Worldmodeldata raises £7M to train AI with video games
▼ Summary
– Worldmodeldata raised £7m from Iona Star Capital to turn gameplay footage into training data for “world model” AI that predicts how environments change in response to actions.
– The company licenses video and engine data from games built on Unreal and Unity, packaging player inputs and 3D state data for labs building physical AI, robots, and self-driving cars.
– It aims to collect one million hours of gameplay data by 2026, claiming that is 25 times the largest existing dataset, but this comparison is unverified.
– The startup has no finalized customer contracts, no revenue, and a team of about ten people, having closed its seed round seven months before the announcement.
– Competitors like Origin Lab and General Intuition are also pursuing this idea, but Worldmodeldata positions itself as a neutral data supplier focused on European sovereign AI.
A Cambridge-based startup believes the key to building smarter artificial intelligence lies not in scraping the internet, but inside the interactive worlds of video games. Worldmodeldata has secured £7 million (€8 million) in seed funding to transform gameplay into structured training data for a new generation of AI systems. The round was led by London’s Iona Star Capital, with Lord Richard Allan, formerly Meta’s top public-policy executive for Europe, stepping in as chairman.
The company is targeting a specific breed of AI known as a world model. Unlike a standard chatbot that simply processes text, a world model attempts to predict how an environment will change in response to an action. This requires a unique type of data: a precise, frame-by-frame record of an action and its direct consequence. Video games naturally generate this exact pairing.
Rather than pulling information from the open web, Worldmodeldata licenses footage and engine data from games built on Unreal and Unity. It then compiles the video, player inputs, and the underlying 3D state into clean, usable datasets. The primary customers are labs working on physical AI, robotics, and autonomous vehicles.
The startup’s central pitch revolves around a bold target: collecting one million hours of gameplay data by the end of 2026. The company claims this would be 25 times larger than any existing dataset, though that comparison is self-reported and remains unverified.
The current reality is more modest. Worldmodeldata has no finalized customer contracts, no revenue, and a team of roughly ten people, including advisors and contractors. The seed round actually closed back in December, nearly seven months before this week’s public announcement. Founder Rhea Loucas explained the quiet period, saying the company “decided to stay still for a little bit” to better understand which specific data formats the labs actually require.
The race is already crowded. San Francisco’s Origin Lab raised $8 million in May, boasting over 20 publisher partnerships. Meanwhile, New York-based General Intuition has pulled in $454 million, though it keeps its data proprietary to train its own models. Loucas positions Worldmodeldata differently, as a neutral data supplier that sells to everyone, similar to a stress-testing firm rather than a model builder.
The company is also making a deliberate European bet. It remains based in Cambridge, and Allan frames this as part of the UK’s broader push for sovereign AI capabilities. The open question is whether a ten-person team can realistically reach a million hours of data before better-funded competitors close the gap. For now, Worldmodeldata is selling a shovel for a gold rush that has only just begun. The prize, its customers hope, is an AI that finally understands cause and effect.
(Source: The Next Web)
