Particle.news
Download on the App Store

Nvidia Opens Vera Rubin AI Rack Design, Details Kyber Roadmap for 576‑GPU Systems

The OCP step signals an 800 VDC, liquid‑cooled path for vendors building denser, more efficient AI data centers.

Overview

  • Nvidia published specifications for the Vera Rubin NVL144 MGX rack and committed the design to the Open Compute Project as an open standard.
  • The NVL144 compute tray uses 100% liquid cooling and a printed‑circuit‑board midplane that replaces cables, with bays for ConnectX‑9 800GB/s networking and Rubin CPX for massive‑context inference.
  • Nvidia outlined the Kyber rack generation targeting 2027 systems that house up to 576 Rubin Ultra GPUs, using vertical compute blades and integrated NVLink switch blades to increase density and simplify scale‑up networking.
  • Industry participants are designing around 800‑volt direct current power, with Foxconn detailing a 40 MW Kaohsiung‑1 project and Vertiv presenting an 800 VDC MGX reference architecture.
  • Nvidia said its NVLink Fusion ecosystem now includes Intel and Samsung Foundry to enable semi‑custom silicon integration into Nvidia‑optimized data center platforms.