RedLectern topic
Energy · emissions · water · infrastructure · measurement boundaries · training and inference.
For: Everyone · operators · policy readers.
After this page you can: Compare footprint claims using matching units and boundaries.
Page reviewed
Part of: Artificial Intelligence › AI and society
The environmental impact of AI is not one number. A footprint claim may cover training, inference, or both; measure only computing equipment or include the wider data centre; and count operational effects or hardware production and disposal as well. Those choices change what the result means. To compare claims, first identify the activity and the unit of useful work, then check which equipment, lifecycle stages and environmental effects the measurement includes.
Energy, greenhouse-gas emissions and water use are related but distinct. Electricity use does not by itself tell you emissions: those depend partly on the electricity supply and the accounting method. Water can be used directly for cooling and indirectly in electricity generation; hardware production can also use water. A useful comparison states its workload, location where relevant, time period, units and boundaries. If a report gives a per-query figure, check what kind of query and what supporting infrastructure it includes. If those details differ or are missing, treat the figures as different views of a footprint rather than a direct ranking.
Read a worked explanation of how serving measurements can include computing equipment, idle capacity, data centre overhead and water.
Start with: Measuring the environmental impact of AI inference
Use the OECD report to see lifecycle stages, direct and indirect impacts, and measurement gaps.
Start with: Measuring the Environmental Impacts of Artificial Intelligence Compute and Applications
A footprint is meaningful only with a boundary: what activity is counted, what equipment and infrastructure are included, and which lifecycle stages are in scope. The AI compute lifecycle can include hardware production, transport, operations and end-of-life. Operational measurement can cover training, when a system is developed, and inference, when a deployed model is used to make predictions or generate outputs. A claim about one stage should not be read as a whole-lifecycle total.
Also identify the functional unit: the useful work to which an intensity is attributed, such as a request or task. Absolute totals and per-unit intensities answer different questions. When comparing two claims, align their workload, functional unit, lifecycle coverage and system boundary first; if any of these are unknown, record that gap instead of treating the numbers as directly comparable.
Maps direct impacts across production, transport, operations and end-of-life, and discusses indirect effects.
Carbon footprint
Greenhouse-gas emissions attributed to an activity within stated boundaries.
Water footprint
Direct and indirect water use attributed to an activity within stated boundaries.
Two reports give emissions per request. One counts only active computing equipment; the other includes idle capacity and data centre overhead. Name two details you would check before comparing their results.
Check whether both measure the same workload and functional unit, and whether both include the same system boundary, such as idle capacity, host equipment and data centre overhead.
Energy use and greenhouse-gas emissions are connected, but they are not interchangeable measures. Emissions attributed to electricity depend on the carbon intensity assigned to the electricity used and on the accounting boundary. A per-task intensity should name the useful work being measured; an absolute total describes a different scale. Some accounting approaches also allocate a share of hardware's embodied emissions to each unit of work.
For a comparison, note whether energy means equipment energy alone or includes supporting infrastructure such as cooling and power distribution. Check whether the emissions figure uses location-based electricity intensity or another method, and whether hardware production is included. A measured value and an estimate should also be labelled as such. If a claim omits these details, its precision may exceed what its boundary supports.
Explains data centre electricity demand and the equipment and infrastructure that contribute to it.
Describes a provider's inference measurement boundary, including supporting data centre systems.
A report gives a carbon intensity per task. List three pieces of information you would look for to understand or compare it.
Identify the functional unit and workload, the electricity carbon-intensity method and location, and whether the boundary includes data centre overhead and embodied hardware emissions.
Water accounting needs its own boundary. Data centre operations can use water directly for cooling and indirectly through electricity generation; hardware production, including semiconductor fabrication, is another lifecycle stage where water can be used. A water claim should say which of these uses it counts and whether it reports withdrawal, consumption or another measure. The location and timing of water use matter when considering local resource pressures.
Infrastructure choices affect more than the computing device doing the work. Cooling, power distribution, host equipment and provisioned capacity can contribute to operational resource use. When a report focuses on AI alone, ask how shared data centre resources are attributed to the workload. Compare water claims only when their units, sources, locations, time periods and included lifecycle stages align; otherwise, describe the differences in scope.
Includes cooling water and supporting infrastructure in its description of an inference measurement approach.
A water-footprint claim includes cooling water but says nothing about electricity generation or hardware production. Name the boundary gap and one detail to request.
The claim covers only some water uses across the lifecycle. Ask whether indirect water used in electricity generation and water used in hardware production are included, and request the location and water measure reported.
No recent updates are available.
Kate Crawford · 2022
A readable account of AI’s material costs, including the supply chains and resources behind its infrastructure. It helps readers see why a footprint boundary can extend beyond model electricity.