Physical AI
Physical AI refers to AI systems operating in or modelling the physical world, as distinct from language or text-only AI. jensen-huang uses this term to emphasise that AI models for biology, chemistry, thermodynamics, and robotics require training on different modalities and data types than language models, and that nvidia specifically builds and open-sources these domain models.
Examples cited (Lex Fridman Podcast #494, 2026):
- Satellite imaging AI — NVIDIA GPUs in orbit process centimetre-scale Earth imagery in real time, discarding unchanged data before downlinking. 24/7 solar power available at polar orbits; cooling via radiation only.
- Biology/drug discovery — NVIDIA’s goal: give Eli Lilly world-class biology AI for drug discovery without NVIDIA itself discovering drugs.
- Humanoid robots — Huang projects humanoid robots using existing tools (reading manuals on first encounter) as the near-term physical AI embodiment.
- Weather modelling, fluid dynamics — domain-specific AI models requiring physics-informed architectures.
Physical AI is also the long-term vision Huang describes for interplanetary travel: a humanoid robot sent ahead, updated via AI upload at the speed of light to “catch up” with the physical vessel.
See open-source-ai for NVIDIA’s strategy of open-sourcing physical AI models across industries.