NVIDIA Cosmos 3 World Models Land on Amazon SageMaker JumpStart

The three Cosmos 3 variants — Edge, Nano, and Super — span a range of compute footprints, from lightweight on-device inference to large-scale simulation. As omnimodal 'world models,' they are designed to give physical-AI systems — robots, autonomous vehicles, and vision applications — the ability to perceive their environment, reason about it, plan, and act, rather than merely process text or images.

Making them available as open frontier models on SageMaker JumpStart lowers the barrier for AWS customers to build embodied and physical-AI applications without wrangling NVIDIA's stack directly. It fits a week thick with physical-AI moves: Anthropic's MHS hardware standard, Google DeepMind's lab-controlling Co-Scientist, and the reported AWS–NVIDIA partnership involving up to 2 million GPUs.

Competitively, world models are an emerging battleground where NVIDIA aims to own the simulation and perception layer for robotics and autonomy — complementing its chips, Cosmos, and now TensorRT Model Connect tooling. Hosting them on AWS extends reach to enterprises already on SageMaker. The open question is real-world validation: world models are promising but unproven at scale for safety-critical robotics and AV deployments. Watch which robotics and AV customers adopt Cosmos 3 and what benchmarks NVIDIA publishes to back the 'perceive, reason, plan, act' claims.