AI could take a keen interest in your video game sessions.

Worldmodeldata would have passed the threshold of one million hours of licensed video game gameplay, three months ahead of the target it had set for the end of 2026. Founded last year in Cambridge by Rhea Loucas, it emerged from stealth on July 6 with a £7 million seed funding round, approximately €8 million, led by the Iona Star fund.It turns gameplay into training manuals for AI systems designed to act in the physical world, from factory arms to self-driving cars.

What a controller learns that the internet ignores

Each hour recorded in games running on Unreal or Unity captures the footage, player inputs and exact state of the game engine, including object positions, frame by frame. Language models have fed on text from the web; systems designed to operate a robotic arm or a vehicle need examples that link an action to its outcome—material that is almost impossible to find online. A world model, the family of AI targeted by these datasets, learns like a child in front of a tower of blocks: it associates a push with a fall, then generalizes to anything that can be pushed.

The largest datasets of this type had previously topped out at around 40,000 hours: the amount claimed multiplies the order of magnitude by 25. V-JEPA 2, published by Meta in June 2025, validated the formula: more than one million hours of video processed, followed by fine-tuning with just 62 hours of real-world robotic manipulation. Transferring these capabilities to the physical world nevertheless still divides researchers, with fine object manipulation remaining the main stumbling block.

Selling to all sides: Cambridge’s bet

General Intuition, its New York-based competitor, raised $454 million and keeps its data to train its own models.Nvidia feeds its own models with Cosmos, an in-house engine that reproduces the physics of the real world rather than that of games. Worldmodeldata takes the opposite approach as a neutral supplier: its raw material is for sale to all laboratories, with no proprietary model to defend.

At the time of its funding round, the fledgling startup of around ten people had not signed a single customer contract, a commercial gamble that remained entirely unproven. It had, however, attracted a heavyweight chairman in the person of Lord Richard Allan, Meta’s former public policy chief.

If the technical promise holds, an automaker will be able to simulate millions of pedestrian crossings without driving a single car. A home-robot manufacturer, meanwhile, will teach its machines to pick up a glass without breaking any crates. As early as March, the CNIL had devoted a feature to world models, highlighting the provenance of the data, and the rights that go with it, as the sector’s blind spot.

👉🏻 Follow tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.