Power Usage Effectiveness (PUE)

PUE is the power a whole data centre draws divided by the power that reaches the computers inside it. A PUE of 1.5 means that for every watt the chips use, another half-watt goes on cooling, power conversion and everything else in the building. 1.0 would mean no overhead at all.

It matters to the price of a token because power, not money, is what limits how many chips a site can switch on. A site's grid connection is fixed; PUE decides how much of it does computing. At 1.09, about 92% of the power bought reaches the chips. At 1.52, about 66% does. On the same connection, the second site serves roughly 28% fewer tokens.

Three numbers, three kinds of evidence

Who	PUE	Period	What kind of number it is
Google fleet	1.09	2025 (TTM); still 1.09 at Q2 2026	Operator's own figure, widest stated boundary (Google)
Microsoft fleet	1.17 (1.16 in FY24)	FY25, to June 2025	Operator's own figure; owned sites only, 12 months operational (Microsoft)
Industry average	1.52 (1.54 in 2025)	2026 survey	Self-reported by 644 operators, averaged per site (Uptime 2026)

The overhead per unit of computing is 0.09, 0.17 and 0.52. The gap between the two hyperscalers is small; the gap between either of them and the average operator is five to six times larger.

Key Claims

• The industry average has barely moved in seven years. 1.52 in 2026, 1.54 in 2025, from 2.50 in 2007; Uptime calls it "a seven-year trend of relative stasis". The cheap fixes (warmer air, aisle containment, blanking panels) were used up by around 2014, and new efficient sites only lower the average slowly because old sites are not closed: 27% of respondents work in data centres over 15 years old. Evidence: moderate (self-reported survey, large sample) (Uptime 2026)
• New builds are far better than the average. Respondents' largest sites average 1.45; new builds "routinely" reach 1.3 or better in suitable climates; 23% of respondents work at a site averaging below 1.3. Evidence: moderate (self-reported survey) (Uptime 2026)
• Google's fleet has been at 1.09–1.12 for a decade and at 1.09 every quarter since Q4 2024. There is almost no overhead left to remove, which is why Google now points to chip efficiency (TPUs, "over 3 times more compute performance per unit of energy than five years ago") rather than the building. Evidence: moderate (operator disclosure, unaudited) (Google)
• Microsoft's fleet got slightly worse while AI load ramped: 1.16 → 1.17 from FY24 to FY25, with Asia Pacific at 1.25 → 1.28. Evidence: moderate (operator disclosure, unaudited) (Microsoft)
• Climate sets the floor. Google's least efficient campuses are in hot or humid places (Singapore 1.15, Storey County NV 1.15, Mesa AZ 1.28 quarterly), and Uptime expects the average to improve more slowly as building moves into subtropical and desert regions. Evidence: moderate (Google; Uptime 2026)
• Both hyperscaler figures leave out their newest sites. Google reports a campus's TTM only after twelve months; Microsoft excludes sites operational under twelve months and all leased capacity. Google's own table shows a new site can sit well outside the fleet figure in its first year, either way: Mesa at 1.28, Fort Wayne at 1.05. Evidence: moderate (read from both pages) (Google; Microsoft)
• PUE can be bought with water. Evaporative cooling lowers PUE and raises water use; closed-loop cooling does the reverse. Microsoft's Asia Pacific row shows both moving the wrong way in one year (WUE 0.03 → 0.25 L/kWh, PUE 1.25 → 1.28), and its "zero water for cooling" design states no power cost at all. A PUE figure read without a water figure is half the picture. Evidence: moderate (operator disclosures) (Microsoft metrics; Microsoft water blog)

Why the boundary matters more than the decimal

Each operator decides what counts as overhead and which sites are in the figure. Google says it includes substations, transformers, water treatment, offices and cafeterias, and that leaving them out would give "1.06 or less". Microsoft states a scope (owned, operational 12 months) but not that list. Uptime's average is each respondent's own site average, counted once regardless of size. So the three numbers are directionally comparable, never precisely comparable, and none of them was measured by anyone outside the operator.

Benchmarks & Data

• Google fleet TTM PUE: 1.09 (2025); Q1 2026 1.08 quarterly / 1.09 TTM; Q2 2026 1.10 / 1.09
• Google campus TTM range, Q2 2026: 1.04 (Lancaster OH) to 1.15 (Storey County NV; Singapore 2nd)
• Microsoft fleet PUE: 1.16 (FY24) → 1.17 (FY25); WUE 0.30 → 0.27 L/kWh
• Uptime industry average PUE: 1.54 (2025) → 1.52 (2026, n = 644); largest sites 1.45; 23% below 1.3
• Overhead per kWh of computing: Google 0.09 · Microsoft 0.17 · industry 0.52
• Share of bought power reaching the chips: ~92% at 1.09 · ~85% at 1.17 · ~66% at 1.52 (arithmetic)

Open Questions

• What is the PUE of a dedicated AI hall at full load? None of the three sources separates AI sites; Uptime says purpose-built AI facilities are still too few to move its average.
• Does liquid cooling lower PUE, or just move the problem? Uptime expects direct liquid cooling at higher temperatures to help "over the longer term"; Microsoft's figure worsened as its AI load rose. No source measures it.
• What do the leased and colocated sites run at? Both hyperscaler figures exclude them, and a large share of AI capacity is leased.
• What is the water price of the hyperscalers' low PUEs? Google's page gives no water figure.

Related Concepts

• Rack Power Density — denser racks force liquid cooling, which changes both PUE and water use
• Grid Connection Queue — when the connection is the limit, PUE decides how much of it computes
• Hyperscaler Capex Cycle — the plant and cooling slice of capex is what buys a lower PUE

Changelog

• 2026-09-24 — Created from four sources: Google's fleet PUE page (new), Uptime's 2026 survey (new), and the two Microsoft sources ingested 2026-09-11 but not compiled until now. Graduated from the "still to compile" list on the frontier page.