the company continues to field inbound interest.

If the numbers hold, that is a near-tripling of the company’s paper value since late June, when it raised $320 million at $2.3 billion. For a company that did not exist two years ago, the trajectory is vertiginous — and it says less about any single startup than about where capital believes the next leg of AI value creation sits: in models that understand space, time, and physical action, not just text.

From gaming clips to large action models

General Intuition was spun out in October 2025 by CEO Pim de Witte from Medal, the video game clip-sharing platform he founded. The unusual asset at the heart of the spin-out was Medal’s archive: hundreds of millions of hours of gameplay video accompanied by “action labels” — timestamped records of exactly which buttons a player pressed and when. Where most training data teaches a model what the world looks like, Medal’s data teaches a model what players actually did in it, frame by frame.

That distinction is the entire thesis. A large language model predicts the next token; a large action model predicts the next action, conditioned on a visual understanding of an evolving three-dimensional environment. General Intuition trains its foundation model on that combined stream of pixels and inputs so that agents can learn to move through space and time — the skill, the company argues, that underlies generalization across tasks the model was never explicitly trained on.

Investor Vinod Khosla told TechCrunch earlier this year that he believes those action labels will be a key part of the “emergence of intuition” — the ability of a model to genuinely generalize rather than pattern-match. It is a claim that would have sounded mystical in 2023 and sounds almost mundane in 2026, as engineers and investors pile into world models, or “large action models,” hoping to do for robotics what transformers did for writing and coding.

Why robotics is the destination

The new capital has a specific destination. Sources say General Intuition intends to spend it pushing its general model into robotic embodiments — meaning substantially more compute (the company has a partnership with CoreWeave) and aggressive hiring.

The strategic logic is straightforward. Video-game environments are physics engines with perfect telemetry: cheap, resettable, and endlessly generative. But the money in physical AI is not in playing games; it is in warehouses, factories, kitchens, and elder care — environments that are messy, expensive to instrument, and impossible to reset. The bet is that a model pretrained on hundreds of millions of hours of labeled human action can transfer to real actuators with far less real-world robot data than competitors who train from scratch on teleoperation recordings.

That bet now has company. The Wall Street Journal’s August feature on world models describes a field crowding with entrants, and the funding environment has begun pricing the scarcity: Valor Equity Partners, best known for backing SpaceX, would be making its first AI-lab investment since SpaceX. Point72 Ventures and Seven Seven Six joining the cap table extends a pattern in which finance-native and founder-native funds are both chasing the same small pool of world-model teams.

The re-rating race

An eight-week, 2.6x step-up invites the obvious question: what changed? Partly momentum — the June round itself validated the category, and WSJ reporting on the company’s “large action models” gave the thesis mainstream visibility in August. Partly scarcity: credible world-model teams with proprietary action data are extremely rare, and General Intuition’s Medal lineage gives it a data moat that cannot be replicated by scraping the web a second time.

But the deeper driver is that physical AI timelines have compressed. Nvidia says its physical-AI business already generates roughly $10 billion in annual run-rate revenue, with Jensen Huang projecting $100 billion within a decade. Chinese robotics firms have become one of the largest customer bases for that stack. When the demand side of embodied AI scales that fast, the model layer feeding it gets repriced accordingly — and the market appears willing to pay venture-stage prices for pre-product companies on the expectation that robotics demand will absorb whatever they produce.

Caveats worth keeping

Skeptics will note what the round is not. It is not a close — TechCrunch describes the deal as still being finalized, and terms can move. It is not revenue-driven; the company remains pre-scale on the commercial side. And a near-tripling in eight weeks is as much a statement about FOMO dynamics in AI venture as about technical progress: oversubscribed rounds at accelerating valuations were the signature of the 2023–2024 foundation-model frenzy, and world models are inheriting that mantle with remarkably little lag.

There is also a competitive risk that the thesis itself gets commoditized. OpenAI made headlines in October 2025 with a reported $500 million data grab aimed at gaming footage, signaling that the giants see the same signal in controller inputs that de Witte saw. If frontier labs can buy or synthesize comparable action-labeled data at scale, General Intuition’s moat narrows to execution speed and team quality.

None of that has cooled the market. The round, once closed, would make General Intuition one of the most valuable world-model startups on paper — and the clearest signal yet that investors have decided the next ChatGPT-scale platform shift will be trained not on the internet’s text, but on humanity’s button presses.