Worldmodeldata secures £7M to develop AI training using video games.
The latest generation of AI has focused on describing the world, while the next must understand how the world reacts. A startup based in Cambridge believes the solution lies within video games.
Worldmodeldata has secured £7m (€8m) aimed at transforming gameplay into training data for AI, as reported by Tech Funding News. The seed funding round was led by Iona Star Capital from London. Lord Richard Allan, former public-policy head for Meta in Europe, is joining as chairman.
What they actually offer
Worldmodeldata targets a new type of AI known as a world model. Unlike chatbots that respond to text, a world model aims to foresee how an environment changes based on actions taken. This requires specific data: an action and its resulting consequence, matched frame by frame.
Video games inherently provide this data. Worldmodeldata acquires footage and engine data from games built on Unreal and Unity instead of scraping the internet. It then organizes the video, player inputs, and the underlying 3D state into structured datasets. Potential customers include laboratories developing physical AI, robots, and autonomous vehicles.
A significant claim, but a caveat
The emphasis is on a noteworthy goal: one million hours of gameplay data by the end of 2026. The company asserts this amount is 25 times larger than the current largest dataset, although this comparison is self-reported and lacks third-party validation.
Currently, the situation is more modest. Worldmodeldata has yet to establish finalized customer contracts or generate revenue, and its team consists of about ten individuals, including advisors and contractors. The seed funding round was completed back in December, seven months prior to this week’s announcement. Founder Rhea Loucas noted the company “decided to pause for a while” to determine the specific data needs of laboratories.
A competitive landscape
Other companies are pursuing similar concepts. Origin Lab in San Francisco raised $8m in May with over 20 publisher agreements. Meanwhile, General Intuition in New York has garnered $454m, though it keeps its data internally to train its own models.
Loucas positions Worldmodeldata as a neutral data supplier that caters to everyone, aligning it more with firms that test AI rather than those that build models.
The approach is intentionally European. The company is remaining in Cambridge, and Allan portrays this as part of the UK’s initiative for sovereign AI. Whether a ten-person team can accumulate a million hours of data before better-financed competitors remains uncertain.
For the time being, Worldmodeldata is providing tools for a gold rush that has hardly begun. The aspiration among its clients is that the resulting AI will truly understand cause and effect.
Other articles
Worldmodeldata secures £7M to develop AI training using video games.
The Cambridge-based startup Worldmodeldata has secured £7 million in seed funding, led by Iona Star Capital, to convert licensed video game data into training datasets for AI world models.
