Nvidia's world model lead says models trained on video game inputs are unlikely to do well at fine-grained motor control such as object manipulation.

“Ming-Yu Liu, who leads world model development at Nvidia, says that models trained on video game inputs are unlikely to fare well with tasks that require fine-grained motor control, like the careful manipulation of objects.”

• Worldmodeldata, which raised a GBP 7M seed in July 2026 to supply gameplay data for world models, says it has licensed almost 1 million hours of data from studios behind popular video games; it names no studio and states no price, and paying individual players is a plan, not yet an arrangement.
• Worldmodeldata is packaging controller inputs and other data collected by video game studios to fill the shortage of world-model training data.
• Worldmodeldata pitches itself as a broker that curates studio data so labs avoid striking individual agreements with many game studios.
• General Intuition and Niantic are already collecting video game data from their own platforms to build models.
• WIRED reports there is no corpus of training material for world model labs comparable to the text available to language models.
• Worldmodeldata's CEO believes video game data will make up the majority of world-model training material, later tuned with task-specific data.

• Datasets, tasks and environments

It is shown on its own, not counted into a figure.

Cited in this guide

• Selling Game Code and Game Data to AI Companies

• All supply facts
• As a table, with the rest of the market
• How every number on the site traces to its records