Google Data Centers — Power usage effectiveness (fleet PUE 1.09, 2025; quarterly data to Q2 2026)
Google fleet TTM PUE 1.09 for 2025 and still 1.09 at Q2 2026 (quarterly 1.10); flat at 1.09-1.12 for a decade; campuses 1.04-1.15 TTM, newest Mesa AZ at 1.28 quarterly. Widest stated overhead boundary (substations, offices, cafeterias). Against Microsoft 1.17 (FY25) and industry 1.52, overhead per kWh of compute is 0.09 vs 0.17 vs 0.52.
What this is
Google's public page on data-centre efficiency. It gives a fleet-wide PUE for 2025, a quarterly and trailing-twelve-month (TTM) PUE for the whole fleet and for each campus, back to 2008, and a note on what Google counts as overhead. It is the operator describing itself.
Key figures (as of 2026-09-24; all read from the page)
- 2025 fleet-wide average PUE: 1.09 — "a comprehensive trailing twelve-month (TTM) PUE of 1.09 across all our large-scale data centers (once they reach stable operations), in all seasons, including all sources of overhead."
- Latest quarters on the page: Q1 2026 quarterly 1.08, TTM 1.09; Q2 2026 quarterly 1.10, TTM 1.09.
- The fleet TTM has barely moved in ten years: 1.11–1.12 through 2016–2017, 1.10–1.11 from 2018 to 2023, 1.09–1.10 in 2024, and 1.09 every quarter since Q4 2024. The quarterly figure follows the weather: it peaks every third quarter (1.11–1.14 across 2016–2025) and is lowest in the first and fourth.
- Per campus, Q2 2026: 35 campuses listed. TTM PUE ranges from 1.04 (Lancaster, Ohio) to 1.15 (Storey County, Nevada, and the second Singapore facility). Three new campuses show a quarterly figure but no TTM yet, because Google reports campus TTM only after twelve months of data: Mesa, Arizona at 1.28 (1.22 in Q1), Red Oak, Texas at 1.14 and Fort Wayne, Indiana at 1.05. Hot and humid sites (Singapore, Arizona, Nevada, Taiwan) are the least efficient.
- What counts as overhead: besides mechanical and electrical losses, Google includes energy used by substations, transformers, water-treatment plants, accessory buildings (cafeterias and gyms) and office space. Google's methodology note (written when the formula was widened in 2013) says stripping these out would give "a PUE of 1.06 or less".
- Google's own comparison: "83% less overhead energy than the industry average", computed as 1 − (0.09 ÷ 0.54) against the Uptime Institute 2025 survey average of 1.54.
- Compute per unit of energy: "over 3 times more compute performance per unit of energy than five years ago", attributed mainly to TPUs, on a methodology Google says it has updated. This is a claim about chips, not buildings, and the page gives no number behind it.
Set against Microsoft's 1.17 and the industry's 1.52
| Operator | PUE | Period | Boundary |
|---|---|---|---|
| 1.09 (TTM) | 2025; still 1.09 at Q2 2026 | Large-scale sites once at "stable operations"; overhead includes substations, offices, cafeterias | |
| Microsoft | 1.17 (1.16 the year before) | FY25 (Jul 2024–Jun 2025) | Fully owned and controlled sites, operational 12 months; leased and colocated excluded (close read) |
| Industry average | 1.52 (1.54 the year before) | 2026 survey | Self-reported by 644 operators (close read) |
Three things follow, and they are the reason to hold this page:
- Google's boundary is wider, not narrower. Its overhead includes things Microsoft does not say it counts (offices, cafeterias). On a like-for-like boundary the gap to Microsoft would, if anything, widen. That is Google's own claim and cannot be checked here.
- Both hyperscalers leave out their newest sites. Google reports campus TTM only after twelve months; Microsoft excludes sites operational under twelve months. Google's own table shows why that matters, though not in one direction: of its three newest campuses, Mesa is running at 1.28, the worst on the page, while Fort Wayne is at 1.05. A new site's first year can sit well outside the fleet figure either way, and the fleet figure never shows it.
- The gap between hyperscalers is small next to the gap to everyone else. Overhead per kWh of computing is 0.09 (Google), 0.17 (Microsoft) and 0.52 (industry average): an 0.08 spread between the two hyperscalers against a 0.35–0.43 spread to the average.
Why it matters here — the token-to-task path
PUE sets how much of a site's power bill reaches the chips. At 1.09, 92% of the power bought does computing; at 1.52, 66% does. Where power, not money, limits how many chips can be switched on, that ratio sets how many tokens a given grid connection can serve, and so how many chips and how much capex one megawatt of connection can carry. Google's fleet is the floor of that multiplier on record in this KB.
The flat ten-year line is the second finding: at 1.09–1.12 there is almost no overhead left to cut. Further cost gains per token at Google have to come from the chips (the TPU claim), not the building.
Limitations
- Operator self-disclosure, unaudited, on a boundary the operator chose. No third party checks these figures on the page.
- Live, undated page. Cite with the period (2025; Q2 2026), never as "current". Earlier campus counts per quarter could not be read reliably from the older tables and are not quoted.
- "Large-scale data centers" only. Any smaller or leased capacity Google uses is outside the figure, and the page does not say how much that is.
- Best case by design. Google has run an efficiency programme longer than anyone; the page is a showcase, and the rung should never use 1.09 as a typical figure.
- No water figure is on this page; the Google-versus-Microsoft comparison is PUE only.
Source: Power usage effectiveness, Google Data Centers. Read 2026-09-24.