General Intuition raises $320M to train robots on video game data

General Intuition US Inc. closed a $320M Series A round on June 26, 2026, reaching a $2.3B valuation. The New York-based company trains AI models for physical agents and robots using billions of gameplay clips with embedded action labels — an approach designed to eliminate costly real-world or simulation-based data collection.

Key takeaways

Series A round: $320M, valuation $2.3B, total funding $454M
Investors: General Catalyst (lead), Jeff Bezos, former Google CEO Eric Schmidt
Data source: Medal platform — billions of gameplay clips with embedded action labels (every button press and timestamp recorded)
Two model types: action models (decide what to do) + world models (predict outcomes of actions)
Public API planned for summer 2026; CEO Pim de Witte is also co-founder of the Medal platform

Video games as training data for robots

The standard path to training physical AI is either collecting teleoperation data (expensive, slow, task-specific) or simulation (cheap, but with a persistent sim-to-real gap). General Intuition proposes a third path: billions of hours of video game footage. Players constantly perceive an environment, make decisions, and act. That loop — perception → decision → consequence — is exactly the cycle that physical intelligence requires.

The critical detail is that Medal clips are not simple screen recordings. The platform logs every button press and its exact timestamp — providing embedded action labels. Instead of inferring actions from pixels (as in typical imitation learning), the model receives pairs: video frame → exact player action. This substantially simplifies the supervision problem.

CEO Pim de Witte — who founded Medal as a gaming clip sharing platform — sits on both sides of this strategy. Through ownership of Medal, General Intuition has unique access to one of the world's largest repositories of labeled action data. According to the company, this amounts to billions of clips spanning dozens of games and hundreds of thousands of environments.

Action models and world models — what the company is actually building

General Intuition describes two pillars of its architecture: action models and world models. An action model decides what the agent should do at any given moment — analogous to a policy in classical reinforcement learning. A world model predicts what will happen after a given action — an internal environment that lets the agent plan without interacting with the physical world. The company has worked on both since 2015, though the scale of the project became public only now.

Comparison with competitors shows the alternative: World Labs (Fei-Fei Li), Wayve GAIA-1, or Google DeepMind World Model — all build world models from robotics data or real vehicle driving. General Intuition's thesis is that video games provide better environmental diversity at a fraction of the cost. An open-world game has more states than a year of teleoperation data collection.

The round and investors

General Catalyst — one of the largest venture capital funds in the US, with a portfolio including Stripe, Airbnb, and Snap — led the round. Jeff Bezos and former Google CEO Eric Schmidt participated individually. The round brings total funding to $454M — General Intuition had previously raised $134M in October 2025. A $2.3B valuation places the company among the highest-valued pre-IPO AI startups in the US.

Why this matters

The biggest problem in robot training is the quantity and quality of labeled data. Collecting teleoperation demonstrations is expensive and slow — one hour of quality data costs hundreds of dollars and requires a skilled operator. Simulations are faster, but the sim-to-real gap remains one of the hardest problems in robotics. General Intuition is trying to bypass both constraints: a billion hours of gameplay represents a scale no robotics lab can generate. If a model trained on game data actually transfers to physical robot behaviors, that is a paradigm shift in data collection. The robotics industry is waiting for concrete results — the company promises public API access in summer 2026, so external validation is near.

What's next?

General Intuition's API is planned for public release in summer 2026 — the first external validation of model quality in robotics applications
The $320M round will fund compute scale-up and pretraining of the next model version — compute scale is critical for world models of this generation
Key question: whether a world model built on game data can predict actions in the physical world — and how large the virtual-to-real gap actually is