Energy and sustainability
What a request actually consumes, how to compute it, and which figures hold up in sustainability reporting.
The idea
Every request makes a graphics card work briefly. Consumption is power times time. A 400-watt card working for one second uses 400 watt seconds, about 0.11 watt hours. If it serves twenty requests concurrently in that second, each accounts for roughly 0.0056 watt hours.
Batching, which lowers cost, therefore lowers consumption per request in the same proportion.
Orders of magnitude
| Operation | Order of magnitude |
|---|---|
| Web search | ~0.3 Wh |
| Short answer from a mid-sized model | 0.1 to 1 Wh |
| Long answer from a large model | 2 to 10 Wh |
| One generated image | 1 to 5 Wh |
| One kilometre by car | ~1,900 Wh |
Measuring it yourself
- 01
Record power draw
Through the card management interface, averaged over the measurement window, not as a peak.
- 02
Measure throughput at the same time
Requests and tokens per second in the same window.
- 03
Convert to per request
Power divided by throughput gives energy per request.
- 04
Add the overheads
Processor, power supply, cooling. The data centre PUE covers this.
Worked through: P = 400 W, R = 12 requests per second, PUE = 1.3:
E = (400/12) × 1.3 / 3600 = 0.0120 Wh per request. At a million requests a day
that is 12.0 kWh per day.
From energy to emissions
The factor I varies substantially: below 50 g per kWh in some European grids and
above 400 g in others. The same service can therefore be eight times more
emission-intensive depending on location. Reporting must state the location-based
factor, not a European average.
What holds up in a report
- Measured energy consumption of your own equipment, not extrapolated from vendor figures.
- The PUE of the data centre actually used, with a reference period.
- The location-based emission factor, with source and year.
- For bought-in services: the provider's figures, explicitly labelled as such.
- Training and operation reported separately.
Not defensible: estimated per-request consumption from third-party publications, vendor peak power figures used as operating values, and any figure without a utilisation statement.
The most effective levers
| Lever | Order of magnitude |
|---|---|
| Increase batching | 5 to 20 times per request |
| Smaller model for easy requests | 5 to 20 times on the affected share |
| Quantisation to int8 | ~2 times |
| Cap prompt and answer length | Linear |
| Site choice by electricity mix | Up to 8 times on emissions |
The last row changes emissions, not consumption. It therefore belongs reported separately and must not be mixed with efficiency measures.