General Intuition raises $320M at $2.3B, using gameplay

General Intuition said it raised $320 million at a $2.3 billion valuation to train agentic world models using gameplay data from Medal, and to scale compute and product access. The new round brings the startup’s total disclosed funding to $454 million after a $134 million round at launch last October.

How does General Intuition train agents?

General Intuition pre-trains a single agentic model on Medal gameplay footage that includes action labels, then refines that model in simulated "world models" and with small amounts of real-world data. The company says Medal supplied "hundreds of millions of hours" of uploaded gameplay together with embedded records of the exact buttons players pressed and when, and the startup demonstrated an agent that played for "100 hours straight" and a quadruped robot fine-tuned with just "eight minutes" of street-collected robotics data.

The firm treats its frame-by-frame simulated environment as a training gym rather than a product, and claims its model learned spatial-temporal rules such as walls blocking movement and ladders enabling climbing. Founders say action labels distinguished the agent's "self" from the environment and helped the model infer causality in ways video-only approaches cannot.

Who backed the $320M round and what will it fund?

Khosla Ventures led the round, with participation from General Catalyst, Jeff Bezos, Eric Schmidt, Nico Rosberg, and researchers at Google DeepMind and MIT. Most of the new capital will go toward scaling compute capacity, the company said, and it has a deal with CoreWeave to support pre-training of the next model. A portion of funding is earmarked to make an API broadly available by the end of summer and to onboard customers across gaming, simulation, and robotics.

Co-founders named in company materials include Pim de Witte, Eloi Alonso, Adam Jelley, and Vincent Micheli. De Witte, who is 31 years old, spun the startup out of Medal and emphasized the value of Medal’s proprietary dataset as a competitive moat. Vinod Khosla said human action and reaction data in games are "the key part to the emergence of intuition," adding that when reasoning emerged in LLMs it represented "a quantum leap."

Why it matters

General Intuition attempts to shortcut the expensive, slow process of collecting vast real-world robotics data by leveraging large-scale game action datasets plus simulation. If the simulation-to-robot transfer holds broadly, it could reduce the need for lengthy physical data collection and make generalized agents easier to train across embodiments. The company already shows cross-embodiment attempts beyond a quadruped, including drones and driving-game tests, and says its model works with any device controllable by a game controller or keyboard and mouse.

The bet is also strategic. Backers and founders frame Medal’s human-action logs as proprietary data that could create a lasting advantage: the startup declined acquisition offers and plans to position its model as an enabling foundation for other companies rather than building verticals like self-driving cars itself. The company has also launched Nerve, a jobs marketplace that lets gamers earn money starting with data labeling and potentially moving toward robot teleoperation.

What to watch

Watch for the company’s API rollout slated by the end of summer and for technical details about the next pre-training run on CoreWeave. Two concrete signals will matter: which customers adopt the API across robotics and simulation, and whether small amounts of real-world data consistently fine-tune agents trained in gameplay and simulated world models at scale.