AWS and NVIDIA partner on SageMaker HyperPod model factories for Cosmos 3 physical AI
The collaboration details how developers can deploy NVIDIA’s Mixture-of-Transformers world model on high-bandwidth AWS clusters to train autonomous robots.
Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.
- AWS and NVIDIA have detailed an infrastructure blueprint for deploying NVIDIA Cosmos 3 on Amazon SageMaker HyperPod clusters.
- NVIDIA Cosmos 3 is an 'omnimodal' world model using a Mixture-of-Transformers design to process video, action, and sound as a single stream.
- The 'model factory' pipeline utilizes the world model to generate high-fidelity synthetic data for closed-loop simulation training of robotic hardware.
Amazon Web Services (AWS) and NVIDIA have published a comprehensive infrastructure blueprint detailing how developers can build and run 'Physical AI model factories' using NVIDIA Cosmos 3 on Amazon SageMaker HyperPod, according to official technical posts from the AWS Machine Learning blog. The partnership addresses the massive computational and network bandwidth bottlenecks involved in training foundation models that can understand physics and safely guide physical hardware.
NVIDIA Cosmos 3 is an 'omnimodal' world model designed specifically for physical AI and robotics applications. Built on a Mixture-of-Transformers (MoT) architecture, Cosmos 3 processes video, images, physical actions, and sound as a single, unified token stream, allowing it to predict how physical environments will react to robotic actions. The 'model factory' pipeline uses Cosmos 3 to generate high-fidelity synthetic video and sensory data, which is then used in closed-loop simulations to train robots before they are deployed to physical hardware.
To support the extreme scale of Cosmos 3, the blueprint details deployment on Amazon SageMaker HyperPod, AWS’s resilient, high-bandwidth container orchestration service. HyperPod integrates Amazon EKS and custom health monitoring to automatically detect and bypass GPU failures during long-running training jobs. This architecture allows developers to maintain high 'goodput' across thousands of distributed GPUs, turning complex physical data ingestion, synthetic world modeling, and simulation-based evaluation into a continuous, automated pipeline.
Transitioning from virtual software to physical robots is the ultimate frontier for AI. NVIDIA's Cosmos 3 world model represents a shift from models that simply generate media to models that simulate the laws of physics. By partnering with AWS to run these 'model factories' on SageMaker HyperPod, the two tech giants are providing a scalable, resilient hardware blueprint that lowers the entry barrier for robotics companies to train and validate physical agents at scale.