Can Lower PUE Hide Higher Water Use in an AI Data Center?
Yes. A lower PUE can accompany higher on-site water use, because PUE measures energy overhead, not water demand. Evaporative cooling can trade water for lower cooling electricity. Separately, a growing AI workload can increase total water use even when both PUE and WUE improve. Read the two ratios alongside absolute energy, water volumes, and the same measurement boundary.
Ask for annual facility kWh, annual IT kWh, annual water liters, and the water numerator's definition. Then ask where and when the water is needed. A better efficiency ratio is useful evidence; it is not a complete answer to a site's resource constraints.
Checked October 28, 2026. This is an interpretation guide, not a measured cooling comparison. All A/B/C figures below are invented annual scenarios. Fifteen arithmetic assertions ran on Node.js v24.14.1 on Windows; no data-center meters, GPU workloads, cooling equipment, or LLMs were operated.
Contents and working definitions
- PUE: total facility energy divided by IT equipment energy, over the same period; a dimensionless ratio.
- WUE: water liters divided by IT energy in kWh. The water boundary and withdrawal/consumption convention must be stated.
- Heat rejection: the step that ultimately moves heat from the cooling system to the outside environment.
- Absolute demand: the actual kWh and liters, before dividing by IT activity.
- Indirect water: water associated with electricity generation or other upstream activities, outside a site-only numerator.
1. PUE and WUE answer different questions
The DOE-sponsored NREL design guide, revised July 2024, defines annual PUE as total facility energy divided by IT equipment energy. It characterizes supporting infrastructure, rather than the usefulness or efficiency of the computation itself. A PUE of 1.20 means 1.20 kWh of facility energy per 1 kWh delivered to IT within the stated boundary.
Microsoft's efficiency page defines its WUE using annual water for cooling and humidification over annual IT electricity. Its disclosed global FY24/FY25 values are PUE 1.16/1.17 and WUE 0.30/0.27 L/kWh. FY25 covers July 1, 2024–June 30, 2025, for fully owned and controlled sites operational for 12 months. These are operator-reported fleet ratios, not a same-building experiment or proof that cooling caused the changes.
Notice that those two ratios moved in different directions. Neither establishes total liters without the corresponding IT-energy denominator. Nor should a fleet average be substituted for a specific AI campus.
PUE = facility energy [kWh] / IT energy [kWh]
WUE = stated water volume [L] / IT energy [kWh]
Facility energy [kWh] = PUE × IT energy [kWh]
Water volume [L] = WUE × IT energy [kWh]The last identity reconstructs only the water volume defined in that WUE. It does not automatically yield drinking-water withdrawal, all site consumption, or a lifecycle footprint. Annual ratios should use sums over the period, rather than an unweighted average of hourly ratios.
2. A closed liquid loop does not tell you where the heat goes
A useful distinction is the liquid moving heat away from chips versus the equipment rejecting that heat outdoors. A recirculating chip loop does not, by itself, identify the external cooling method.
In its December 2024 design explanation, Microsoft explicitly said replacing evaporative systems with mechanical cooling would raise PUE, while warmer chip-level cooling and efficient chillers could mitigate the energy impact. It also distinguished cooling water from administrative uses. That is evidence of a possible design tradeoff, not a universal percentage penalty.
The company's June 24, 2026 update describes cooling towers that evaporate water, hybrid systems that switch modes, air-cooled chillers, and new closed-loop direct-to-chip AI designs without operational water evaporation. These descriptions concern different configurations; they do not establish that every existing Microsoft building has adopted the new design.
① Chip → recirculating coolant loop
Ask: what carries the heat away from the processor?
② Facility → outdoor environment
Ask: dry heat rejection, mechanical refrigeration, evaporation, or a hybrid?
Only after both answers: compare the measured energy and water boundaries.
Conceptual diagram, not a piping design or a universal system layout. All essential distinctions are also stated in the text.
3. Two counterexamples expose what a ratio can hide
Synthetic example, not vendor performance: A and B each use 10 million IT kWh in a full year. We assume comparable useful work, service quality, climate, and accounting boundaries solely to isolate the ratios. The numbers are stipulated; they were not derived from a cooling simulation.
| Metric | A | B |
|---|---|---|
| IT energy | 10 GWh | 10 GWh |
| PUE | 1.20 | 1.10 |
| Facility energy | 12 GWh | 11 GWh |
| WUE | 0.20 L/kWh | 0.60 L/kWh |
| Stated site water | 2 million L | 6 million L |
B uses about 8.3% less facility energy than A, but three times the water. There is no contradiction: PUE contains no water term. This demonstrates an accounting possibility; it does not prove a particular evaporative system will deliver these values.
