General Intuition raises $220M at a $6.2B valuation to turn gameplay footage into robot training data

General Intuition announced on Monday that it has raised $220 million at a $6.2 billion valuation, tripling the company's worth in roughly three months. The round was led by Valor Equity Partners, with Atreides, 776, Point72, Khosla Ventures, and General Catalyst participating.

The New York startup is betting on a simple, unproven thesis: the missing dataset for physical AI isn't in the real world at all. It's in video games. General Intuition trains world models on billions of action-labeled gameplay videos, teaching AI to understand how actions change an environment over time — the kind of understanding a robot needs before it can be trusted in a factory.

A data moat built on gameplay#

General Intuition spun out of Medal in 2025. Medal is a platform where players capture and share gameplay clips, and that lineage is the core of the investment case: the startup claims leading robotics and world models are currently trained on less than 1% of the amount of action data it can access through Medal.

Medal is on track for roughly three billion uploaded videos per year, according to the company — an ever-growing stream of human behavior recorded inside interactive, simulated environments. General Intuition argues that games are unusually good training fodder: players constantly encounter unpredictable situations and respond in real time, which is exactly the skill a robot needs.

CEO Pim de Witte, Medal's co-founder, made the case vividly in a LinkedIn post announcing the round. “On any given day, we see more accidents in sim than across the entire USA,” he wrote — and at any moment, more people are playing driving games with steering wheels than Waymo has cars on the road.

MIRA: long, endless synthetic video#

The technology at the center of the pitch is MIRA, the company's world model, released in June. Most video-generation systems produce short clips before the quality collapses; MIRA claims to generate indefinitely without diverging, rendering at 20 frames per second in 720×576 resolution on a single Nvidia B200. It does this with 5.6 billion parameters — a modest size next to the trillion-parameter frontier models that dominate headlines — working in a compressed latent space rather than raw video frames.

That's the honest caveat, and it's a significant one: the current MIRA remains a research demonstration centered on a single game, and its broader commercial capabilities are not yet established. The company says it is testing the technology with a limited group of customers in robotics, simulation, and entertainment. The $6.2 billion valuation is, in effect, a bet that simulated accidents transfer to physical competence.

A funding pace that tripled the valuation#

The financing history reads like a startup being priced on momentum. General Intuition raised a $133.7 million seed round in October 2025, led by Khosla Ventures and General Catalyst. In June, it closed $320 million at a $2.3 billion valuation. The new $220 million pushes disclosed funding past $650 million — and the valuation up roughly 2.7 times in about a quarter.

The company plans to expand its team in New York and Europe as it scales model training, compute capacity, and commercialization. That hiring note matters: the pitch is no longer just a research project, but a product push.

Why physical AI is thirsty for data#

The backdrop is a real bottleneck. Robots learn by trying things and failing, but every physical failure costs time, hardware, and occasionally something important. Recording enough real-world robot experience is slow and expensive. Gameplay data is cheap, abundant, and — crucially — already labeled with the player's actions: turn, brake, jump, dodge.

If that action knowledge transfers from simulated worlds to physical ones, synthetic video becomes the default training feedstock for every robot program built on top of it. That's the prize Valor, Khosla, and the rest are paying up for. World models have attracted heavy capital all year from labs hunting for training data that doesn't come from humans typing — this round is the latest and one of the largest expressions of it.

What to watch#

Three things decide whether this bet pays. First, whether gameplay-trained models actually transfer to physical robots — the claim at the center of the thesis has not been demonstrated at scale. Second, the first partner deployments in robotics and simulation, where “testing with a limited group of customers” has to become repeatable revenue. Third, the competitive lane: the world-model thesis is crowded, and the startup that owns the gameplay-to-simulation pipeline could end up owning a piece of every robot program built on top of it — or discover that billions of game videos are worth far less outside simulated environments.