Cambridge’s Worldmodeldata raises £7M to turn video games into AI training data

The last wave of AI learned to describe the world. The next needs to learn how the world pushes back. A Cambridge startup thinks the answer hides inside video games.

Worldmodeldata has raised £7m (€8m) to turn gameplay into AI training data, Tech Funding News reported. London’s Iona Star Capital led the seed round. Lord Richard Allan, Meta’s former public-policy chief for Europe, joins as chairman.

What it actually sells

The target is a new kind of AI called a world model. Where a chatbot reacts to text, a world model tries to predict how an environment changes when you act on it. That demands a rare kind of data: an action, and the consequence that follows, aligned frame by frame.

Games happen to record exactly that. Worldmodeldata licenses footage and engine data from titles built on Unreal and Unity, rather than scraping the web. It then packages the video, the player inputs and the underlying 3D state into clean datasets. The buyers would be labs building physical AI, robots and self-driving cars.

A big number, and a catch

The pitch centres on one figure: one million hours of gameplay data by the end of 2026. The company says that is 25 times the largest existing dataset, though that comparison is its own, and unverified.

The reality today is smaller. Worldmodeldata has no finalised customer contracts and no revenue, and its team runs to about ten people, including advisors and contractors. The seed even closed back in December, seven months before this week’s announcement. Founder Rhea Loucas says the company “decided to stay still for a little bit” to work out which data the labs actually need.

A crowded, fast-moving race

Others are chasing the same idea. San Francisco’s Origin Lab raised $8m in May with more than 20 publisher deals. New York’s General Intuition has pulled in $454m, though it keeps its data in-house to train its own models.

Loucas frames Worldmodeldata differently, as a neutral supplier that sells data to everyone, closer to the firms that stress-test AI than to a model builder.

The bet is European by design. The company is staying in Cambridge, and Allan casts that as part of the UK’s push for sovereign AI. Whether a ten-person team can reach a million hours before better-funded rivals is the open question.

For now, Worldmodeldata is selling a shovel for a gold rush that has barely begun. The gold, its customers hope, is an AI that finally grasps cause and effect.