AI Workload Energy and Carbon Calculator
Estimate the electricity, emissions and wholesale power cost of an AI training or inference job on live data from 15 European grids, and find the cleanest hour to start it.
How much electricity does a GPU job use, and how much CO₂ does it emit? It depends on the hardware, the facility and, above all, the grid it runs on and the hour it starts. This calculator prices one job against live carbon intensity and day-ahead prices in fifteen European markets, using vendor board-power figures and assumptions you can change.
AI workload energy and carbon calculator
Method, defaults and sources
The calculator holds power constant for the whole job and prices it against the hourly carbon intensity and day-ahead price of the chosen market, taken from the live grid.
IT power = accelerators × board power × server overhead × average loadFacility power = IT power × PUEElectricity = facility power × hoursEmissions = electricity × mean carbon intensity over the jobCost = electricity × mean day-ahead price over the job
The best start is the hour in the next 48 hours, among those where the whole job still fits, that gives the lowest mean carbon intensity. A job longer than the forecast is priced on the mean of the hours that remain, and the result says so.
Defaults. Server overhead 1.8×: the NVIDIA DGX B200 (8 × B200 at 1,000 W) is rated at about 14.3 kW, 1.79 times its GPUs' 8 kW of board power (NVIDIA). That is a nameplate maximum, so typical draw is lower. PUE 1.56: Industry-average annual PUE, Uptime Institute Global Data Center Survey 2024 (Uptime Institute). Average load 70%: an assumption, set your own.
Cost is the wholesale day-ahead component only. Network charges, levies, taxes and contract prices are not included. Where a source licence does not allow prices to be republished, cost is left blank rather than estimated.
| Accelerator | Board power | Source |
|---|---|---|
| NVIDIA B200 (SXM) | 1,000 W | Lenovo Press, ThinkSystem NVIDIA B200 180GB 1000W GPU (LP2226) |
| NVIDIA H200 (SXM) | 700 W | NVIDIA H200 product specifications |
| NVIDIA H100 (SXM) | 700 W | NVIDIA H100 product specifications |
| NVIDIA H100 NVL (PCIe) | 400 W | NVIDIA H100 product specifications (350–400 W, configurable) |
| NVIDIA A100 (SXM) | 400 W | NVIDIA A100 product specifications |
| NVIDIA L40S (PCIe) | 350 W | NVIDIA L40S product specifications |
| NVIDIA L4 (PCIe) | 72 W | NVIDIA L4 product specifications |
| AMD Instinct MI355X | 1,400 W | The Register, 12 June 2025, reporting AMD's launch specifications |
| AMD Instinct MI325X | 1,000 W | AMD Instinct MI325X product specifications |
| AMD Instinct MI300X | 750 W | AMD Instinct MI300X product specifications |
| Intel Gaudi 3 (OAM) | 900 W | Intel Gaudi 3 HL-325L OAM product brief |
Board power excludes host CPUs, memory, networking and cooling. Figures retrieved 2026-10-10.