Power and cooling: the hidden cost of AI servers
A modern GPU server can draw more power than a small office. Power and cooling, not the hardware, are often the largest ongoing cost. Here is how to plan for it.
Why GPUs are power-hungry
High-end GPUs can each draw 300–700W under load. A multi-GPU box, plus CPU, drives and fans, can pull several kilowatts continuously. That energy becomes heat that must be removed, which costs more energy again.
Metered vs flat power
Some providers bundle a flat power allowance, which means light users subsidise heavy ones. We meter power and bill actual consumption at €0.45/kWh, so you only pay for what your hardware genuinely uses.
Planning your power budget
- Estimate typical draw, not just peak, since servers rarely run flat out.
- Account for redundancy if you need dual power feeds.
- Remember idle draw, GPUs still consume power between jobs.
We help estimate realistic power draw during planning, so your monthly cost holds no surprises.
In short: Power and cooling are often the biggest ongoing cost of a GPU server; we meter and bill it at actual usage. See colocation pricing