The details

• Foundation Models: The platform includes world foundation models (WFMs) that generate synthetic training data and create physics-based videos from text, images, video, and sensor inputs, focusing on industrial and driving scenarios.

• Processing Pipeline: Using NeMo Curator, the platform reduces video processing from 3+ years to 14 days on the NVIDIA Blackwell platform. The Cosmos Tokenizer provides 8x better compression and 12x faster processing.

• Industry Adoption: Leading companies including 1X, Agile Robots, Figure AI, Uber, and Waabi are using Cosmos for robotics and autonomous vehicle development.

"The ChatGPT moment for robotics is coming. Like large language models, world foundation models are fundamental to advancing robot and AV development, yet not all developers have the expertise and resources to train their own," said Jensen Huang, founder and CEO of NVIDIA.

Why it matters

NVIDIA's platform addresses challenges in physical AI by cutting costs and simplifying data needs. Through its open foundation models and streamlined tools, NVIDIA lets more developers build advanced robots and self-driving vehicles without massive budgets or specialized expertise.