AI World Models: How Video Games Train Neural Networks

Worldmodeldata, a startup building on video games, has raised £7 million to train AI. The company turns gameplay into data for a new class of artificial intelligence, world models, which do not merely describe the world but understand how it reacts to actions.

Why world models need different data

Large language models such as ChatGPT, Claude and Gemini were trained mainly on text. They picked up language, knowledge and reasoning patterns. But this approach has a fundamental limit: such systems have no built-in understanding of physics.

"World models" are the next frontier. These are AI systems able to predict the consequences of physical actions: what happens if you push an object, pull a lever or drop something from a height. This kind of understanding is essential for robots, industrial agents and any system that operates in real, rather than virtual, space.

The main difficulty is data. Recording real physical interactions at scale is very expensive: it takes robots with sensors, laboratories, skilled operators and thousands of hours of experiments in varied conditions. Worldmodeldata found an alternative route.

Why video games are a source of data for AI

Game engines are essentially physics simulators. They model gravity, collisions, fluid dynamics, material elasticity and the behavior of objects under load in real time. Every hour of gameplay contains millions of "action → environment response" pairs: a car hits a wall and deforms according to the laws of physics, a character jumps and lands at a predictable speed, a liquid spills and spreads according to simulated hydrodynamics.

Variety matters too: modern games create thousands of types of environments and objects, from driving simulators to strategy games with physical systems. That is a ready-made database of interaction with a physical environment, already structured by the game logic itself.

Key facts of the round:

• Amount: £7 million (~€8 million)
• Funding type: seed round
• Lead investor: Iona Star Capital, London
• Startup: Worldmodeldata, Cambridge, UK

What this means

The market for AI training data is changing. After years dominated by text corpora and multimodal datasets, the biggest labs, including Google DeepMind, Meta and OpenAI, are actively researching world models for robots and agentic systems. Demand for physically faithful, interactive training data will grow in step with the market for humanoid robots and autonomous agents.

Worldmodeldata's bet is simple: decades of accumulated gameplay are a scalable source of physical-interaction data that exists nowhere else at comparable volume. If the forecast holds, video games will go from entertainment to training infrastructure for the next generation of AI.

*Meta has been designated an extremist organization and banned in Russia.

What are AI world models?

They are a new class of neural networks that do not merely describe the world but understand how it reacts to actions. Unlike large language models (ChatGPT, Claude, Gemini), which were trained on text, world models are meant to reproduce the physics and logic of interaction between objects in the surrounding world.

Why does AI need data from video games?

Video games contain billions of frames of gameplay with known world physics and known environmental reactions to player actions. That makes them an ideal source of training data for neural networks to learn cause-and-effect relationships and predict how events unfold.