Platform
GPUs
Power available per rack
Data hall, roughly to scale
air-cooled GPU serversliquid-cooled NVL72 racknetwork switches
A GPU cluster is far more than GPUs. Every GPU needs a server slot, a network port, power, cooling and floor space, and the numbers grow fast. Choose a platform and how many GPUs you want.
Assumptions: HGX/DGX H100 server 8 GPUs, about 10.2 kW; HGX B200 server 8 GPUs, about 14.3 kW; GB200 NVL72 rack 72 GPUs, about 120 kW, liquid cooled. One 400 Gb/s compute-network port per GPU, rail-optimised, on 64-port switches in a non-blocking leaf-spine design (2 tiers up to 2,048 GPUs, 3 above). Network, storage and management add about 10% power; facility overhead (PUE) 1.3. A sketch for planning, not a design.