Calculate continuous megawatt power consumption, monthly and annual operating electricity bills, and required cooling tonnage for NVIDIA HGX H100 8-GPU 640GB SXM5 at US National Commercial Average ($0.12/kWh).
At 10.2 kW each, PUE 1.25 and $0.12/kWh, about $107,222 a year ($8,935 a month), assuming full power all year.
PUE is total facility energy divided by IT energy. Industry surveys have put average reported PUE around 1.5–1.6. Modern liquid-cooled sites are designed for much lower values.
The heat from 81.6 kW of IT load is about 278,000 BTU/h, or 23.2 tons of refrigeration. The calculator's 29.0 tons also includes facility overhead, which is conservative.
NVIDIA HGX H100 is the 8-GPU baseboard used in most H100 servers: eight H100 SXM5 GPUs, each with 80 GB of HBM3 (640 GB in total) and a configurable power limit of up to 700 W, joined by NVLink and NVSwitch. Server makers build complete systems around it. NVIDIA's own DGX H100, built on the same board, lists a maximum system power of about 10.2 kW, which is the per-node default here. The GPUs alone account for up to 5.6 kW; CPUs, memory, network adapters, storage and, above all, fans make up the rest.
Power Usage Effectiveness, defined by The Green Grid and standardised as ISO/IEC 30134-2, is total facility energy divided by the energy delivered to IT equipment. A PUE of 1.25 means that for every 1 kWh the servers use, the building uses another 0.25 kWh on cooling, power conversion losses and lighting. Industry surveys by the Uptime Institute have put the average annual PUE of reporting data centres at roughly 1.5 to 1.6 for several years, while new liquid-cooled facilities are designed for values well below that.
| Quantity | Working | Result |
|---|---|---|
| IT load | 8 × 10,200 W | 81.6 kW |
| Facility load at PUE 1.25 | 81.6 × 1.25 | 102.0 kW |
| Annual energy | 102.0 × 8,760 | 893,520 kWh |
| Annual cost at $0.12/kWh | 893,520 × 0.12 | $107,222 |
| Monthly cost | ÷ 12 | $8,935 |
| Heat and cooling (as calculated) | 102.0 kW × 3,412.142 ÷ 12,000 | 348,038 BTU/h, 29.0 tons |
This is the bill if every node draws its maximum rating all year. Training jobs can come close to that; inference fleets usually do not. At 70% of rated power the same eight nodes would cost about $75,000 a year. Use measured average power from your servers' BMC or rack PDUs when you have it.
| PUE option | Facility load | Annual cost at $0.12 |
|---|---|---|
| 1.08 (immersion) | 88.1 kW | $92,640 |
| 1.15 (direct-to-chip liquid) | 93.8 kW | $98,645 |
| 1.25 (efficient air) | 102.0 kW | $107,222 |
| 1.45 (older chilled-water/CRAC) | 118.3 kW | $124,378 |
Electricity price moves the result more than PUE does. At PUE 1.25 the same cluster costs about $71,500 a year at $0.08/kWh and about $178,700 at $0.20/kWh. The $0.12 default is close to recent averages for US commercial customers in EIA data (roughly 12 to 13 cents per kWh), but large data-centre contracts, regional prices and demand charges vary widely, so use your own tariff.
Nearly all electricity used by IT equipment ends up as heat in the room, so cooling is normally sized from the IT load: 81.6 kW is about 278,000 BTU/h, or 23.2 tons of refrigeration. The calculator applies PUE before converting, which gives 29.0 tons. That is a conservative figure, because much of the facility overhead (chiller compressors, outdoor fans, transformer losses) rejects heat outside the white space. Size computer-room cooling from IT load plus design margin, and let the mechanical engineer size the plant.
Density is the other constraint. 81.6 kW in eight nodes is roughly two to four racks depending on layout, which is far more power per rack than many older colocation halls were built for. Check both the power and the cooling available per rack before ordering hardware.