A Physical AI system, such as a robot or autonomous vehicle, translates real-world data into physical actions. It cannot be built in a single training job, so it needs a continuous pipeline of generating synthetic data, post-training, and evaluating in closed-loop simulation.

What makes NVIDIA Cosmos 3 different from other world models?

Cosmos 3 uses a Mixture-of-Transformers design with per-layer joint attention, so one model can both read and generate across video, image, action, and sound. It runs the same transformer trunk in three modes: forward-dynamics for video generation, inverse-dynamics for action labeling, and a deployable action policy.

What does the cost metric GPU goodput mean?

GPU goodput is the useful pipeline progress per reserved GPU-hour across the whole loop. AWS says it is the metric that governs cost, not the peak throughput of any one job, because you pay for reserved capacity whether or not the pipeline is making progress.

RoboticsAutonomous DrivingOpen-Source AIAmazon AI BlogSep 4, 2026Sep 5, 2026, 04:00 JST

AWS shows Physical AI model factory with NVIDIA Cosmos 3

Send on LINE

3 Key Points

What happened
AWS published a guide to building a Physical AI model factory using NVIDIA Cosmos 3 on Amazon SageMaker HyperPod, with code in the awsome-distributed-ai GitHub repository.
Why it matters
Cosmos 3 unifies generation, post-training, and evaluation into one model running in different modes, letting a robot or AV team use a single persistent GPU pool instead of separate capacity per stage. This raises GPU goodput—the useful pipeline progress per reserved GPU-hour—which the post identifies as the real cost driver.
What to watch
The test is whether the single unified model truly eliminates per-stage GPU capacity waste, as AWS claims—gains hinge on teams adopting the MoT design and pipeline walkthrough. Watch the 16B, 64B, and 4B parameter tiers to see which deployment scale delivers the claimed goodput.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

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.

FAQ

What is Physical AI?: A Physical AI system, such as a robot or autonomous vehicle, translates real-world data into physical actions. It cannot be built in a single training job, so it needs a continuous pipeline of generating synthetic data, post-training, and evaluating in closed-loop simulation.
What makes NVIDIA Cosmos 3 different from other world models?: Cosmos 3 uses a Mixture-of-Transformers design with per-layer joint attention, so one model can both read and generate across video, image, action, and sound. It runs the same transformer trunk in three modes: forward-dynamics for video generation, inverse-dynamics for action labeling, and a deployable action policy.
What does the cost metric GPU goodput mean?: GPU goodput is the useful pipeline progress per reserved GPU-hour across the whole loop. AWS says it is the metric that governs cost, not the peak throughput of any one job, because you pay for reserved capacity whether or not the pipeline is making progress.

Amazon AI BlogRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.