Measuring energy and water efficiency for Microsoft datacenters

What this is

Microsoft's public efficiency metrics page: fleet-wide PUE and WUE for two fiscal years, with a three-region split and an explicitly stated scope boundary. The KB already holds Google's 1.09 fleet PUE (unpicked, 2026-08-30), which is a best-case outlier from the operator with the longest-running efficiency programme. This is a second hyperscaler on a stated methodology, which converts a single boast into a comparison.

Definitions, as the operator states them

• PUE — "an industry metric that measures the energy efficiency of a datacenter", computed as total energy needed for facility ÷ total energy used for computing. Closer to 1.0 is more efficient.
• WUE — "measured in liters per kilowatt hour", computed as annual liters of water used for humidification and cooling ÷ total annual kWh used to power IT equipment.

Note what each denominator is: PUE's denominator is compute energy; WUE's denominator is IT equipment energy. They are not the same base, so the two ratios cannot be composed.

Results

Global fleet

By region

Methodology and scope

Covers "datacenters that Microsoft fully owns and controls and that were operational for 12 months at the time of calculation." FY24 = 1 Jul 2023–30 Jun 2024; FY25 = 1 Jul 2024–30 Jun 2025.

Two consequences of that boundary, and they are the whole reason to read this page carefully:

1. Leased and colocated capacity is excluded. A large and growing share of AI capacity is leased. The disclosed fleet is therefore the part Microsoft designs and runs, not the part it rents.
2. The 12-month-operational filter excludes ramping sites. Microsoft says so directly: "Global and regional PUE and WUE are expected to continue improving as datacenters reach full operational capacity." A site's worst efficiency year is its first, and this metric never sees it.

Why it matters here

• A PUE that degrades while AI load ramps is the signal. 1.16 → 1.17 is small in absolute terms, but the direction is the finding: liquid-cooled, high-density AI halls are not (yet) making the facility overhead ratio better, even under a scope that filters out new sites. This bounds how much of the buildout can be presented as efficiency-constrained rather than generation-constrained — which is the assumption the three published power calls rest on.
• PUE and WUE trade against each other. The APAC row is the cleanest evidence in the table: WUE jumps 0.03 → 0.25 L/kWh and PUE worsens 1.25 → 1.28 in the same year. Evaporative cooling buys power efficiency with water; closed-loop buys water with power. Any argument that one number is improving should be checked against the other.
• Ceiling, not floor. Against Google's 1.09, a 1.17 fleet PUE says the efficiency headroom between hyperscalers is roughly 7 percentage points of facility overhead — nowhere near enough to absorb the demand growth the buildout assumes.

Limitations

Operator self-disclosure, unaudited here, on a boundary the operator chose. No per-campus figures — Microsoft publishes region, not site, so nothing here can be matched to a specific interconnect queue or a specific power call. The page is undated and live: the FY labels are the only anchor, and the numbers can change under the same URL. The APAC WUE move (0.03 → 0.25) is large enough to suggest a fleet-composition or methodology change rather than an operational one; the page gives no explanation, and none should be inferred.

Source: Measuring energy and water efficiency for Microsoft datacenters, Microsoft