What matters in AI.

Subscribe

ASRock Rack shows an NVIDIA rack with 72 Rubin GPUs and liquid cooling

The rack has 20.7TB of HBM4 memory and a liquid cooling system for all of its trays.

At the Open Compute Project Global Summit 2026 in San Jose, ASRock Rack showed its version of the NVIDIA Vera Rubin NVL72 rack. The rack has 72 Rubin GPUs, which are AI processors, and 36 Vera CPUs, which are general processors. NVIDIA gives 3,600 PFLOPS of NVFP4 inference. Liquid cooling covers all trays. No price was given.

How it works

The rack has processor trays and switch trays in one frame. 4 power shelves supply it, and each shelf gives 110kW. ASRock Rack also showed a 2U server with 8 Rubin GPUs. It has no fans. Liquid cooling covers its CPUs, GPUs and drives.

The numbers

  • NVIDIA gives 3,600 PFLOPS for inference and 2,520 PFLOPS for training.
  • The rack has 20.7TB of HBM4 memory.
  • The 2U server has 400 PFLOPS of NVFP4 inference.

What is not known

No price or release date was given. The numbers are the numbers that NVIDIA gives. They are not test results from the booth. It is not known which of the 4 cooling units customers select.

Insight: NVIDIA sets the rack that others offer

ASRock Rack shows its version of the NVIDIA Vera Rubin NVL72 rack. Google Cloud says it opens servers on the same rack next month. AMD says it gives its Helion rack specifications to the Open Compute Project. Many firms supply the same NVIDIA rack, while AMD supplies an open rack that is different. A customer selects who supplies the rack.

Sources

  1. Liquid-Cooled NVIDIA Servers at the ASRock Rack OCP 2026 Boothservethehome.com
AI MATTER · NEWS · AI MATTER · NEWS ·2 NOV2026

Posted

Tags

More in Compute

All Compute news