AI servers by OEM and ODM

Most models here hold 8: 49 of the 68 list eight GPUs, on NVIDIA HGX, AMD Instinct or Intel Gaudi 3 baseboards. Rack-scale systems put 72 GPUs in one NVLink domain: GB200 and GB300 NVL72 racks from NVIDIA, Supermicro, HPE, Lenovo, GIGABYTE, ASUS and Ingrasys all state 72. Four-GPU servers such as the Pegatron AS205-2T1 sit below that.

A GB200 or GB300 NVL72 rack runs 120 to 140 kW: NVIDIA rates its DGX GB200 NVL72 at about 120 kW, Supermicro states 132 kW, Inventec 136 kW, and HPE and GIGABYTE up to 140 kW. One 8-GPU node is a tenth of that: 14.3 kW for a DGX B200, 15 kW for a DGX B300 or Supermicro’s SYS-822GS-NB3RT. Makers that print no kW figure keep a blank cell.

26 models are liquid cooled throughout, including the NVL72 racks from NVIDIA, Supermicro, HPE, Lenovo, GIGABYTE, ASUS, Inventec and Wiwynn, Dell’s XE9780L and XE9785L, Supermicro’s LCC and ALC nodes, Lenovo’s SR780a V3 and SC777 V4 Neptune, and QCT’s 3U D75L. Five more, among them HPE’s XD685 and Dell’s XE9712, ship in air or liquid versions. The liquid cooling list links to the CDU and cold plate vendors behind them.

OEMs such as Dell, HPE, Lenovo and Cisco sell branded servers with their own support. ODMs such as Wiwynn, Wistron, Inventec, Ingrasys and QCT design and build servers, often rack-scale ones, for hyperscalers and for other brands. The table lists both on one sheet so a Dell XE9780 and an Inventec P9000IG6 read side by side.

From the maker’s own product page, datasheet or user guide, with one source link on each figure and the date we checked it. A field stays empty when the maker does not print the figure. The methodology page lists every field definition.