We Are the Frontier

Liquid-Cooled NVIDIA Racks Show Up at OCP 2026

ASRock Rack's displayed Vera Rubin systems are shown hardware at a conference, with no availability dates given in the source.

Updated on Nov. 2, 2026 in Tech

Liquid-Cooled NVIDIA Racks Show Up at OCP 2026

ASRock Rack showcased liquid-cooled NVIDIA systems at the OCP Global Summit 2026 in San Jose, including a full NVIDIA Vera Rubin NVL72 rack. The systems are presented as shown hardware at a conference booth; the extraction does not establish availability or shipping dates.

Why it matters

The booth signals that rack-scale liquid cooling is becoming the default packaging for NVIDIA's next-generation GPU platforms. Direct-to-chip liquid cooling and fanless server designs point to data centers being built around power and heat density rather than air.

The vendor lists 3,600 PFLOPS of sparse NVFP4 inference and 2,520 PFLOPS of dense NVFP4 training for a full 72-GPU rack, backed by 20.7TB of HBM4 memory. These are manufacturer specifications shown at a booth, not independently measured results.

The players

ASRock Rack

A server and motherboard subsidiary of ASRock that builds rack-scale and GPU server systems for data center customers.

NVIDIA

The dominant maker of AI accelerators, whose GPU, NVLink and MGX platforms underpin most large-scale AI training clusters.

The details

The XGMRACK-VR/NVL72 is a 48RU NVIDIA MGX rack with a 5000A busbar and a liquid manifold, holding 18 Vera Rubin compute trays and nine NVLink 6 switch trays. Each compute tray is a 1U design with two Vera Rubin Superchips, and the rack's power shelves are rated at 110kW each. The smaller 2U16X-GNR2/DLC RUBIN places a Rubin NVL8 board with eight GPUs into a chassis that is fully liquid-cooled with no fans and powered over a 54V DC busbar.

Timeline

  1. ASRock Rack's OCP Global Summit 2026 booth coverage was published on November 2, 2026.

  2. The booth showcase took place at the OCP Global Summit on October 30, 2026.

The Tech Race

ASRock Rack's displays extend NVIDIA's MGX modular rack design to the Vera Rubin generation, with the rack built around a 5000A busbar and liquid manifold. The showcase places a third-party integrator alongside NVIDIA's own platforms at the OCP Global Summit, where hardware makers compete to offer the most complete turnkey AI rack.

Availability and pricing for these systems were not announced in the source, so end users cannot yet order or deploy them. The shift to liquid-cooled, fanless racks primarily affects data center operators who will need to plan power and cooling infrastructure for them.

The takeaway

Watch for pricing and first customer deployments of the Vera Rubin NVL72 rack, which will show whether liquid-cooled rack-scale systems move from conference displays into production data centers. Independent benchmarks of the 3,600 PFLOPS sparse inference figure would also confirm or qualify the vendor's claims.

Further reading

For broader context on the facilities housing this hardware, see Data Centers coverage.

Source note: This article includes information reported by ServeTheHome.