General Intuition, the New York startup developing AI systems that learn how to navigate space and time, is in talks to raise about $300 million in new funding, according to people familiar with the discussions.

The round would value the company at a little over $2 billion, the sources said. If completed, the deal would mark a major jump for the company, which spun out of Medal only eight months ago with a $134 million seed round.

The startup is led by Medal co-founder Pim de Witte, who founded General Intuition with researchers Eloi Alonso, Adam Jelley, and Vincent Micheli. The company is focused on foundation models for embodied AI and world modeling, with the goal of teaching agents how to understand and act in physical or simulated environments.

A key part of the company’s pitch is its access to Medal’s large trove of video game footage. General Intuition says it trains on a dataset built from 2 billion videos a year generated by Medal’s 10 million monthly active users. The company argues that first-person gameplay data offers a useful training ground for systems that need to learn spatial and temporal reasoning, including the ability to perceive what is happening, predict what comes next, and respond in real time.

Sources said the fundraising effort has drawn interest from prominent backers including Jeff Bezos and Eric Schmidt, along with existing investors Khosla Ventures and General Catalyst. The company’s dataset has also reportedly attracted attention from OpenAI, which previously tried to buy Medal, according to reporting cited by TechCrunch. The source material also says other major AI labs have shown interest.

General Intuition is entering a crowded and fast-moving segment of the AI market. Recent releases from Runway, Decart, and World Labs have added momentum to the world model category, while Google’s Genie 3 has recently incorporated Google Maps data to improve real-world simulation.

Still, General Intuition is taking a somewhat different approach from some rivals. Rather than focusing on selling world models directly, the company’s strategy is to build models that can train agents, with those agents serving as the product. The startup believes its unique data advantage gives it a path toward commercial viability.

The company plans to use the new capital to expand computing capacity, with a new product expected by the end of summer or early fall, according to a source familiar with the matter.

The fundraising discussions remain ongoing, and terms could still change before any deal is finalized.