ASRock Rack Shows Vera Rubin NVL72 Rack and 2U HGX Rubin at OCP 2026

By Grace Kim

ASRock Rack brought a full NVIDIA Vera Rubin NVL72 rack to OCP Global Summit 2026 in San Jose, rated at 3,600 PFLOPS of NVFP4 inference (sparse) and 2,520 PFLOPS of NVFP4 training (dense). The XGMRACK-VR/NVL72 sat at the center of a booth packed with Rubin-generation hardware: HGX Rubin NVL8 servers in 2U and 5U, two NVIDIA MGX RTX PRO servers, a dual NVIDIA Vera CPU server, the liquid-cooled Spectrum-X SN6600-LD switch, and the passively cooled RTX PRO 4500 Blackwell Server Edition GPU.

Liquid cooling was the through-line. ASRock Rack applied direct-to-chip liquid cooling (cold plates mounted directly on silicon) across its rack-scale system, its HGX servers, and an 8x RTX PRO 6000 PCIe server.

What does the NVL72 rack actually contain?

The XGMRACK-VR/NVL72 is a 48RU NVIDIA MGX rack with a 5000A busbar and a liquid manifold. Inside are 18 Vera Rubin compute trays and nine NVLink 6 switch trays — 36 NVIDIA Vera CPUs and 72 NVIDIA Rubin GPUs in total.

Memory capacity is 20.7TB of HBM4 across the rack. Each Rubin GPU gets 288GB.

Power arrives through four 3RU 19-inch power shelves. Each shelf is rated at 110kW and uses six 18.3kW PSUs plus bulk capacitors. ASRock Rack offers four cooling-distribution options: in-rack, in-row, RDHx (rear-door heat exchangers), or liquid-to-air sidecar CDUs.

Each compute tray is a 1U MGX design carrying two Vera Rubin Superchips. A Superchip pairs two Rubin GPUs with one Vera CPU over NVLink-C2C (NVIDIA's chip-to-chip interconnect). Networking per tray: eight 800Gbps OSFP ports from ConnectX-9 SuperNICs, plus two 400Gbps QSFP112 ports from a BlueField-4 DPU (data processing unit). Management switches handle lower-speed networking up top.

The trays are fully liquid-cooled. That cuts total rack power draw, and it's a direction NVIDIA has pushed with partners for several generations now.

What does the 2U HGX option change?

Buyers wanting a traditional 8-GPU HGX form factor got one, with a twist: ASRock Rack fit an NVIDIA HGX Rubin NVL8 board, with eight Rubin GPUs, into a 2U chassis — the 2U16X-GNR2/DLC RUBIN, arguably the most impressive system at the booth.

It's fully liquid-cooled with no fans at all. The cold plates cover the E1.S drives, the CPUs, and the GPUs. Networking runs on ConnectX-9 SuperNICs and BlueField-4 DPUs. Power comes in over a 54V DC busbar.

The host side is a dual-socket Intel Xeon 6 platform. It supports the Xeon 6700P, 6500P, and 6700E series in LGA 4710 sockets, with 16+16 DIMM slots at 2DPC (two DIMMs per memory channel). Other series work too, but given the cost of the rest of the system, most buyers will likely spec the 6700P.

Compute output is quoted by NVIDIA at 400 PFLOPS of NVFP4 (sparse) for HGX Rubin NVL8. I/O includes eight 800Gbps OSFP ports via ConnectX-9 and eight hot-swap E1.S NVMe bays.

Why the tray redesign matters

The HGX 8-GPU platform went through major changes this generation, including a change in GPU orientation within the tray. That change is what enables the platform to work in high-density, liquid-cooled racks — the same deployment model as the NVL72.

The booth didn't stop at fully liquid-cooled boxes. Hybrid-cooled servers rounded out the display, and the 5U HGX Rubin NVL8 variant gave buyers a larger-chassis alternative to the 2U system.