The pitch sounds straightforward enough: extract precision data from video games, license it to AI labs, and watch world models learn how physics actually works. Simple. Except Worldmodeldata, a Cambridge startup that surfaced publicly earlier this month, closed a £7 million seed round in December—and as of its July 6 announcement, had not a single finalized customer contract.
Not one.
Founded by Rhea Loucas and incorporated last August, the company spent the better part of seven months in quiet mode before its July 6 announcement—time, the founders say, spent refining a product that didn't yet have paying users. The round was led by Belfast-based Iona Star Capital, with Gerry Buggy from the fund taking a board seat. No other investors were named in the disclosure, which is perhaps more telling than the company would prefer.
Lord Richard Allan, Meta's former VP of Public Policy for EMEA and a onetime UK Member of Parliament, signed on as Chairman. Companies House records show Allan joined the board back in November, well before the public rollout. His involvement lends the venture a certain gravitas—though whether that translates to traction with AI labs remains an open question.
The Technical Pitch
What Worldmodeldata is actually selling is this: licensed gameplay data pulled directly from game developers through formal agreements, not scraped or reverse-engineered. The datasets include 1080p multi-view video synchronized frame-by-frame with controller inputs, per-entity 3D positions, rotations, velocities, semantic labels, camera pose, and depth information. All of it delivered in formats like Parquet, HDF5, and JSON-L—the kind of structured, ground-truth data that AI researchers building world models ostensibly crave.
The target customers are labs working on physical AI systems, robotics, autonomous vehicles, and anything else that requires a model to understand how objects move through space and time. Unlike competitors such as General Intuition—which trains its own models using gameplay data—Worldmodeldata positions itself as a neutral infrastructure provider. It'll license to anyone willing to pay, assuming the anyone in question actually materializes.
The company claims the largest publicly available gameplay dataset today contains around 40,000 hours of footage. Worldmodeldata says it aims to assemble a library exceeding 1 million hours by year-end, though this target remains unverified. Multiple outlets noted at the time of the announcement that the company declined to specify how many licensing deals are currently in negotiation.
A Crowded, Capital-Heavy Space
Worldmodeldata's emergence coincides with a broader frenzy around world models—the AI systems that learn to predict how environments evolve over time. Capital has poured in. San Francisco's Origin Lab raised $8 million from Lightspeed in May to license game worlds to AI builders. General Intuition, which takes a different approach by training its own models atop gameplay data, closed a $320 million Series A in June at a roughly $2.3 billion valuation. Odyssey pulled in $310 million the same month to accelerate simulation research. Paris-based AMI Labs secured a staggering €1.03 billion in March.
The timing suggests venture investors see licensed gameplay data as foundational infrastructure for the next wave of AI development—akin to what scraped web text was for large language models. But there's a wrinkle. The economics of data licensing at scale remain murky, especially when game developers hold the keys and frontier labs are notoriously price-sensitive about infrastructure costs.
What Happens Next
With roughly 10 people on the team as of July—including advisors and contractors—Worldmodeldata operates out of The Bradfield Centre in Cambridge Science Park. The £7 million (approximately €8.2 million) will fund product development, team expansion, and the licensing deals required to hit that 1 million hour target. A browser demo called "Cat Field" showcases the company's capture pipeline, though enterprise trials are locked behind request-access forms.
The real test, of course, is conversion. Can the startup convince game developers to license their proprietary data? And will frontier AI labs—already spending heavily on compute and talent—pay for curated gameplay datasets when alternatives exist?
For now, Worldmodeldata is making a calculated wager: that gameplay data becomes as indispensable to spatial AI as ImageNet once was to computer vision. The company has the capital and a well-connected chairman. What it doesn't have yet is revenue, customers, or proof that the market it's building for will actually pay. That's a gap £7 million is meant to close, preferably before the runway runs out.
