NVIDIA Cosmos3 world models land on Amazon SageMaker JumpStart for physical AI

NVIDIA made its Cosmos 3 family of world models — Cosmos3-Edge, Cosmos3-Nano and Cosmos3-Super — available on Amazon SageMaker JumpStart on August 27. Positioned as open, frontier omnimodal world models for physical AI, Cosmos 3 targets robotics, autonomous vehicles and vision systems that must perceive, reason, plan and act in the real world.

World models differ from language models in that they learn dynamics — how scenes evolve over time — enabling simulation, prediction and planning for embodied systems. Offering three size tiers (Edge for on-device, Nano for constrained compute, Super for maximum capability) lets developers match the model to deployment constraints, from a robot's onboard compute to a data-center training pipeline. Availability on SageMaker JumpStart means AWS customers can deploy them with minimal setup inside existing ML workflows.

The release dovetails with the week's physical-AI theme. Anthropic previewed its Model Hardware Standard for controlling robots and lab tools, and the AWS-NVIDIA 2M-GPU expansion was explicitly framed around 'agentic and physical AI.' NVIDIA is planting world models as the perception/planning layer beneath that build-out, complementing its Isaac robotics stack and Omniverse simulation tools.

Competitively, this pits NVIDIA against Google DeepMind's robotics and world-model research and a growing field of physical-AI startups. The open, omnimodal framing is a bid to make Cosmos the default substrate for robotics developers the way CUDA became default for GPU compute. Skeptical questions center on real-world generalization — world models trained largely in simulation can struggle with the messiness of physical deployment, the same 'lacks physical intuition' caveat Anthropic raised about LLMs. Watch which robotics and AV customers adopt Cosmos 3 on AWS and whether benchmark or demo evidence backs the perceive-reason-plan-act claims.