A GPU is one load inside a system
A component wattage is not the same as the AC input of a server. CPUs, memory, fans, network adapters, storage and power conversion all sit outside a GPU-only number. At the next boundary, the data center also needs non-server IT and facility infrastructure.
Use the most complete input available
If you have a complete system power specification, use it and document what it includes. Do not add GPU watts again. If all you have is a chip-level assumption, label it as such and build the rest of the IT budget explicitly. Nameplate maximum, measured average and expected peak should not be mixed without explanation.
An illustrative scenario
Suppose there are 100 systems at an assumed 10 kW each, plus 100 kW for other IT. That is 1.1 MW of modeled IT load. With a planning PUE of 1.25, the facility scenario becomes 1.375 MW. These are hypothetical inputs, not a specification for any particular server.
Energy adds a time dimension
A power envelope does not tell you the annual bill. You need operating hours and an average electrical load assumption. GPU compute utilization is not automatically the same as electrical load fraction; idle draw and different workloads can change that relationship.
Turn uncertainty into scenarios
Try a low, central and high electrical-load case. Keep the system count and power boundary consistent, and change one assumption at a time. The purpose is to expose sensitivity—not to imply that a rough input becomes accurate because the calculator shows decimal places.
Before handing off the model
- Label every value as measured, specified or assumed.
- Document whether networking and storage are included.
- Use consistent energy boundaries when discussing PUE.
- Separate annual energy planning from peak capacity and transient behavior.
- Ask an engineer to evaluate electrical architecture, cooling and system interactions.