In short
Measure a baseline first, then set two kinds of target together: an intensity target (energy or emissions per useful task) and an absolute target or cap on the total. Intensity shows efficiency; the absolute figure shows whether growth in use is eating the savings. Review them at least yearly as models and usage change.
Steps
- Measure or estimate a baseline for your main AI uses with a written method.
- Choose a functional unit, such as energy per request or per document processed.
- Set an intensity target per unit and an absolute target or cap on the total.
- Add process targets you control directly, such as the share of requests served by small models.
- Review progress and the method every year and restate the baseline if the method changes.
Set an intensity target and an absolute target together, from a measured baseline. Either alone can mislead: per task figures can fall while the total climbs, and a total can look fine because use simply has not grown yet.
Why two targets?
Efficiency per request can improve very fast. Google reports that the energy of its median Gemini text prompt fell 33-fold over a year. But cheaper AI tends to get used more, which Luccioni, Strubell and Crawford discuss as the rebound or Jevons effect. An intensity target captures the first; an absolute target or cap captures the second.
What should the targets look like?
| Type | Example |
|---|---|
| Intensity | Energy per document processed down by a set share against the baseline |
| Absolute | Total AI electricity, or estimated emissions, held at or below a cap |
| Process | Share of requests handled by small models; share of suppliers giving per-request data |
The Green Software Foundation's Software Carbon Intensity specification gives a ready-made intensity measure: emissions per functional unit, using location-based grid data. See calculate an SCI score.
Common mistakes
- No baseline. See set a baseline.
- Targets met by certificates. Buying certificates can change a market-based figure without changing energy use. Keep at least one target in kilowatt-hours.
- Changing the method silently. If the method improves, restate the baseline.
- Targets nobody controls. Pair outcome targets with process targets that teams can act on.
The method behind the numbers is in how to measure AI energy.
Sources
- Software Carbon Intensity specification, Green Software Foundation (2024)
- From efficiency gains to rebound effects: the problem of Jevons' paradox in AI's polarized environmental debate, Luccioni, Strubell and Crawford, ACM FAccT (2025)
- Measuring the environmental impact of delivering AI at Google Scale, Elsworth et al., Google (2025)
Related
How do I set a baseline to prove AI energy savings?
Show a real AI energy saving: measure before any change on the same set-up and load, repeat runs, report per unit of work, and check quality held.
How do I track AI energy per task rather than in total?
Pick a functional unit such as per document or per successful task, log energy and quality against it, and report it next to the total.
How do I stop AI efficiency gains being cancelled out by more use?
How to keep AI efficiency savings from being swallowed by growing use: track totals, cap or budget use, retire freed capacity and ask if it is needed.
How to measure the energy and carbon of AI
A practical guide to metering or estimating the energy and carbon of AI work: tools, system boundaries, per-task metrics and honest uncertainty.