Huang: Industrials Become Robots, Nvidia Drops AI Stack

🔑 Enhanced Key Takeaways

• •NVIDIA announced the Vera Rubin platform at GTC 2026, featuring Vera CPUs, Rubin GPUs, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch, and Groq 3 LPU for full-stack AI factories[1].
• •Vera Rubin NVL72 system delivers up to 10x higher inference throughput per watt with 72 Rubin GPUs and 36 Vera CPUs, reducing GPU count by one-fourth compared to Blackwell for training large models[1].
• •NVIDIA introduced NVIDIA Cosmos world models, Isaac Lab 3.0, GR00T N models, and previewed GR00T N2 to enhance humanoid robot dexterity and performance in unfamiliar environments[3].
• •Partnerships with ABB Robotics, FANUC, KUKA, Universal Robots, YASKAWA, Agility, Figure, Boston Dynamics, and others integrate NVIDIA's Omniverse, Isaac, and Jetson for real-world robotics deployments[3].
• •Physical AI Data Factory Blueprint automates robotics training data pipelines using NVIDIA Cosmos and integrates with Microsoft Azure and Nebius for scalable data production[2][7].

🛠️ Technical Deep Dive

• •Vera Rubin NVL72: 72 Rubin GPUs + 36 Vera CPUs, 10x inference throughput per watt, 1/4 GPU count vs. Blackwell for training[1].
• •Vera CPU Rack: 256 liquid-cooled Vera CPUs via MGX architecture, supports 22,500+ concurrent CPU environments with ConnectX SuperNICs and BlueField-4 DPUs for networking/security/storage[1].
• •Physical AI Data Factory Blueprint: Integrates NVIDIA Cosmos open world foundation models and coding agents to generate synthetic training data from limited real data for robotics, vision AI agents, and AVs[2][7].
• •Isaac Lab 3.0 and GR00T N models: Enhance robot learning speed, dexterity; GR00T N2 previewed for higher success rates in novel environments[3].
• •Jetson modules: Embedded for real-time AI inference on production line controllers[3].

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline