In short
The Software Carbon Intensity (SCI) score, published as ISO/IEC 21031:2024, is ((E × I) + M) per R: energy used, times the location-based grid carbon intensity, plus a share of the hardware's embodied emissions, divided by a functional unit such as one request or one document. Choose the unit, draw the boundary, gather or estimate each term, and recalculate after every change.
Steps
- Choose a functional unit that matches the feature's value, such as per request or per document.
- Write down the boundary, saying which hardware and services are included.
- Measure or estimate the energy (E) used by the feature over a period.
- Multiply by the location-based grid carbon intensity (I) for where and when it ran.
- Add the share of embodied hardware emissions (M) for the same period.
- Divide by the number of functional units (R) delivered in that period.
SCI turns an AI feature's footprint into a rate you can track and improve. It is a rate rather than a total, so it shows whether each request is getting greener even as use grows.
What goes into the formula?
SCI = ((E × I) + M) per R
| Term | Meaning | Where to get it for AI |
|---|---|---|
| E | Energy used, in kWh | Meter your own servers or GPUs; estimate hosted APIs from tokens |
| I | Location-based grid carbon intensity, gCO₂e/kWh | Regional, ideally hourly, grid data |
| M | Embodied emissions allocated to the period | Hardware manufacturer or cloud provider data, shared by use time |
| R | Functional unit | Per request, per answer, per document, per successful task |
The Green Software Foundation's guidance on the I term recommends the most frequent, region-specific data you can get, falling back to averages only when nothing finer exists.
What the specification doesn't allow
Market-based measures such as renewable certificates, power purchase agreements and offsets do not reduce an SCI score. The score only falls when the software uses less energy, runs on cleaner electricity at the time, or uses less hardware.
Common mistakes
- Choosing a unit that hides waste. "Per call" ignores retries; "per successful task" doesn't.
- Leaving M out silently. If you can't estimate embodied emissions, say so in the boundary.
- Comparing scores with different boundaries or units across teams.
Related: estimating energy of AI API calls and accounting for embodied carbon. The full measurement method is in how to measure AI energy.
Sources
- Software Carbon Intensity (SCI) specification, Green Software Foundation (2024)
- ISO/IEC 21031:2024 Software Carbon Intensity (SCI) specification, ISO (2024)
- SCI guidance: carbon intensity (I), Green Software Foundation (2024)
Related
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 account for the embodied carbon of AI hardware?
Include the emissions from making AI chips and servers: allocate a share by use time, ask suppliers for data, and keep hardware in service longer.
How do I find the carbon intensity of my electricity grid?
Where to get grid carbon intensity data for AI footprints: regional, hourly sources such as the GB Carbon Intensity API and Electricity Maps.
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.