Physical AI at NVIDIA

NVIDIA is building a dedicated physical-AI stack (Cosmos, Curator, Jetson Thor) because training and deploying world models requires fundamentally different hardware than LLMs.

• Two-tower Cosmos 3 architecture: A slow reasoning/planning tower and a fast action tower share a latent space via cross-attention, avoiding encode-decode bandwidth loss between models.
• Cosmos 3 Edge runs on-device: A 4-billion-parameter distilled version runs real-time on NVIDIA's Jetson Thor, a passport-sized module with 2,000 teraops, powering robots like 1X and Boston Dynamics humanoids.

Themes, sentiment and reasons are AI-written paraphrases, checked so they never quote an episode. They may contain errors. Not investment advice. Episodes remain the property of their creators.