On June 25, 2026, General Intuition, an embodied AI startup spun out of game clip 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 driver Nico Rosberg, and researchers from Google DeepMind and MIT. Including its $134 million seed round from October 2025, the company's total public funding now stands at $454 million.
The vast majority of the capital will go toward expanding compute capacity through a partnership with CoreWeave, focused on pretraining next-generation models. A portion is reserved to open the API more broadly to external customers by late summer 2026.
What Makes General Intuition Different
- One model, Fortnite and a quadruped robot. When a TechCrunch reporter visited the New York R&D floor, a screen showed an AI agent that had been playing 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 description of a "single brain."
- The data source isn't video—it's keystrokes. Medal hosts hundreds of millions of hours of player-uploaded gameplay footage, but what General Intuition actually consumes is the embedded action labels: which key was pressed on each frame, when, and why. De Witte believes inferring actions from video alone is "not enough."
- Only 8 minutes of real-world data. Researcher Josh Duplantis revealed in a TechCrunch demo that fine-tuning a Medal-pretrained model to an unfamiliar quadruped robot required just 8 minutes of real street data—street data, not office data.
- Not just robots. The same model has been tested on drones, driving games, and other scenarios—anything controllable "by keyboard, mouse, or gamepad."
- Physical-world data collection (Figure, Tesla Optimus, Chinese humanoids): real-machine teleoperation, expensive per-sample data, slow to scale.
- Simulation (NVIDIA Isaac/Cosmos, GR00T, Galaxea): high-fidelity rendering, but the sim-to-real gap remains an engineering challenge.
- Video self-supervision (Veo, Sora, Genie): learns world dynamics from video, but with weak action grounding.
- Sim-to-real ceiling: the 8-minute fine-tune was on one robot in a relatively structured office; stairs, outdoors, and collaboration remain unproven. TechCrunch itself notes that scaling sim-to-real transfer "is a question nobody can fully answer yet."
- Data moat vs. data breadth: game scenarios are limited (FPS, driving, platformers); coverage of industrial, agricultural, and household long-tail settings is key.
- Compute dependence: with most of the $320M going to CoreWeave compute, maintaining single-model multi-scenario capability will require renewed coordination of compute, data, and architecture.
- Productized ethics: the company has publicly committed to "no lethal autonomous weapons" and launched Nerve, letting gamers transition into robot teleoperation—values that may close doors with defense customers.
- No independent verification: the "8-minute fine-tuning" and "100-hour gameplay" figures are company claims without third-party replication. Competitors including Odyssey ($1.45B valuation), Decart, and Fei-Fei Li's World Labs Marble are vying for the same "world model training ground" position.
Strategic Positioning
The core thesis: embodied AI training data doesn't need to come from the physical world—game keystroke streams already contain the causal structure of self-environment interaction. This contrasts with three mainstream approaches:
General Intuition represents a fourth path: games as "video self-supervision with action grounding," preserving visual diversity while freely acquiring scarce action labels. De Witte called it "the next stage of future pre-training—a single model that can respond to Fortnite screen information and act, and also respond to real-world dynamics, which LLMs can never do." Investor Vinod Khosla framed it as the emergence of human-like intuition in world models.
Notably, the company's world model (a frame-by-frame simulated environment built from Medal data) is not a product but "the gym"—an RL training environment rather than a consumer offering.
Why It Matters
1. The "games → embodied" path now has a $2.3B valuation backing it, addressing the sector's scarcest resource: high-quality, action-labeled real-world data. 2. If the "8-minute fine-tuning" claim scales, it could upend embodied training economics, moving data costs from hours of real-machine demonstrations to minutes of fine-tuning. 3. It offers a real data point for the game-synthetic-data route—a middle ground between pure video and pure simulation, potentially reusable across the whole embodied training stack. 4. Unlike Figure 03, Tesla Optimus, or Unitree, General Intuition builds no hardware and no single-scenario operations—making it the closest sample to a "foundation model company" in embodied AI.
Risks and Open Questions
> Source: TechCrunch