Now suppose annual IT energy doubles to 20 GWh, PUE becomes 1.10, and WUE falls to 0.15 L/kWh. Facility energy is 22 GWh and water is 3 million L. Compared with A, total energy rises about 83.3% and water rises 50%, despite better PUE and WUE.
C isolates expansion, not a failure of efficiency. The ratios still identify improvements per IT kWh. They simply do not say whether demand growth exceeds those gains. Since useful workload was not measured, these examples also cannot establish liters per successful request or per training run.
4. Check location, season, and the missing water boundary
A site-only WUE can omit water associated with producing the electricity it buys. A June 2025 research abstract hosted by Lawrence Berkeley National Laboratory identifies server efficiency, utilization, grid water factors, cooling, and climate among determinants of workload water use. The abstract supports considering multiple factors, not assigning a universal water footprint to every AI request.
For a broader operational estimate, a transparent starting point is stated site water plus electricity consumption multiplied by a relevant electricity-water factor. Use the electricity quantity that factor actually covers, align period and geography, and avoid double counting. This article does not supply a grid factor, calculate lifecycle water, or equate withdrawal with consumption.
An annual ratio also cannot reveal a hot week's peak water draw, the availability of reclaimed water, or pressure on a particular watershed. Treat those as separate evidence requests. Water replenishment elsewhere should be reported separately from physical demand at the campus; it does not remove the need to examine local timing and supply.
5. What this means for AI infrastructure growth
Our interpretation: energy and water constraints can coexist. A campus that fits an annual electricity budget may still need a different heat-rejection design or water supply arrangement. Conversely, reducing evaporation may shift more attention to electricity, equipment cost, and operating conditions. A PUE headline alone cannot identify the lowest-cost or most expandable option.
For an operator or infrastructure investor, request the site-specific cooling configuration, energy and water totals, seasonal peak requirements, and the scope of commissioned capacity. Those inputs matter more than assuming that a lower ratio guarantees an easier buildout. No supplier orders, revenue forecasts, stock-price effects, or investment recommendation follow from the synthetic examples here.
At the workload layer, our guide to aligning GPU telemetry with request timing explains another boundary problem: board-level measurements do not represent whole-facility electricity. Keep that distinction when connecting AI service activity to PUE and water demand.
Comparison checklist
- Match the period: full year against full year; retain monthly and peak-season evidence separately.
- Match the meters: facility electricity, IT electricity, and the explicitly defined water numerator.
- Show the totals: report kWh and liters beside PUE and WUE.
- Identify the system: chip heat transfer and outdoor heat rejection are separate questions.
- Keep unknowns: grid water, local scarcity, useful work, and costs remain unknown until measured or sourced.
The stronger question is not simply “Which campus has the lowest PUE?” It is “Which configuration can deliver the required AI service within this site's energy, water, and operating constraints?”
Appendix · sources and reproducibility
Checked October 28, 2026. Official operator disclosures describe their own scope; they are not an independent comparative benchmark. Fifteen offline arithmetic assertions verified unit conversion, scenario totals, percentage changes, and one energy-weighted PUE calculation. No hardware, facility telemetry, cooling test, or paid API was used. The examples exclude electricity prices, embodied resources, water treatment, construction, and measured service output.
- Revised July 2024; hosted in a 2025-07 directory — NREL/DOE guide, section 8: PUE is annual total facility energy divided by IT energy; infrastructure efficiency is not overall compute efficiency.
- Undated live page; FY24/FY25 table — WUE annual cooling/humidification liters divided by annual IT kWh.
- 2024-12-09 — Evaporative-to-mechanical cooling can raise PUE; warmer chip cooling mitigates energy impact.
- 2026-06-24 — New AI direct-to-chip design recirculates coolant without operational evaporation; cooling tower, hybrid, air-cooled chiller and chip cooling distinctions.
- June2025 journal abstract — Workload water use depends on server efficiency/utilization, electricity-grid water factors, cooling, climate and infrastructure.
댓글
댓글 쓰기
안녕하세요 :)