GPU Cost per Hour

Every GPU turns its purchase price into an hourly cost whether anyone uses it or not. This is the arithmetic behind hyperscaler depreciation debates, cloud GPU pricing, and the build-vs-rent decision — with utilization doing most of the work.

GPUs in the fleet

Capex per GPU · $ incl. server share

Depreciation period · years

Utilization · % of wall-clock hours doing paid work

Power per GPU · kW incl. cooling share

Electricity · $ per kWh

Annual depreciation

$48,000

Annual power

$8,410

Annual total (depreciation + power)

$56,410

Per utilized GPU-hour

$1.34

The machine costs the same whether it works or waits — depreciation and idle power run around the clock. Every point of utilization you lose is carried by the hours that do work: idle hours make busy hours dearer. This is the whole economics of renting versus owning compute.

How to read it: compare your per-utilized-hour number against cloud GPU rates. If your realistic utilization is low, renting wins even at prices that look expensive; if you can keep the fleet busy, owning gets cheap fast. The depreciation period is the honesty dial — the industry argues about 3 versus 6 years because it moves this number by half.

Embed this tool

Paste this snippet into any page to run the calculator there.

<iframe src="https://menfem.com/embed/tools/gpu-cost" width="100%" height="720" style="border:0" title="GPU Cost per Hour — MenFem"></iframe>