AI environmental impact concept
How to measure AI energy use
ReferenceTo measure AI energy use, define the workload and boundary, measure the equipment included over a stated period, then divide the energy by the work completed if you need a per-query result. Training means building a model by learning from data; inference means using that trained model to make predictions or generate responses. Measure and report those activities separately, because their workloads and boundaries differ.
Reviewed 2026-10-09
Choose what the measurement means
Start by naming the activity: model training, fine-tuning, or inference, which is the use of a trained model to produce an output. State whether you are measuring one experiment, a batch of requests, or a service over a reporting period. A measurement has a boundary: it may cover only the accelerator doing the computation, or also the host computer, idle capacity, and data centre overhead. Write down what is inside that boundary and what is left out before collecting data. That makes the result interpretable and helps readers compare like with like.
Measure, then calculate the unit you need
First, describe the workload: the task, input set, output limit, hardware, and whether requests run one at a time or together. Next, collect energy for the chosen boundary while that workload runs; use direct equipment or system readings when available, or clearly label a calculation based on runtime and estimated power as an estimate. For a per-query result, divide the measured energy for a defined request batch by the number of completed requests in that same batch. For training, report the energy for the training run and its run boundary separately from serving requests.
Common ways estimates go wrong
A device reading can miss energy used by the host computer, machines held idle for reliability, or data centre cooling and power distribution. Conversely, multiplying a chip’s rated power by elapsed time is an estimate, not a measurement of its actual workload energy. A per-query average also depends on which requests were counted and how much work each request caused; a median and a mean answer different questions. If you convert energy into emissions, that is a separate calculation that needs its own location, time period, electricity carbon-intensity source, and boundary.
Check whether the result can be understood
Before sharing a result, check that another reader can identify the workload, measurement period, hardware or service boundary, measurement method, and denominator used for any per-query figure. Confirm that training and inference are not combined without explanation, and that idle equipment or facility overhead is either included or named as excluded. Keep the raw total alongside a per-request average so the division can be checked. If you used an estimate, show its inputs and label it as an estimate; if a provider supplied the figure, identify the provider and the scope it says the figure covers.
Questions people ask
How do I calculate energy per AI query?
Measure or estimate energy for a defined batch of requests, then divide by the completed requests in that batch. State the request set and which equipment and facility overhead the energy figure includes.
How is AI training energy different from inference energy?
Training builds a model by learning from data; inference uses a trained model to make predictions or generate outputs. Report them separately because they are different workloads, and define the boundary for each.
Can I work out energy use from runtime alone?
Runtime alone does not show actual energy use. A runtime-and-power calculation can be an estimate, but it depends on the power assumption and which computing equipment the estimate covers.
Should data centre cooling be included?
Say whether it is included. Equipment energy and the energy used by supporting data centre infrastructure are distinct boundaries; PUE is a data centre metric relating total facility energy to IT equipment energy.
Learn more
- Measuring the environmental impact of delivering AI at Google Scale
A production example of defining a serving boundary that includes active accelerators, host energy, idle capacity, and data centre overhead.
- Green Algorithms: Quantifying the carbon footprint of computation
Explains an estimate based on processing time, computing resources, memory, and facility efficiency and location.
- ISO/IEC 30134-2:2026: Power Usage Effectiveness
Explains PUE and how it relates computing equipment energy to total facility energy.
What changed
Sources
Part of AI environmental impact: Energy and emissionsWhat does a footprint include?