The power and water footprint of generative AI cannot be reduced to a per-question figure

The environmental impact of generative AI involves training, inference, search, storage, networking, cooling, power supply, and semiconductor manufacturing. The meaning of the provider's figures changes depending on whether they cover only server power or the entire facility, whether they include renewable energy certificates, and whether they refer to water withdrawal or consumption. First, clearly specify the target process, region, period, and allocation method.

The unit for utilization decisions should also be business processes rather than a single question. Even with the same model, processing volumes differ for short classification, large-scale document summarization, long inference, image/video generation, and always-on agents. Include monthly volume, input/output scale, retry rate, caching, and rework due to quality failures, and compare with current business processes, smaller models, and rule-based processing.

Compare unit values based on one business result that meets quality conditions, not per request. Do not hide retries, excessively long inputs, or unused generated outputs from the denominator, and measure current procedures and candidate models under the same completion conditions. If measurement boundaries differ among providers, do not finalize rankings; instead, align on how much of the server, facility, cooling, and power supply is included before making judgments.

The IEA's Key Questions on Energy and AI indicates a central forecast that global data center power consumption will approximately double from 485 TWh in 2025 to 950 TWh by 2030, accounting for about 3% of global power demand that year. While consumption in AI-specialized data centers will triple during this period, efficiency, adoption, and the high-load nature of uses will influence the forecast.[1]

Berkeley Lab's 2025 Update estimates U.S. data center power consumption at 649 TWh (11.8% of total power) for the base case in 2030, showing a range of 521–843 TWh in the compound uncertainty scenario. Since results vary based on assumptions such as AI chip shipments, utilization rates, and equipment lifespan, a single prediction cannot be treated as a fixed value for budgeting or site selection.[2]

An increase in overall data center demand forecasts does not indicate the consumption of a single model chosen by your company.[1] [2] Use demand forecasts as scenarios for capacity, price, and regional constraints, and separately request the provider's measured values and estimation methods for individual procurement. Where a plausible adverse scenario threatens continuity under relevant regional and supplier conditions, prioritize constraints on location, processing time, and usage volume.

Li et al.'s research divides water use into scope 1 for facility cooling, scope 2 for power generation, and scope 3 for equipment manufacturing, among others. For the GPT-3 example, they estimated operational water consumption (scopes 1 and 2) for training at U.S. data centers at 5.4 million liters, with 700,000 liters attributed to scope 1 within the facility. This is an estimate based on model and equipment conditions, not actual measurement disclosure.[3]

The same study shows 500 mL for inference as a range for moderate answers of approximately 10–50 turns, emphasizing differences in water efficiency by location and time. This value should not be applied as a fixed unit value for all models. In regions with high water stress, even small amounts have significant impact, and it is necessary to distinguish whether withdrawn water is returned or consumed through evaporation, etc.[3]

The IEA and Berkeley Lab show demand forecasts for different scopes, and water research handles conditional estimates dependent on equipment, region, and time.[1] [2] [3] Instead of converting this heterogeneity into precise rankings for individual models, create a disclosure reconciliation sheet that lists the numerator, denominator, process, region, period, and allocation method for each candidate.

The reconciliation sheet does not convert each candidate's figures into a single conversion value but divides them into common parts observed under the same conditions and specific parts with different conditions. First, test whether the approval decision changes based on the common parts alone; if it does not, do not add further estimates. Only add missing information that could change the decision to the subject of additional inquiry, linking response deadlines to their impact on judgments.

Classify comparison results into three states: directly comparable with aligned boundaries, comparable only in direction, and unusable for decision-making. Two reviewers independently assign states; do not average discrepancies, but record in the adjudication log which boundary perceptions differed. If judgment is impossible, choose additional disclosure, usage limits, or regional restrictions, and reopen the same decision sheet when data becomes available later.

Environmental values used for approval include premises that drive results, such as activity volume, retries, facility efficiency, processing region, and water classification. Replace each premise with appropriate lower and upper bounds, and check whether the value per business result meeting quality conditions and the total period volume exceed the approval cap. Do not decide based solely on a single central value, and retain the source of the premises.

As a hypothetical example, if the total period volume exceeds the cap by placing the upper bound on retries even if inference efficiency improves, do not confirm the improvement. Compare separate proposals for reducing retries, limiting target business, or halting expansion until disclosures align. Do not mix changes in price or model performance as reasons for environmental value improvements; place them in a separate decision column.

The sensitivity sheet should list premises that reversed the conclusion, those that did not, and unmeasured scopes. If activity volume or processing region falls outside the assumed range after operations begin, update only that premise and re-evaluate. Instead of relaxing quality floors when environmental caps cannot be met, review the usage scope and processing method, and set a re-evaluation date.

• BOUNDARY: Align numerator, denominator, process, region, period, and allocation method
• MISSING: Record exclusions, non-disclosure, and unmeasurable items separately from zero
• SENSITIVITY: Test for reversal of unit values and total period volume with the range of each premise
• DECISION: Determine corresponding measures for comparable, direction-only, and unusable comparisons

During procurement, inquire about inference power by model and region, facility PUE, cooling WUE, water source, water stress, temporal alignment of renewable power, measurement period, third-party verification, and estimation methods. If non-disclosed, do not treat as zero but register as risk with uncertainty, and check cost, capacity, and regional impact under the upper scenario.[1] [2] [3]

In operations, track power, water, and cost per successful business instance, including those with low quality and increased retries. Reduce usage through smaller models, batch processing, caching, shifting processing time and region, and combination with manual labor, while simultaneously maintaining minimum quality and safety conditions. To monitor rebound effects from total volume increases, list both efficiency metrics and monthly totals.[1]

Operational responsibility is divided such that the AI infrastructure manager handles processing volume and retries, the procurement manager handles power and water boundaries and regions, and the business manager holds completion quality. Efficiency improvements do not count as reductions if monthly totals increase. If the provider's disclosure scope changes, processing shifts to a region with high water stress, or retries increase due to quality issues, re-approve model selection and deployment.[1] [3]

• Measurement boundary: Clearly specify how much of training, inference, facility, power, and manufacturing is included
• Activity volume: Record monthly instances, tokens, images, retries, and idle time by use case
• Location and time: Link power composition, water stress, and cooling conditions
• Decision-making: Simultaneously approve minimum conditions and caps for quality, safety, cost, power, and water
• Infrastructure manager: Measure processing volume and retries per successful result
• Procurement manager: Confirm measurement boundaries, regions, and distinctions between estimates and actual measurements
• Business manager: Maintain quality floors even if environmental impact is reduced