Worked planning examples
- H100 electricity cost: GPU vs complete-system power
- GPU rack capacity: power, cooling tons, and space
Found a problem? Send a correction with the hardware reference and assumptions.
Infrastructure planning / v1
AI Data Center Power & Capacity Calculator
Size the infrastructure behind a GPU cluster—or work backwards from a facility’s limits to find its practical AI capacity.
Hardware deployment
Model design load, cooling, rack density, and average energy use.
Use custom hardware values
Custom values are used only when “Custom GPU power” is selected. Complete-system and rack-system presets use their source-backed total input power without GPU multiplication.
Estimated capacity
—
Adjust inputs to see a planning estimate.
Constraint signal—
Planning estimate only. Validate electrical, mechanical, thermal, and code requirements with qualified engineers.
How the calculator works
Forward mode combines hardware power, quantity, auxiliary load, PUE, reserve, rack dimensions, cooling, and operating assumptions. Reverse mode calculates whole-system limits for power, cooling, rack space, and rack density, then selects the smallest limit.
Design vs average load
Design/max load sizes electrical, cooling, and racks. Average load applies your load factor for energy and cost.
PUE and heat
PUE multiplies IT load to estimate facility load. Heat uses BTU/hr = watts × 3.412142, then tons = BTU/hr ÷ 12,000.
Reverse sizing
Power, cooling, rack density, and rack space are checked together; the next violated constraint is reported.
Hardware power reference
| Hardware | Type | Memory | Max power | Source |
|---|---|---|---|---|
| NVIDIA H100 SXM | gpu | 80 GB | 700 W | Reference |
| NVIDIA H100 NVL | gpu | 94 GB | 400 W | Reference |
| NVIDIA H200 SXM | gpu | 141 GB | 700 W | Reference |
| NVIDIA H200 NVL | gpu | 141 GB | 600 W | Reference |
| NVIDIA B200 SXM | gpu | 180 GB | 1,000 W | Reference |
| NVIDIA DGX H100 | server | 8 × 80 GB | 10,200 W | Reference |
| NVIDIA DGX H200 | server | 8 × 141 GB | 10,200 W | Reference |
| NVIDIA DGX B200 | server | 8 × 180 GB | 14,300 W | Reference |
| NVIDIA GB200 NVL72 | rack-system | — | 120,000 W | Reference |
| NVIDIA GB300 NVL72 | rack-system | — | 142,000 W | Reference |
| AMD Instinct MI300X | gpu | 192 GB | 750 W | Reference |
| AMD Instinct MI325X | gpu | 256 GB | 1,000 W | Reference |
| AMD Instinct MI350X | gpu | 288 GB | 1,000 W | Reference |
FAQ
What does PUE mean?
Power Usage Effectiveness is total facility energy divided by IT equipment energy. A PUE of 1.2 means 20% overhead beyond IT load.
Are current estimates code-compliant?
No. Electrical current outputs are planning estimates. Final distribution, protection, and redundancy require engineering review.
Can I share a calculation?
Yes. Use a URL fragment such as #mode=reverse&reverseHardware=h200-sxm&availablePower=500&reversePue=1.2; fragments are not crawlable query-string pages.
Sources and methodology
Hardware references link to manufacturer pages where available. Values are labeled design or maximum inputs and were last verified on 2026-08-20. Missing specifications are intentionally left unspecified.