Particle.news
Download on the App Store

Nvidia Debuts Vera Rubin, a Rack-Scale AI Platform, at CES 2026

Cloud providers plan rollouts later this year pending real-world verification of the performance claims.

Overview

  • CEO Jensen Huang unveiled the platform at CES weeks earlier than Nvidia’s usual GTC timeline, a move positioned to outpace AMD’s Helios and Intel’s Gaudi 3 efforts.
  • Rubin is a full-stack, rack-scale system that integrates CPUs, GPUs, networking and software, featuring the NVL72 design linking 72 GPUs and 36 CPUs and HBM4 memory with over 20 TB/s bandwidth.
  • Nvidia says Rubin delivers up to 5x faster inference and 3.5x faster training than Blackwell, with up to 10x lower inference cost and up to 75% fewer GPUs needed for training.
  • Availability is targeted for the second half of 2026, and Microsoft, Google Cloud, AWS, Oracle and CoreWeave said they plan to offer Rubin systems as they come to market.
  • The company tied Rubin to "physical AI" use cases and announced Alpamayo for autonomous driving, with Mercedes‑Benz planning to integrate the technology into future vehicles.