Worldmodeldata is building a large, licensed dataset of action-conditioned gameplay data from Unreal and Unity to train world models that can predict and act in real-world environments.

The company licenses gameplay footage and engine data through partnerships with game developers and communities, ensuring high-quality, safer data for AI training without web scraping.

The goal is to assemble aligned datasets of actions, outcomes, and 3D states to enable AI models that understand environmental dynamics when specific actions are taken.

Leadership includes founder and CEO Rhea Loucas and Lord Richard Allan as chairman, marking a strategic UK anchor in sovereign AI and data infrastructure development.

A stated North Star is one million hours of training data, illustrating ambition while noting uncertainty about reaching it based on current milestones.

The firm targets reaching one million hours of data by the end of 2026, up from tens of thousands today, to support frontier labs, physical AI, and robotics.

Cambridge, UK, serves as the base to leverage local game industry ecosystems, with plans for a Series A in 2026 but no disclosed details yet.

The company emerged from stealth before the funding round, addressing a growing shortage of suitable training data for world models.

Business model centers on selling datasets to labs and companies building physical AI, robotics, and self-driving systems, rather than offering model development services.

As of now, Worldmodeldata has not generated revenue or finalized data contracts, though there is talk of a potential pipeline with a leading AI lab.

The market is competitive, with peers like Origin Lab and General AI Intuition; Worldmodeldata differentiates itself as a neutral data supplier rather than an in-house model builder.

The global market for synthetic and licensed training data is expected to grow significantly, underscoring strong demand for world-model data infrastructure.