NVIDIA Unveils Vera Rubin and Next-Generation Feynman at GTC 2026, Expands Physical AI Platform

In live updates dated March 19, 2026, NVIDIA reported on announcements made during its March 16 keynote at GTC 2026. The centerpiece, Vera Rubin, is a computing platform for agentic AI comprising seven chips, five rack-scale systems and one supercomputer. NVIDIA also unveiled its next-generation Feynman architecture and highlighted collaborations with automakers and robotics companies in physical AI.

NVIDIA’s announcements showed an approach that treats AI computing infrastructure as more than individual chips, encompassing rack-scale systems and pre-deployment simulation. These offerings could provide options for organizations considering AI factory deployments, but reference designs and demos alone do not establish real-world costs, uptime or performance.

The scale NVIDIA emphasized at GTC was less about the performance of individual chips and more about bringing Vera Rubin together as a computing platform spanning seven chips, five rack-scale systems and a supercomputer. Feynman was positioned as the next generation, with a configuration that includes CPUs, LPUs, networking and storage. The new element, then, was the scope NVIDIA laid out for designing multiple components as a single AI computing system. At the same time, its strategy of co-designing software and silicon around CUDA and accelerated computing extends an approach already discussed in the keynote. The AI factory reference design and DSX Air bring design and simulation before deployment into NVIDIA’s product portfolio. For physical AI, NVIDIA outlined a path from simulation in Omniverse to deployment of robots and autonomous vehicles in the real world. But the collaborations and demos highlighted in the live updates do not show the scale or operational track record of any specific deployment. At GTC, NVIDIA CEO Jensen Huang said computing demand had increased 1 million times over several years and forecast at least $1 trillion in revenue from 2025 to 2027. These should be understood as his claims and projections. The supply of computing resources is a separate issue from their availability for actual use. Reuters reported that about 600,000 NVIDIA H200 units were stranded in Hong Kong awaiting Chinese import approval. NVIDIA’s announcements outline rack and system designs but do not explain how much of the announced capacity will be supplied or brought into operation. The need to distinguish product introductions from performance verification in real-world settings also applies to Humanoid.Guide’s coverage of Geek+’s wheeled humanoid, Gino 1. The industry implication is that NVIDIA has outlined a vision connecting computing infrastructure, simulation for factory design, and physical AI models and tools. But this live update alone does not establish how extensively outside companies can use the offerings, how real-world robot data will be collected and fed back into retraining, or the overall system’s cost, power consumption and safety in operation. The next points to watch are supply timelines and actual deployment scale for each generation, the terms of access to DSX Air, and the scope of implementation at each partner.

Vera Rubin includes the new Vera CPU and the BlueField-4 STX storage architecture. NVIDIA described it as a plan to vertically integrate and optimize the entire system, including software. For the Feynman generation, the company showed a configuration that connects the NVIDIA Rosa CPU, LP40 LPU, BlueField-5 and CX10 using NVIDIA Kyber. For AI factories, NVIDIA unveiled the Vera Rubin DSX AI Factory reference design and Omniverse DSX Blueprint. It said DSX Air can simulate an AI factory in software before construction. For physical AI, NVIDIA named BYD, Hyundai, Nissan and Geely as new automaker partners, and also highlighted collaborations with Uber, ABB, Universal Robots and KUKA. In the keynote, NVIDIA said the Olaf demo used its physical AI stack, the Newton physics engine and simulation with Omniverse.