NVIDIA Cosmos 3 on AWS: Omnimodal World Models for Physical AI

This article introduces NVIDIA Cosmos 3, a unified omni-model that combines vision reasoning, world generation, and action prediction for physical AI applications, with reference architectures for deployment on AWS.

• Cosmos 3 unifies three separate models (Predict, Transfer, Reason) into a single Mixture-of-Transformers architecture with one inference call
• Offers three variants: Super (64B) for research, Nano (16B) for production, and Edge (4B) for on-device robot inference
• Dual-tower design combines autoregressive reasoning tower with diffusion-based generation tower for coherent video and action output
• AWS reference architecture spans data curation on S3/Batch, post-training on SageMaker HyperPod, and inference via EKS or SageMaker JumpStart
• Two production deployment patterns: Amazon EKS with Cosmos NIM for real-time control loops, or SageMaker JumpStart for managed serverless inference
• Edge deployment uses Cosmos 3 Edge (4B) on NVIDIA Jetson via AWS IoT Greengrass with OTA model updates
• Recommended instances range from g7e.2xlarge for single-GPU inference to p6-b200.48xlarge for Blackwell-based training

Cosmos 3 simplifies physical AI pipelines by eliminating multi-model orchestration, reducing latency, and enabling end-to-end deployment from data ingestion through edge inference on AWS.