Limitless: An AI Podcast NVIDIA GTC: Jensen Huang's 5 Biggest Announcements
30 snips
Mar 17, 2026 A fast recap of Jensen Huang's bold trillion-dollar orders target and market reaction. Debate over the controversial DLSS 5 and its uses beyond gaming. Deep dive into the Vera Rubin platform and its claimed 35x performance leap. Discussion of new inference chips, strategic acquisitions like Grok, robotics demos, and ambitions for self-driving and space-deployed AI.
AI Snips
Chapters
Books
Transcript
Episode notes
Vera Rubin Packs Multiple Chip Types For 35x Gain
- Vera Rubin promises ~35x aggregate performance uplift by combining seven new chips into five new racks to form an AI supercomputer.
- Ajaz attributes the 10x perf-per-watt gain over Blackwell partly to new LPUs using SRAM from the Grok integration.
Performance Per Watt Is The Multiplier For AI Scale
- Rubin improves compute energy efficiency roughly 10x vs Blackwell, meaning each gigawatt of power yields far more model training/inference capacity.
- Josh predicts data-center gigawatt scale will rapidly multiply AI "intelligence" delivered to the market.
NVIDIA Building Specialized Inference Hardware
- NVIDIA is diversifying beyond general GPUs into hyper-specific inference LPUs to capture the surging market for prompting/querying models.
- Ajaz links this shift to the Grok acquisition enabling SRAM-based inference chips inside Rubin.
