IT load or facility power? The question that breaks GPU budgets

Operators often quote capacity as total facility power, while buyers size their clusters by IT load. The difference is roughly 30% of the number, and it decides whether a deployment fits the hall at all.

Where the gap comes from

IT load is what the servers, switches and storage draw. Facility power adds the cooling plant, power conversion losses, lighting and the rest of the building. In a dense air-cooled deployment, that overhead is large, and it grows as the hall gets hotter.

A worked example

Take 256 air-cooled 8-GPU HGX B200 nodes, 2,048 GPUs in all, at roughly 14 kW per node. That is about 3.6 MW of IT load. Add cooling and electrical overhead, and the site needs around 5.2 MW. A hall marketed as “4 MW available” may or may not take this cluster, depending on which of the two numbers the operator meant.

The fix

State which figure you mean, every time, and require operators to do the same in writing before any price is discussed. In every RFI we send, the power basis is a mandatory field.