NVIDIA Introduces Cosmos to Accelerate Physical AI Development

NVIDIA Introduces Cosmos to Accelerate Physical AI Development

At CES 2025, NVIDIA introduced Cosmos, a comprehensive platform designed to accelerate the development of physical AI systems through advanced simulation and world modeling capabilities. The platform marks NVIDIA's latest push to make autonomous vehicle and robotics development more accessible to developers worldwide.

Key Points:

  • Cosmos introduces world foundation models for physics-based simulations
  • The platform is available under an open model license, making it accessible to developers of all sizes
  • Major companies including Uber, Waabi, and Hillbot are already implementing Cosmos in their development pipelines

Physical AI applications like robots and autonomous vehicles have traditionally faced significant hurdles in development, primarily due to the massive amounts of real-world data needed for training and the computational resources required to process it. NVIDIA's release of Cosmos aims to address these challenges head-on by providing pre-trained models that can generate physics-aware synthetic data and predict how objects interact in virtual environments.

The platform's core offering is a suite of world foundation models (WFMs) trained on an impressive 9,000 trillion tokens from 20 million hours of real-world data. These models come in three variants - Nano for edge deployment, Super for baseline performance, and Ultra for maximum fidelity. Each variant is designed to serve different use cases, from real-time inference to high-quality synthetic data generation.

Several major players in the physical AI space are already putting Cosmos to work. Waabi, an autonomous vehicle company, is evaluating the platform for data curation and simulation development. Robotics startup Hillbot is using Cosmos to generate terabytes of high-fidelity 3D environments for robot training. Uber is partnering with NVIDIA to accelerate autonomous mobility development by combining their rich driving datasets with Cosmos's capabilities.

What sets Cosmos apart is its focus on practical applications. The platform includes tools that can process 20 million hours of video data in just 14 days using NVIDIA's latest Blackwell GPUs – a task that would typically take over three years using traditional CPU-based systems. Its video tokenizers achieve 8x more compression than current methods while processing data 12x faster.

NVIDIA has also addressed safety and responsibility concerns by building in guardrails that help mitigate harmful inputs and screen generated content. The platform includes watermarking capabilities for AI-generated sequences, promoting transparency in synthetic content creation.

"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, during his CES keynote.

The move to make Cosmos freely available under an open model license marks a significant shift in democratizing physical AI development. By removing barriers to entry, NVIDIA is enabling smaller companies and independent developers to compete in a field traditionally dominated by well-resourced organizations.

Chris McKay is the founder and chief editor of Maginative. His thought leadership in AI literacy and strategic AI adoption has been recognized by top academic institutions, media, and global brands.

Let’s stay in touch. Get the latest AI news from Maginative in your inbox.

Subscribe