Worldmodeldata from Cambridge raised £7 million to train AI on video games

Worldmodeldata, a neural network for video games, has attracted £7 million in investment for training AI: the company transforms gameplay into data for training a new class of artificial intelligence — world models that not only describe the world, but understand how it responds to actions.

Why World Models Need Different Data

Large language models — ChatGPT, Claude, Gemini — were trained primarily on text. They have mastered language, knowledge, and reasoning logic. But this approach has a fundamental limit: such systems do not have built-in understanding of physics.

"World models" — the next frontier. These are AI systems capable of predicting the consequences of physical actions: what happens if you push an object, pull a lever, or throw an item from height. Such understanding is necessary for robots, industrial agents, and any systems operating in real, rather than virtual space.

The main difficulty is data. Recording real physical interactions at scale is very expensive: it requires robots with sensors, laboratories, qualified operators, and thousands of hours of experiments in diverse conditions. Worldmodeldata has found an alternative path.

Why Video Games Are a Data Source for AI

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

Diversity is also important: modern games create thousands of environment and object types — from driving simulators to strategies with physical systems. This is a ready-made database of interactions with the physical environment, already structured by game logic itself.

Key facts of the round:

• Amount: £7 million (~€8 million)
• Type of funding: seed round
• Lead investor: Iona Star Capital, London
• Startup: Worldmodeldata, Cambridge, United Kingdom

What This Means

The market for training data for AI is transforming. After years of text corpus and multimodal dataset eras, major laboratories — Google DeepMind, Meta, OpenAI — are actively researching world models for robots and agent systems. Demand for physically accurate and interactive training data will grow alongside the market for humanoid robots and autonomous agents.

Worldmodeldata's bet is simple: accumulated decades of gameplay — this is a scalable source of data about physical interaction that does not exist anywhere in comparable volume. If the forecast is correct, video games will transform from entertainment into the infrastructure for training the next generation of AI.

*Meta is recognized as an extremist organization and is banned in Russia.

What are world models in AI?

These are a new class of neural networks that not only describe the world, but understand how it responds to actions. Unlike large language models (ChatGPT, Claude, Gemini) trained on text, world models must reproduce physics and the logic of interaction between objects in the surrounding world.

Why does AI need data from video games?

Video games contain billions of gameplay frames with known world physics and environment reactions to player actions. This is an ideal source of training data for neural networks to learn causal relationships and predict event progression.