- Worldmodeldata closed its £7 million seed round last December, but the Cambridge startup waited seven months before announcing it publicly this week.
- The company aims to collect one million hours of gameplay data by the end of 2026, but so far, it has no finalised customer contracts and a team of about ten people, including advisors and contractors.
- Worldmodeldata is joining a fast-changing industry. Since its fundraising, competitors Origin Lab and General Intuition have raised $8 million and $454 million for similar gaming data projects.
Worldmodeldata, a Cambridge-based deeptech startup, aims to build the largest library of AI training data from video games, targeting 1 million hours by the end of 2026. So far, it hasn’t made any sales.
The UK startup just announced its £7 million seed, led by Iona Star Capital. Lord Richard Allan, Meta’s former vice president of public policy for Europe, the Middle East and Africa, joined as non-executive chairman.
The round actually closed in December 2025. Founder and CEO Rhea Loucas tells Tech Funding News, “We decided to stay still for a little bit.” She said the company used this time to determine which data was needed for AI training before launch.
Gerry Buggy, co-founder of Iona Star, adds, “That coupling between action and consequence is the scarcest resource in AI today. Worldmodeldata is manufacturing that missing ingredient, and every company building physical AI or digital AI will eventually face the same challenge.”
Loucas agreed, saying, “It’s not hard, I would say, compared to others. We didn’t do a roadshow. When we talked to our current investor, they saw this, and it was done very quickly.”
The story behind Worldmodeldata
Loucas spent four to five years in the video games industry before starting the company. The idea came from seeing large language models hit their limits in late 2024.
“Many capabilities we’re lacking, like common sense and causality understanding, don’t come from pure text token understanding,” she says.
World models, which are AI systems trained to predict how environments respond to actions rather than just reacting to inputs, require a different kind of data: recorded actions and their outcomes. Loucas notes that this data is rare outside video games, where billions of players naturally create a record of actions, video frames, and engine states.
The company published a paper on its concept in August 2025, was founded soon after, and closed its seed round just a few months later.
Competition is a key issue
Seven months later, the industry has changed a lot. San Francisco’s Origin Lab raised an $8 million in a seed round in May for a similar idea, licensing gameplay data from studios to sell to AI labs. It says it already has partnerships with over 20 publishers. New York’s General Intuition has raised $454 million across two rounds in less than a year, backed by its Medal platform, which has tens of millions of monthly users.
Loucas points out a key difference. “General Intuition is a spin-off from Medal AI. They keep all their data in-house, train their own AI models on it, and then sell the models. They’re competing with the likes of OpenAI or Anthropic. We are data providers, more like Scale AI. Our job is to provide high-quality, video game-generated data to the labs that need it. General Intuition doesn’t provide its data to anyone,” she explains.
This difference matters for anyone trying to understand the busy AI infrastructure sector, which TFN has covered in reports on Odyssey’s $310 million round and Encord’s $60 million Series C for physical AI data infrastructure.
Still, the main question is: who is paying Worldmodeldata for its data?
The global synthetic and licensed training data market was worth an estimated $710 million in 2026 and is projected to reach $3.67 billion by 2031, at a 39% compound annual growth rate, according to Mordor Intelligence.
No contracts yet, a small team, and one big goal
Loucas says the company has a customer pipeline and is in close discussion with one of the world’s leading labs, but no contracts have been finalised, and she did not share any names.
The company has not yet started earning revenue. The team is about ten people, including employees, advisors, and contractors, and most of the new funding will go toward hiring.
This shows a gap between the public goal of one million hours, which Loucas says is 25 times bigger than the current largest dataset at 40,000 hours, though this number comes from the company and isn’t independently verified, and the company’s current situation: no finalised contracts, a ten-person team, and a seed round that’s already seven months old, while competitors have raised much more in less time.
Loucas calls the one million-hour goal deliberate positioning: “The one million is kind of our north star. That says a lot.” Whether this means rapid growth or just big ambitions will become clearer in the next few quarters.
The company will stay based in the UK and Europe by choice. Cambridge was picked on purpose, and Loucas believes the UK’s video games industry offers the right ingredients. The company may serve global customers once the product is ready, but Loucas declined to comment on plans for a 2026 Series A.
