GPU Cluster Deployment Estimator_
Enter a GPU count and platform to estimate the racks, power, cooling, and physical deployment time for a cluster—from H100/H200 HGX through GB200 and GB300 NVL72. Useful for early scoping and budget planning.
Planning estimate only. Actual rack density, power, cooling, and schedule depend on the exact SKU, rack configuration, site power/cooling, and network design — confirm against the OEM spec and a site assessment.
How this is calculated_
Rack count comes from the platform's GPU-per-rack density; IT power from the rack count times a representative per-rack draw; facility power adds a rough loss/cooling factor on top. Cooling follows the platform—NVL72-class racks are liquid-cooled by design. Timeline is a range keyed to scale. These are planning estimates; the real figures depend on the exact SKU, rack configuration, site conditions, and network design, and should be confirmed against the OEM spec and a site assessment.
Common questions_
How many racks do I need for a GPU cluster?
It depends on the platform's rack density. Air-cooled H100/H200 HGX systems fit roughly 32 GPUs per rack; liquid-cooled configurations reach higher. GB200 and GB300 NVL72 place 72 GPUs in a single liquid-cooled rack. So an 8,192-GPU cluster is roughly 114 NVL72 racks or more air-cooled racks — the estimator computes this from your GPU count and platform.
How much power does a GPU cluster draw?
On the order of the rack count times the per-rack draw. An NVL72 rack is in the ~120 kW range (GB300 higher), while an air-cooled HGX rack is far lower. Facility power runs above the IT load once cooling and losses are included. Confirm exact figures against the OEM power summary before any electrical work.
When do I need liquid cooling instead of air?
GB200 and GB300 NVL72 racks are 100% liquid-cooled — there is no air-cooled fallback for the GPUs. Air or rear-door heat exchangers can serve lower-density H100/H200 HGX racks, but as rack density climbs, direct-to-chip liquid becomes required.
How long does it take to physically deploy a GPU cluster?
The physical build — rack landing, assembly, cabling, cooling, and commissioning to an accepted cluster — ranges from a few weeks for a small deployment to several months (often phased) for tens of thousands of GPUs. The estimator gives a range from the rack count; the real schedule depends on site readiness and crew size.