Build / Reference data
Step 2 · Build
AI workload energy and carbon calculator
Describe one AI job and see what it draws, emits and costs in each European market, now or at the cleanest start in the next 48 hours.
The grid data is unavailable right now, so emissions and cost cannot be computed. Electricity still updates.
- Electricity
- 264 kWh
- 11.0 kW for 24 h (IT 7.1 kW × PUE 1.56)
- Emissions, start now
- –
- Emissions, best start
- –
- Wholesale cost, start now
- –
Method and defaults
IT power is accelerators × board power × server overhead × average load, and facility power multiplies it by PUE; energy, emissions and cost then follow from the job’s hours and the market’s hourly carbon intensity and day-ahead price. The best start is the hour in the next 48 hours, among those where the whole job fits, with the lowest mean carbon intensity. Defaults are 1.8× server overhead, PUE 1.56 and 70% average load; cost is the wholesale day-ahead component only, and board power excludes host CPUs, memory, networking and cooling.
| Accelerator | Board power |
|---|---|
| NVIDIA B200 (SXM) | 1,000 W |
| NVIDIA H200 (SXM) | 700 W |
| NVIDIA H100 (SXM) | 700 W |
| NVIDIA H100 NVL (PCIe) | 400 W |
| NVIDIA A100 (SXM) | 400 W |
| NVIDIA L40S (PCIe) | 350 W |
| NVIDIA L4 (PCIe) | 72 W |
| AMD Instinct MI355X | 1,400 W |
| AMD Instinct MI325X | 1,000 W |
| AMD Instinct MI300X | 750 W |
| Intel Gaudi 3 (OAM) | 900 W |
Figures checked 2026-10-10. Open the accelerator power dataset.