Is Power Usage Effectiveness (PUE) Still Fit for Purpose? An Integrative Review of Data Center Metrics for Assessing the Environmental Impact of Modern AI Infrastructure

Overview

This study by Waseq Billah, Jasmine Suarez and Erika Wolters, published in 2026 in the Journal of Science Policy & Governance (vol. 1, issue 28), is a policy-oriented integrative review. The search screened 206 hits to 54 peer-reviewed articles published between 2017 and 2025. The paper is free to read on the journal site, but no open licence is stated in the PDF, so it is not redistributed in this library. The central question is whether PUE still works for high-density AI facilities.

The study identifies four blind spots. First, PUE is carbon blind: it treats every kWh alike regardless of source. Grid intensity varies from a world average of 475 gCO2e/kWh to below 60 on clean grids and above 700 on coal-heavy grids, so a low-PUE site on a coal grid can have a worse impact than a high-PUE site on renewables. Second, PUE is purely operational and ignores embodied and end-of-life impacts. Third, it ignores local water and grid scarcity. Fourth, it treats IT energy as a single block, with no view of idle hardware or wasted computation.

Evaporative cooling can improve PUE at a high water cost, creating a perverse incentive in water-stressed regions; for example, a Microsoft data centre in North Holland used 84 million litres of drinking water in 2021, nearly five times its 12-20 million litre design estimate. WUE is annual on-site water use divided by IT energy, and WUEsource adds the water consumed in generating the electricity. Citing Lei and Masanet (2022), in a hot, dry climate a further 10% cut in water use can raise total energy use by 25%, whereas in a cool, humid climate cutting water use costs little energy.

The authors propose a four-axis dashboard instead of a single metric: infrastructure efficiency (PUE), climate impact (CUE), water footprint (WUE and WUEsource) and computational productivity (workload goodput per watt). Two sites with the same PUE can show very different shapes on the spider chart. Embodied carbon was left out because manufacturer disclosures are inconsistent. On policy, the revised EU EED requires public reporting that includes water use and waste-heat use; Singapore pairs PUE limits with Green Mark certification; the authors suggest a 100+ MW threshold for applying interconnection criteria to hyperscale sites.

Key points

1. PUE is carbon blind: grid intensity runs from a 475 gCO2e/kWh world average to below 60 on clean grids and above 700 on coal-heavy grids, so a low-PUE site on coal can be worse than a high-PUE site on renewables.
2. PUE derivatives such as aPUE, bPUE and cPUE add temporal or environmental sensitivity but do not fix the core limitation that PUE measures only power overhead.
3. PUE treats IT energy as monolithic and cannot separate productive computation from idle hardware or wasted work, which motivates goodput metrics.
4. Evaporative cooling can improve PUE at a high water cost; a Microsoft data centre in North Holland used 84 million litres of drinking water in 2021 against a 12-20 million litre design estimate.
5. WUE is annual on-site water use divided by IT energy; WUEsource adds water consumed in generating the electricity, and both are needed for siting decisions.
6. Water-energy tipping point: in a hot, dry climate a further 10% cut in water use can raise total energy use by 25%; in a cool, humid climate the cost is low.
7. Proposed dashboard: PUE, CUE, WUE and WUEsource, and workload goodput per watt; embodied carbon left out because of inconsistent disclosures.

Using it in design

• Give the sustainability panel at least four axes (PUE, CUE, WUE/WUEsource and a placeholder for IT productivity) and show them as a radar chart, so a low PUE never appears alone.
• For CUE, use a site grid-intensity input, defaulting to 475 gCO2e/kWh with a source tag and a range of 60-700.
• When the user picks evaporative or adiabatic cooling, compute WUE, flag high WUE in water-stressed locations and warn that PUE gains are being bought with water.
• Optionally add a water-versus-energy trade-off curve by climate zone for cooling-system selection.

Who should read it

Engineers, policy makers and sustainability specialists who select, report or regulate data centre sustainability metrics.

Related standards

• ISO/IEC 30134-2
• ISO/IEC 30134-8
• ISO/IEC 30134-9
• EU Energy Efficiency Directive (EU) 2023/1791

Publication details

Authors

Waseq Billah; Jasmine Suarez; Erika Wolters

Publisher

Journal of Science Policy & Governance (Journal of Science Policy and Governance, Inc.)

Year

2026

Volume

1

Issue

28

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.38126/jspg280101

Licence

Licence conditions not established — confirm the conditions with the publisher

Citation