AWS details how to run a Physical AI model factory with NVIDIA Cosmos 3 on SageMaker HyperPod
An AWS engineering post shows how Cosmos 3 runs synthetic data generation, post-training and closed-loop evaluation as three workloads on one persistent GPU pool.
AWS published a guide on building a "Physical AI model factory" by running NVIDIA Cosmos 3 on Amazon SageMaker HyperPod. The premise: a robot or autonomous vehicle that turns real-world data into physical actions cannot be built in a single training job, it needs a continuous loop.
That loop has four stages. Real Physical AI data such as DROID, BridgeData2 and AV sensor logs is ingested and curated into a shared corpus on Amazon S3 and Amazon FSx for Lustre. A Cosmos3-Super teacher then generates synthetic data to augment it. The combined synthetic and real corpus post-trains a deployable Cosmos3-Nano policy, with vision fine-tuning applied across both tiers. The policy is finally evaluated in closed-loop simulation.
One trunk, three modes
Cosmos 3 is an open omnimodal world foundation model that treats video, image, action and sound as a single token stream. The same transformer trunk runs as a forward-dynamics world model for synthetic video generation, an inverse-dynamics action labeler and a deployable action policy. What changes between modes is only which tokens start as noise. Because one model family covers generation, post-training and evaluation, AWS schedules the stages as time-shared workloads on a single persistent, resilient GPU node pool instead of standing up separate capacity per stage.
The family has two tiers: Cosmos3-Nano at 16B parameters on a dense 8B Qwen3-VL backbone, and Cosmos3-Super at 64B on a dense 32B Qwen3-VL backbone. NVIDIA also released Cosmos3-Edge, a 4B tier for on-device deployment benchmarked on Jetson Thor and Orin. Edge is built on a dense ~2B backbone trained from scratch rather than initialized from Qwen3-VL, making it a separate weight lineage. Task variants such as Cosmos3-Nano-Policy-DROID are specialized to policy mode at 15 Hz with a 32-step horizon.
NVIDIA released the model under the Linux Foundation's OpenMDW-1.1 license. AWS argues the cost metric that matters is not peak throughput of any single job but GPU goodput: useful pipeline progress per reserved GPU-hour across the whole loop. Infrastructure templates and job manifests sit in the awsome-distributed-ai GitHub repository. Full write-up in the AWS blog post.
