The next generation of artificial intelligence cannot simply describe the world; it must learn to be pushed back by it. Cambridge-based startup Worldmodeldata has just raised £7 million (about €8 million) in a seed round led by Iona Star Capital to turn video games into interactive training data for AI. The idea is as simple as it is ambitious: video games provide one of the few environments where a model can experience cause and effect in real time, receiving immediate feedback from the system — exactly the kind of learning that models trained solely on static text, images or videos lack.

For years, generative AI has fed on enormous textual and visual datasets, learning linguistic patterns and recognising objects. But to operate in physical contexts — robotics, autonomous driving, industrial operations — you need an understanding of the world's dynamics: how an object rolls, how a surface reacts to pressure, how an environment changes in response to an action. Large Language Models (LLMs) excel at mimicking language but remain blind to physical causality. Video games, especially those with advanced physics engines, become a training ground for models that will have to interact with reality. It is not just about generating scenarios: every frame becomes an opportunity for reinforcement learning, with a closed loop of action-reaction impossible to obtain from traditional datasets.

Worldmodeldata signals a structural paradigm shift: training moves from centralised, static datasets to distributed simulation platforms. For companies developing robots or embedded systems, this opens a strategic path: generating training data in-house, using simulations customised for their own operational environments — factories, warehouses, construction sites. Those adopting an on-premise approach can control the entire training cycle, protecting the intellectual property of the simulation data (the layout of a warehouse, the mechanics of a production process) without having to transfer it to external clouds. In this scenario, hardware for graphics rendering and physical simulation — GPUs with high VRAM, self-hosted inference and training clusters — becomes a critical asset, no longer confined to research labs.

The ripple effect also concerns cloud service providers: if simulated data generation and training of interaction models become workloads that can be managed locally, the value of the cloud shrinks to a computational commodity, while differentiation shifts towards data quality and ownership. On the sovereignty front, regulated sectors such as manufacturing, logistics and defence could find in the on-premise approach the answer to GDPR and data residency requirements for sensitive operational data. At the same time, a new competition emerges among simulation engines: whoever offers the most realistic and customisable physical environments could become the main gatekeeper of embodied AI training, displacing traditional big data accumulators.

Who wins and who loses? Beneficiaries are hardware manufacturers for high-performance workstations and servers, as well as startups developing orchestration frameworks for local workloads. Storage providers and static data annotation platforms risk seeing their markets eroded, while large cloud platforms may have to rethink their AI factory services to integrate simulation pipelines instead of simple labeling.

The starting point is a seed investment, but the question it raises is far from small: are we ready to rethink AI infrastructure to support worlds that push back?