WORLD · aws-gpu

Cosmos 14B

Cosmos 14B is the world-model path for physical AI generation, policy evaluation, and robot-centric video worlds.

Runtime Target

Use Max-plan cloud GPU planning first; prefer Nebius H200/H100 capacity for the initial high-memory runtime target.

Plan

MAX

Provider

nebius

Accelerator

nvidia-h200-141gb

Simulator

cosmos-physical-ai-generation

Embodiment

robot-centric multiview simulation

Placement

cloud-only

Status

Working Now

Nebius H200 smoke validation has run Cosmos Predict 2.5 14B end to end and saved a generated video artifact. Catalog metadata, paid gating, autoscaling, and Max-plan cloud-session responses are deployed.

Ready To Deploy

The CDK-managed Nebius lifecycle service launches the bounded H200 text2world smoke, persists status, and cleans up expired instances through the janitor.

Runtime Notes

Cosmos is currently a video/world-generation job path. One-H200 latency is too high for a pi0.7-style low-latency robot subgoal loop.

Measured H200 Timing

17 frames

7.724s

17 frames + guidance

13.213s

9 frames

4.008s

1 frame

1.814s

Cache size

43 GiB

GPU

NVIDIA H200