AI Giants Race to Build World Models for Physical Intelligence

World models seek to move artificial intelligence beyond predicting the next word by learning how spaces, objects and actions interact, then simulating what happens next. The approach can let autonomous vehicles and robots rehearse decisions in synthetic environments before acting in the real world, reducing the cost and danger of trial and error. Researchers view that capacity for spatial reasoning, physical causality and planning as a potential bridge from large language models to embodied AI and, eventually, artificial general intelligence.

The race intensified in 2025. NVIDIA unveiled its Cosmos world-foundation-model platform at CES in January, while Google DeepMind announced Genie 3 on Aug. 5, generating interactive 720p environments at 24 frames per second. Funding followed: Fei-Fei Li’s World Labs raised $1 billion in February 2026, and Yann LeCun’s AMI Labs secured $1.03 billion in March. The investments underscore competing bets on spatial intelligence and predictive architectures, though visual realism still does not guarantee accurate physics, and training costs, data quality and safety remain unresolved.