High density starts where a standard cabinet stops. Most retail racks are built for a few kilowatts; GPU and HPC racks need 20, 40 or more. The questions that matter change with it: how much power the facility will commit per rack, how it removes the heat, and when the space is ready.
Start from the circuit
A circuit's usable capacity is its voltage times its amperage times 0.8 for continuous load, times 1.732 if it's three-phase. That's the number to compare with your servers' peak draw, not the breaker size on the quote.
| Circuit | Usable | DGX H100s per circuit |
|---|---|---|
| 208V single-phase, 30A | 5.0 kW | 0 |
| 208V three-phase, 30A | 8.6 kW | 0 |
| 208V three-phase, 60A | 17.3 kW | 1 |
| 415V three-phase, 32A | 18.4 kW | 1 |
| 415V three-phase, 60A | 34.5 kW | 3 |
| 415V three-phase, 100A | 57.5 kW | 5 |
Server draw: NVIDIA, DGX H100/H200 user guide: introduction and specifications, 10.2 kW max per 8U system.
With 2N power, every rack gets two feeds and each must carry the full load alone, so you pay for twice the circuit capacity you draw. With N+1, ask how the facility shares the redundant path and what happens to your rack in a single failure.
Density by server count
| Servers | Peak draw | Space | What it needs |
|---|---|---|---|
| 1 per rack | 10.2 kW | 8U | High density, air with containment or rear-door cooling |
| 2 per rack | 20.4 kW | 16U | High density, air with containment or rear-door cooling |
| 3 per rack | 30.6 kW | 24U | High density, air with containment or rear-door cooling |
| 4 per rack | 40.8 kW | 32U | Liquid-assisted cooling, confirm in writing |
| 5 per rack | 51.0 kW | 40U | Liquid-assisted cooling, confirm in writing |
How power is billed at high density
At 30 kW and up, the power line is most of the bill. A breakered contract bills the full circuit whether you draw it or not, so at high density it pays to be precise about committed kW. A metered contract bills what you draw at the facility's billed PUE, so ask for that PUE in writing. An all-in rate hides both: compare it with the metered equivalent before you accept it.
| Metro | Industrial rate | 40 kW rack per month | 36 months |
|---|---|---|---|
| Houston | 7.07 cents/kWh | $2,023 | $72,833 |
| Dallas / Fort Worth | 7.07 cents/kWh | $2,023 | $72,833 |
| Austin | 7.07 cents/kWh | $2,023 | $72,833 |
| Seattle | 7.51 cents/kWh | $2,149 | $77,366 |
| Charlotte / Raleigh | 8.37 cents/kWh | $2,395 | $86,226 |
| Hillsboro / Portland | 8.92 cents/kWh | $2,553 | $91,892 |
| Salt Lake City | 9.10 cents/kWh | $2,604 | $93,746 |
| Miami | 9.17 cents/kWh | $2,624 | $94,467 |
EIA Electric Power Monthly, Table 5.6.A, July 2026; 70% draw, PUE 1.4.
What to get in writing
- Committed kW per rack, and the circuits that deliver it
- Cooling method: air with containment, rear-door heat exchangers or direct-to-chip liquid
- Floor loading for heavy racks
- The date power and cooling are ready, not just the space
- The price per kW when you add the next rack
Questions
Anything well above a standard cabinet's few kilowatts. In practice, buyers use the term from about 15 to 20 kW per rack, and GPU racks commonly run 30 to 100+ kW.
Usually less per kW on space, because more power fits in the same floor area, but the power and cooling lines grow. Compare on a 36-month landed cost, not on rent per cabinet.