AI's energy and water footprint: what numbers can tell us
AI infrastructure has physical costs, but a single headline number rarely explains the footprint of a specific model or workload.
Topic diagram
Define the boundary before comparing footprints
Key takeaways
- Check the system boundary behind every footprint estimate.
- Grid mix, cooling design and workload affect environmental impact.
- Ask vendors for methodology and comparable disclosures.
- Measure quality and resource use together when optimizing.
Why this question matters
AI services run on computing infrastructure that consumes electricity and requires cooling. Some facilities also use water directly, while electricity generation can have indirect water impacts. Public estimates differ because they measure different systems, locations, workloads and accounting boundaries.
Know what is being measured
Training one model, serving a million requests and operating a data center are different units of analysis. Estimates can include or exclude hardware manufacturing, cooling, electricity generation and idle capacity. A number without a boundary and time period is difficult to compare.
Location changes the result
Electricity carbon intensity varies by grid and time. Cooling systems also differ according to climate, facility design and water availability. Two facilities with similar computing demand may have different environmental impacts because they use different power sources and cooling approaches.
Inference is a growing question
Once a model is deployed, repeated inference can accumulate significant energy use. The footprint depends on model size, output length, batching, hardware utilization and the amount of computation required for each request. More capable systems may use more compute, but optimization can reduce resource use per task.
Ask vendors for useful disclosures
For procurement, request the scope of reported emissions, measurement period, location-level information where available, renewable electricity accounting, water metrics and methodology. Treat unsupported claims such as “carbon neutral” as prompts for clarification rather than proof of impact.
Efficiency is more than model size
Teams can reduce waste by choosing the smallest model that meets quality requirements, caching repeated work, limiting unnecessary context, batching suitable jobs and retiring unused deployments. Measure quality and resources together; a smaller system that causes repeated retries may not be more efficient overall.
Avoid false precision
Public figures can help establish scale, but they may not map to a particular user's request. Prefer transparent ranges and disclosed assumptions. The useful question is not only how much energy AI uses, but what task it performs, what alternatives exist and how its resource use can be reduced.
Common questions
Can every AI prompt be assigned one energy figure?
Only with a stated method and assumptions. Workload, hardware, utilisation, output length and system boundaries change the result.
What should a business ask a provider?
Ask how energy and water are measured, what boundary the figures cover, where workloads run and whether the data is measured or estimated.
Build a comparison that readers can reproduce
When publishing an environmental comparison, state the unit, time period, workload and system boundary. Explain whether the figure covers model training, inference, cooling, electricity generation or only a portion of the process. If a provider gives a company-wide figure, do not divide it by an arbitrary number of users and present the result as a measured per-request value. Where the method is incomplete, say so.
For an internal project, start with a baseline of requests, average input and output sizes, retries and the model or service used. Test whether shorter context, caching, a smaller model or a deterministic alternative can achieve the same outcome. Measure quality at the same time. If an efficiency change causes more retries or human correction, the apparent saving may not survive at the system level.
Include local conditions in infrastructure decisions
Electricity availability, grid carbon intensity, water stress and cooling practices vary by location and over time. A general industry estimate cannot settle the impact of a specific facility. Ask providers how they account for direct and indirect water use, whether figures are measured or modelled, and how frequently they are updated. Compare like with like and avoid turning a complex infrastructure question into a single ranking.
Responsible reporting is transparent about uncertainty. It identifies the assumptions, separates measured facts from estimates and explains what the numbers cannot establish. This gives readers a basis for comparison without overstating precision or suggesting that one efficiency measure resolves every environmental concern.
Questions for a credible environmental comparison
Does the figure cover training, inference or the wider facility? Is the value measured or estimated? What workload and time period does it describe? Are direct water use and indirect electricity-related impacts separated? Can the assumptions be compared with those used for another provider?
If the source cannot answer these questions, present the figure with clear limitations or leave it out. Transparent uncertainty is more useful than an exact-looking number that readers cannot reproduce.
Separate efficiency from total demand
A more efficient model can lower the resources needed for one task, yet total consumption can still rise if the organisation dramatically increases the number of tasks it runs. Track both the resource intensity of a workload and the total volume of activity. Explain the time period and system boundary when reporting progress, and avoid claiming a reduction from a single benchmark that does not reflect production use.
For a public claim, preserve the methodology and source data that support it. If the figures are estimates, label them and explain the assumptions. Clear boundaries make comparisons more useful and help readers see where uncertainty remains.
Questions to ask about footprint estimates
Before comparing two energy or water figures, check whether they describe model training, one inference, a whole data center or a wider supply chain. Look for the time period, workload, hardware, utilization, cooling approach and local electricity or water context. Distinguish direct measurements from modeled estimates and avoid treating a number from one workload as a universal rate for every AI request. When reporting a footprint, state the boundary and assumptions alongside the result. These details make comparisons more honest and help teams identify which operational choices they can actually change.
Founder analysis
Applying a consistent measurement boundary
An AI footprint estimate is only meaningful when its boundary is clear. Distinguish training from inference, direct facility energy from broader electricity impacts, and water withdrawn from water consumed. Workload, model size, hardware utilization, location, cooling design, and the electricity mix can change the result substantially. Prefer measured, disclosed methods and ranges over a single number presented without assumptions.
Sources and further reading
These references provide additional context for the frameworks discussed in this article. Check the source directly for its scope, date and limitations.
- International Energy Agency: Energy and AI
- U.S. Department of Energy: Data center electricity-demand report
Related Ninth Post reading
- RAG vs fine-tuning: which approach does your AI product need?
- An AI data privacy checklist for product teams
- How to evaluate an AI feature before launch