Video games may hold the missing training data that helps AI models understand the physical world. A British startup called Worldmodeldata is turning game inputs and other information from video game studios into datasets designed to teach AI how actions create consequences.
That goal reaches beyond better game-playing systems. Worldmodeldata wants to improve world models, a class of AI systems designed to understand environments and navigate them. If the plan works, data built inside virtual worlds could become the foundation for models that operate in the physical world.
The Missing Link Is Cause and Consequence
World models need more than visual information. They need examples that connect an action with what happens next: a movement, a decision, a change in the environment, or an unexpected result. Worldmodeldata is aiming to solve that lack of cause and consequence data by packaging controller inputs and other data collected by video game studios into training datasets.
Video games offer an unusual advantage because they generate both the environment and the actions taken inside it. The data is available in the quantities needed for large AI training systems, and it covers enough different situations to capture important corner cases. Those unusual situations matter because a model trained only on familiar patterns can fail when the world behaves in an unfamiliar way.
“The corner cases are the ones to actually get right,” said Nicole Fraenkel, a partner at Khosla Ventures. That idea gives video game data a powerful role: it can expose AI models to a broad range of environments and outcomes before those models are optimized for a specific real-world setting or task.
A Data Broker for the World-Model Race
Worldmodeldata pitches itself as a broker that curates and organizes this information. Instead of asking AI labs to strike individual agreements with tons of game studios, the startup aims to bring that data together in training datasets that labs can use.
The company has licensed almost 1 million hours’ worth of data from studios behind various popular video games. That scale gives Worldmodeldata a large base for its strategy and shows why video game environments have become attractive sources of AI training material.
Worldmodeldata is not alone in collecting video game data. General Intuition and Niantic are already gathering data from their own platforms to build models. Nvidia is also developing world models, placing the field inside a broader push to create AI systems that can understand and respond to environments.
Rhea Loucas, CEO of Worldmodeldata, sees the opportunity in the sheer volume and variety of games. “There are millions of great games, and they are more and more similar to the real world,” Loucas said.
Why Scale Could Shape the Next AI Breakthrough
Researchers broadly assume that world-model performance will improve as training datasets grow, much as large language models improve with more training data. That belief makes the supply of useful data a central challenge, not a side issue.
Fei-Fei Li and Yann LeCun are among the researchers mentioned in connection with world models, while Khosla Ventures is part of the investment picture through Nicole Fraenkel. The interest reflects a simple question with major consequences: can AI learn enough about environments from virtual worlds to make sense of the physical ones?
Worldmodeldata expects video game data to make up the majority of training material for world models. Later, those models would be optimized with data tied to a particular real-world environment or task. That approach separates broad learning from focused adaptation, using games to build a wide base before applying more specific information.
Loucas described the possibility in ambitious terms: “This could well lead to the GPT moment for world models—making them really useful.”
Games Can Teach Movement, But Not Everything
The strategy also has limits. Video game environments can provide huge quantities of varied data, but virtual physics do not capture every detail of the physical world. That gap matters most when an AI system must manipulate objects, where small differences in force, contact, and movement can change the result.
Ming-Yu Liu offered a clear warning: “I would be more conservative on using video game data for manipulation. The physics for manipulation is much more involved.”
That distinction points to the likely path forward. Video game data can provide the broad training foundation, while data from a specific physical environment or task can refine the model where precision matters most. The games supply range; focused real-world data supplies the final layer of accuracy.
Worldmodeldata is betting that this combination can push world models toward practical use. With almost 1 million hours’ worth of licensed data, a broker model for game studios, and growing interest from companies including General Intuition, Niantic, and Nvidia, the next major AI training ground may be hiding inside the worlds people already play.
