GPU platforms compared

Platform	Power per GPU	Power per system	Power per rack	Cooling and weight
H100 (HGX/DGX)	700 W	10.2 kW per 8-GPU system	About 41 kW with four systems in an air-cooled rack (NVIDIA). Up to about 70 kW with rear-door cooling (Dell).	Air or rear-door. About 288 lb per system.
H200 (HGX)	700 W	10.2 kW	Same as H100	Same chassis as H100
B200 (HGX/DGX)	Up to 1,000 W	14.3 kW	About 55 kW air-cooled (GIGABYTE). About 103 to 107 kW with direct liquid cooling (GIGABYTE, Dell).	Air or direct-to-chip
GB200 NVL72	About 1,200 W (estimate)	One 72-GPU rack	About 120 to 132 kW nominal. HPE lists an estimated peak of about 192 kW.	Direct-to-chip hybrid: about 115 kW to liquid and 17 kW to air (QCT). About 3,245 lb.
GB300 NVL72	About 1,400 W (estimate)	One 72-GPU rack	132 kW nominal and about 155 kW peak. HPE recommends 192 kW of busway.	Direct-to-chip. About 3,300 lb. An in-row 1.3 MW CDU serves up to eight racks (HPE).

What this means for the site

Plan power around the peak figure, not the nominal one. For GB300 NVL72, that means busway sized for 192 kW per rack.
H100, H200 and B200 can run on air at lower density. B200 at full density and every NVL72 rack need direct-to-chip liquid cooling and coolant distribution units (CDUs).
NVL72 racks weigh over 3,200 lb. No vendor publishes a floor rating, so get the site's floor load rating in writing.
At 132 kW per rack, 1 MW of IT load is about seven NVL72 racks, before networking, storage and cooling.

Sources: HPE GB300 NVL72 QuickSpecs; HPE GB200 NVL72 QuickSpecs; NVIDIA DGX documentation; Dell; GIGABYTE; QCT; Vertiv; SemiAnalysis (per-GPU estimates for NVL72). Figures vary by vendor and configuration, so confirm with your hardware vendor. GB300 guide → · Requirement template →