An AI rack power budget should answer two different questions: what electrical load belongs to this rack, and what wider infrastructure must support it? Start with a list of complete systems and additional IT equipment. Keep facility overhead and redundant supply capacity in separate calculations. The worked example below produces 43 kW for one rack and 440 kW for a ten-rack cluster. All example inputs are hypothetical, not a hardware recommendation.
1. Make a load schedule before using a calculator
For each item, record its quantity, electrical input power, evidence type and location. A manufacturer maximum, a measured operating peak and a measured average answer different questions. Do not silently use an average for one server and a maximum for another. If the purpose is an early capacity discussion, state the chosen planning condition in the schedule header.
Complete-system documentation is a useful starting point. For example, NVIDIA’s DGX H100 deployment guide lists a maximum system power value separately from networking, storage and management infrastructure. Use the documentation for your exact platform and configuration; the hypothetical 10 kW input below is not presented as that product’s specification.
| Item in one example rack | Quantity | Input per item | Subtotal |
|---|---|---|---|
| Complete AI server | 4 | 10 kW | 40 kW |
| Rack network switch | 2 | 1 kW | 2 kW |
| Other rack IT equipment | 1 group | 1 kW | 1 kW |
| Total modeled rack IT input | — | — | 43 kW |
The arithmetic is 4 × 10 + 2 × 1 + 1 = 43 kW. Because the server entry already represents the complete system’s AC input, do not add its internal GPU, CPU, fan or power-conversion loads again. Keep evidence beside the inputs so that a future server replacement changes one traceable row.
2. Put shared equipment in exactly one place
Now suppose ten identical racks share an additional storage and management installation modeled at 10 kW. The cluster IT load is 10 × 43 + 10 = 440 kW. That last 10 kW appears once, not once per rack. A spreadsheet that adds a shared installation to every rack would overstate this example by 90 kW.
You may allocate shared load across racks for reporting: 10 kW divided by ten racks adds 1 kW per rack, giving 44 kW per allocated rack. That accounting value is not the physical draw of each rack’s feeds. Keep the location-based 43 kW value available for electrical review. Shared storage may sit on an entirely different distribution branch.
Use the AI data center power calculator in complete-system mode with 40 systems, 10,000 W per system and 40 kW additional IT. The additional field contains 30 kW of rack switches and other rack IT, plus 10 kW of shared equipment. The expected IT result is 440 kW.
3. Separate rack load from facility overhead
At an assumed planning PUE of 1.25, this cluster produces a facility power scenario of 440 × 1.25 = 550 kW. The difference is 110 kW of modeled non-IT overhead across the chosen facility boundary. It is not another 110 kW to add to the rack power strips.
PUE is an energy ratio in measured reporting. The US Department of Energy’s explanation uses annual facility energy divided by annual IT energy. Applying a selected PUE to a power envelope here is a planning simplification, not proof of coincident maximum cooling and electrical demand. Review the actual cooling and distribution systems separately.
4. Do not add redundant feed ratings as though they were consumption
A rack receiving more than one supply path does not necessarily consume the sum of those paths’ ratings. Conversely, spare nameplate capacity does not prove that the remaining connections can support every server after a failure. The question is which loads remain connected to which sources, and at what input demand, during the defined event.
NVIDIA’s electrical design documentation illustrates why this is platform-specific: its DGX H100 example includes particular requirements for energized power supplies and upstream paths. Do not transplant a generic two-feed diagram to another server without checking that server’s requirements. This article does not select connectors, circuits, breakers or conductors.
5. Carry IT load into the capacity planner
The value to transfer to the data center capacity planner is 0.440 MW of IT load, together with the separate PUE assumption of 1.25. Choose Use IT load in capacity planner, review those values, then select Apply IT load & PUE. Other planning settings remain unchanged. Do not enter the already expanded 0.550 MW as IT load: multiplying it by PUE again would produce 0.6875 MW and count overhead twice.
Power factor, reserve and block size belong to the next planning stage. Record them explicitly rather than hiding them in the server wattage. For annual energy, also supply an electrical average-load fraction or a time-resolved profile; peak rack kW alone does not establish an annual electricity bill.
A useful handoff checklist
- Identify the server configuration, count and power evidence.
- Locate rack IT and shared IT separately, with no duplicated rows.
- Label results as measured, specified or hypothetical.
- Keep IT kW, facility kW and installed supply capacity distinct.
- Review cooling, space, weight and normal/failure-state distribution with the appropriate specialists.
The output is a transparent planning worksheet, not an engineered rack deployment. Revisit it when equipment, workload assumptions or the required failure scenario changes.
Sources & further reading
- NVIDIA: Planning a Data Center Deployment — DGX H100 ↗
- NVIDIA: Electrical Specifications — DGX H100 ↗
- US DOE: Cooling Water Efficiency Opportunities for Federal Data Centers ↗