NVIDIA Unveiled Vera Rubin: Chip with 50 Petaflops Performance

At the GTC 2026 conference, Jensen Huang presented Vera Rubin — a GPU with inference performance of 50 PFLOPS, which is 5 times greater than the capabilities of the previous Blackwell generation. The chip is already in full production as of the first quarter of 2026.

NVIDIA officially launched the Vera Rubin platform with record-breaking specifications for artificial intelligence inference. The company also presented the Feynman architecture on TSMC's 1.6nm process and the NemoClaw platform for enterprise AI agents. This announcement comes at a time when investors were questioning the pace of AI industry development.

Vera Rubin: New Standard for Inference

Each Vera Rubin NVL72 GPU delivers 50 PFLOPS of inference performance in NVFP4 format — a fivefold increase compared to Blackwell. At the rack level, the NVL72 configuration delivers 3.6 EFLOPS of computing power. The chip is built from two reticle-sized dies, contains 336 billion transistors, and uses HBM4 memory with bandwidth up to 22 TB/s. The Rubin Ultra rack with 600 kW power will ship in 2027. Cloud partners, including AWS, Google Cloud, Microsoft Azure, and Oracle Cloud, will begin receiving equipment in the second half of 2026.

NemoClaw: Platform for Enterprise AI Agents

NVIDIA introduced NemoClaw — an open platform for deploying AI agents in enterprise environments. Unlike OpenClaw, which is oriented toward individual users, NemoClaw is specifically designed for companies. The platform integrates three existing NVIDIA components: the NeMo framework for model training and agent reasoning pipelines, the Nemotron model family (released in December 2025), and NIM microservices for inference deployment. This announcement positions NVIDIA as a player in the growing market for enterprise AI software.

Physical AI and Robotics

Jensen Huang devoted significant attention to physical AI — the convergence of artificial intelligence with robotics and the physical world. NVIDIA demonstrated the evolution of its robotics platform since GTC 2025, when the company first introduced the Isaac GR00T N1 humanoid robot foundation model, the Blue robot (developed jointly with Disney Research and Google DeepMind), and the Newton physics engine for training robot movements at scale. At GTC 2026, new deployments and platform updates were showcased, demonstrating robots performing increasingly complex manipulation tasks in real industrial environments.

Что это значит для Казахстана

For companies in Kazakhstan and Central Asia working with cloud computing and AI, these NVIDIA announcements are critically important. Cloud service providers using AWS, Google Cloud, and Microsoft Azure infrastructure will gain access to Vera Rubin in the second half of 2026. This will enable local enterprises and startups in the region to leverage advanced AI inference capabilities without requiring their own equipment investments. Companies like Alashed IT (it.alashed.kz) can integrate NemoClaw into solutions for corporate clients, providing access to enterprise-grade AI agent tools.

Vera Rubin delivers 50 PFLOPS of inference — 5 times more than Blackwell — and is already in full production as of Q1 2026.

NVIDIA demonstrated that the AI industry is far from slowing down by presenting a chip already in production with record inference specifications and an enterprise platform for deploying AI agents. The success of these announcements will depend on whether real production workloads can match the stated specifications and whether corporate clients will adopt NemoClaw at expected scales.