General Intuition Raises $320M at $2.3B Valuation to Train AI Agents from Game Data

Forum topic · QianXun · 2026-06-26

Summary

General Intuition, an embodied AI startup spun out of game-clipping platform Medal, announced on June 25, 2026 that it raised $320 million at a $2.3 billion valuation. The round was led by Khosla Ventures, with participation from General Catalyst, Jeff Bezos, Eric Schmidt, F1 driver Nico Rosberg, and researchers from Google DeepMind and MIT, bringing total disclosed funding to $454 million. The company trains a single model on Medal's keypress-annotated gameplay footage—hundreds of millions of hours of player recordings with embedded action labels—and has demonstrated the same model playing Fortnite for 100 consecutive hours and controlling a quadruped robot walking autonomously with one camera. Researchers say fine-tuning the pretrained model to an unfamiliar robot requires only 8 minutes of real-world street data. Most of the new funding will go to compute via CoreWeave, with broader API access planned by late summer 2026.

On June 25, 2026, General Intuition—an embodied AI startup spun out of game-clipping platform Medal—announced a $320 million funding round at a $2.3 billion valuation. The round was led by Khosla Ventures, with participation from General Catalyst, Jeff Bezos, Eric Schmidt, F1 champion Nico Rosberg, and researchers from Google DeepMind and MIT. Combined with a $134 million seed round in October 2025, the company's total disclosed funding now stands at $454 million.

Most of the capital will fund compute expansion through a partnership with CoreWeave, focused on pretraining next-generation models. A portion is reserved for opening the API more broadly to external customers by late summer 2026.

What the company actually does

One model, two domains. When a TechCrunch reporter visited the New York R&D floor, a screen showed an AI agent that had played Fortnite continuously for 100 hours; a few meters away, a large quadruped robot walked autonomously through the office using a single camera, bumping into chair legs and trash cans but never stopping—matching CEO Pim de Witte's claim of a "single brain."
The data isn't video—it's keypresses. Medal hosts hundreds of millions of hours of player-uploaded gameplay footage, but what General Intuition consumes is the embedded action labels: which key was pressed on each frame, when, and why. De Witte argues that inferring actions from raw video is "not enough."
Only 8 minutes of real-world data needed. Researcher Josh Duplantis demonstrated that fine-tuning a Medal-pretrained model to an unfamiliar quadruped robot required just 8 minutes of real street data—not office data.
Not just robots. The same model has been tested on drones and driving games—anything controllable via keyboard, mouse, or gamepad.

Why this matters

The core thesis: embodied AI training data doesn't need to come from the physical world. Game keypress streams already contain the causal structure between self and environment. This contrasts with three mainstream embodied AI approaches:

Physical-world collection (Figure, Tesla Optimus, domestic humanoid robots): expensive real-robot demonstrations, slow to scale.
Simulation (NVIDIA Isaac/Cosmos, GR00T, Galaxea): high-fidelity rendering, but the sim-to-real gap remains hard.
Video self-supervision (Veo, Sora, Genie): learns world dynamics from video, but weak action grounding.

General Intuition's path is a fourth option: games as "video self-supervision with action grounding," preserving visual diversity while gaining scarce action labels for free. The company also uses its world model—generated frame-by-frame from Medal data—not as a product but as "the gym," an RL environment for training.

De Witte framed it as "the next stage of future pre-training": a single model that responds to both Fortnite screen input and real-world dynamics. Investor Vinod Khosla tied this to the emergence of "human-like intuition" in world models, analogous to the reasoning leap in LLMs.

Key takeaways

1. The "games to embodiment" path now has a $2.3B valuation stamp of approval, addressing the field's scarcest resource: high-quality action-labeled real-world data. 2. If the "8-minute fine-tune" scales, it could collapse embodied training economics from hours-level real-robot demonstrations to minutes-level fine-tuning. 3. The company positions itself as the embodiment sector's closest analogue to a foundation model company—no hardware, no single-scenario operations, just a general pretrained brain.

Risks and open questions

Sim-to-real ceiling. The 8-minute fine-tune was demonstrated on one robot in a structured office; transfer to stairs, outdoors, or collaborative settings is unproven, and no third-party replication exists yet.
Data coverage. Game scenarios (FPS, driving, platformers) may not extend to industrial, agricultural, or household long-tail tasks.
Compute dependence. Most of the $320M goes to CoreWeave compute; scaling must not break single-model multi-domain capability.
Ethical stance. The company commits to no lethal autonomous weapons and runs the Nerve platform to retrain gamers as robot teleoperators—a values choice that closes doors to defense customers.
Competitive field. Rivals like Odyssey ($1.45B valuation), Decart, and Fei-Fei Li's World Labs (Marble) are competing for the same "world model training ground" position; the landscape has not converged.

Bottom line: General Intuition is betting that the next stage of embodiment is pretraining. If it succeeds, robot developers may rethink building world models from scratch; if not, it becomes another high-profile 2026 sample that never fully crossed the sim-to-real gap.