AI Training Energy Calculator
The electricity a GPU fleet draws over a training run or a month of serving — with the data-centre overhead, the power bill and the carbon it implies at your grid's intensity — and the household-years it equals.
Energy is GPUs times their draw times hours, scaled up by the data centre's PUE for cooling and power losses.
How the ai training energy calculator works
Energy is GPUs times their draw times hours, scaled up by the data centre's PUE for cooling and power losses. A thousand accelerators at 700 W for a month is half a gigawatt-hour before overhead, which is a mid-six-figure power bill and hundreds of tonnes of CO₂ on an average grid — or a tenth of that on a clean one.
The grid intensity is an input because it is the dominant uncertainty: the same run emits ten times more in one region than another.
Formula: kWh = GPUs × W × hours × PUE / 1000; cost = kWh × tariff; CO₂e = kWh × grid intensity
Worked examples
| Inputs | Energy used (MWh) | Note |
|---|---|---|
| A thousand GPUs for a month | 805.11 | hundreds of tonnes |
| A clean grid | 805.11 | an eighth of the emissions |
| A single 8-GPU node for a week | 1.468 | a household-year |
FAQFrequently asked questions
How much energy does training a model use?
Accelerators times their draw times hours, plus the hosts around them and the data centre's overhead. A thousand accelerators for a month is roughly 700 MWh at the meter.
What is PUE?
Power usage effectiveness — total facility power over IT power. Cooling and distribution losses push it above 1; good hyperscale sites run 1.1 to 1.2, older facilities 1.5 or more.
What grid intensity should I use?
Your region's average, from the grid operator or a public dataset: roughly 50 g/kWh for hydro- and nuclear-heavy grids, 400 for a typical mix, 700 or more for coal-heavy ones.
Why include host overhead?
CPUs, memory, network switches and storage draw real power alongside the accelerators — typically 20 to 40% on top. Leaving it out understates the bill.
Does this cover embodied emissions?
No — manufacturing the hardware carries its own footprint, amortised over its life. This page is operational energy only.
Where these figures come from
- Vaswani et al. (2017) — Attention Is All You Need — the transformer architecture the memory arithmetic follows
- Kaplan et al. (2020) — Scaling Laws for Neural Language Models — the compute relationship used for training estimates
- Hoffmann et al. (2022) — Training Compute-Optimal Large Language Models — the tokens-per-parameter guidance ("Chinchilla")
- IEEE 754 — Standard for Floating-Point Arithmetic — the numeric formats behind bytes per parameter
- National AI Centre — Australia's national AI body
Last checked: September 2026. The relationships here are architectural, not vendor-specific: bytes per parameter follow the numeric format, KV-cache size follows the transformer definition, and token-per-word ratios come from published tokeniser behaviour.