AI data center power calculator
How much electricity does an AI cluster draw? Enter the number of accelerators (GPUs). The defaults are published figures for one common setup, each linked below, and you can change every one.
Your cluster
Estimate
Where the defaults come from
- 1.275 kW per accelerator: NVIDIA rates its eight-GPU DGX H100 server at 10.2 kW maximum (DGX H100 user guide). 10.2 ÷ 8 = 1.275 kW, which covers the GPUs (up to 700 W each for an H100 SXM) plus processors, memory, storage and networking in the same box. Network switches and storage outside the server are not included.
- PUE 1.54: the average power usage effectiveness reported in the Uptime Institute Global Data Center Survey 2025. A PUE of 1.54 means that for every 1 kW the servers use, the building draws 1.54 kW in all. Sites differ; type your own.
- 100% load: the rated maximum, which is what sites plan power for. Real draw is lower: one study measured an eight-GPU H100 node peaking at about 8.4 kW while training, about 82% of its 10.2 kW rating (arXiv 2412.08602).
- 10,380 kWh a year per home: the US Energy Information Administration's average of 865 kWh a month per residential customer in 2024, times 12 (EIA).
This is an estimate from rated figures, not a measurement of any real site. Other accelerators and servers draw different amounts; change the server power to match the hardware you have in mind.
Build one without the power bill
Data Center: AI Frontier is an offline strategy game about building the computers that train an AI, in a pixel-art valley that reacts to what you build. Solar follows the day and the weather, wind and batteries cover the night, and industry pollutes the valley the four villages live in. Free on the App Store for iPhone and iPad, with no account, ads, tracking or purchases.