Worldmodeldata raises £7 million to build the training data layer for physical AI

Worldmodeldata has emerged from stealth with a £7 million Seed round (approximately €8.2 million), backed by Iona Star Capital. The Cambridge company was founded in 2025 by Rhea Loucas (CEO) and is building what it describes as the largest library of video-game-generated training data for physical AI. The round closed in December 2025; the company announced it publicly in July 2026 after seven months of stealth. Lord Richard Allan, former Vice-President of Public Policy at Meta and Facebook, chairs the company.

What physical AI needs that the internet cannot provide

Language models were trained on the internet — an enormous, cheap corpus of human-written text that existed before the models were built. Physical AI — models that understand, predict, and plan interactions with the physical world — needs fundamentally different data: footage of how objects move, collide, fall, stack, and behave under physical constraints. That kind of data does not exist in documents, search indexes, or social media feeds.

Video games are an unusually efficient source. Modern game engines simulate physics in real time, produce photorealistic rendering, and generate data at scale — already encoded with the physical interactions that physical AI needs to learn from. Worldmodeldata's thesis is that this synthetic-but-physics-faithful data can serve as the substrate for training the next generation of embodied AI systems: robots, autonomous vehicles, and machines that operate in the physical world rather than the digital one.

The scale ambition is specific: 1 million hours of gameplay data by the end of 2026. That figure would represent one of the largest single corpora of physically-grounded synthetic training data assembled anywhere.

Seven months in stealth

The decision to close funding in December and announce only in July is worth noting. In a fundraising environment where many companies announce rounds before the ink is dry, sitting on a closed round for seven months is a signal. For a data infrastructure company whose product is the data, announcing without a library would be announcing without a product.

Gerry Buggy of Iona Star Capital frames the investment in architectural terms: "AI spent the last few years learning to describe the world. Now it needs to learn how to interact with it. Worldmodeldata is building the infrastructure to make that possible."

Loucas is equally direct about what the round is for: "World models represent a fundamental paradigm shift in AI. We are building the dataset infrastructure to make this technology real."

Whether gaming data generalises reliably to real-world physical AI — the texture and physics of a rendered game environment versus the noise and variability of the real world — is the technical open question the field has not yet settled. The commercial question is whether 1 million hours of gameplay is a meaningful fraction of what future physical AI training will require, or a starting point that future models will rapidly outscale. Both answers are worth watching.