AWS shows Physical AI model factory with NVIDIA Cosmos 3

The blog post addresses a practical bottleneck in robotics and autonomous driving development: the need to run a perpetual improvement loop, not a single training job. Traditionally, teams provision separate GPU clusters for each stage—data generation, post-training, and evaluation—each with its own setup and teardown lifecycle. AWS argues this fragments capacity and lowers GPU goodput, the measure of useful pipeline progress per reserved GPU-hour.

Cosmos 3's design is central to AWS's proposal. Because one model can operate as a forward-dynamics world model, an inverse-dynamics action labeler, and a deployable policy, all three stages can run as workloads on one persistent GPU pool. AWS states this eliminates re-provisioning steps and terabyte-scale data migrations between stages. The cluster runs on Amazon SageMaker HyperPod with EKS, with all stages sharing a single Amazon FSx for Lustre file system backed by S3.

The post walks through a complete end-to-end robot-policy stage on the public DROID dataset, with runnable code in the awsome-distributed-ai GitHub repository. AWS emphasizes that capacity should be committed to the whole loop, whether through a flexible training plan for a bounded campaign or a capacity reservation for an open-ended one, to avoid variability in GPU availability and lead times.