Uptime Institute Global Data Center Survey 2026 (executive summary)

What this is

The 16th yearly survey of data-centre owners and operators by the Uptime Institute, an industry body that certifies data-centre designs. It ran online from April to May 2026 and collected answers from more than 800 owners and operators (and over 800 vendors and consultants, reported separately). Nearly one in four respondents work for third-party providers (colocation, wholesale, cloud); 52% are in North America and Europe. Not everyone answers every question, so each figure carries its own sample size.

This is a self-reported survey. Each operator states its own site's numbers. They are not metered or audited by Uptime, and the sample is whoever chose to answer. Uptime itself says PUE is "best suited to evaluate a single site over time or to observe directional motion in large sample sizes, rather than comparing between only a few differing facilities."

Key figures (all read from the executive summary)

Efficiency (PUE — total facility power ÷ IT power; 1.0 would mean zero overhead)

• 2026 industry average PUE: 1.52 (question: "What is the average annual PUE for your data center?", n = 644). Uptime calls it "right in line with a seven-year trend of relative stasis".
• The series starts at 2.50 in 2007; the 2025 value was 1.54. The chart's intervening values run between 1.52 and 1.67 [the pairing of each value to its year could not be read cleanly from the extracted chart text, so no year-by-year series is quoted].
• The average PUE of respondents' largest data centres is 1.45; new builds "routinely produce PUE of 1.3 or better" where climate allows fresh-air or evaporative cooling; 23% of respondents work in a data centre averaging below 1.3.
• 27% of respondents work in data centres over 15 years old. Uptime names legacy sites, kept open because capacity is scarce and equipment is expensive and slow to get, as what holds the average up. It expects improvement to "remain slow", made slower by new building in hot (subtropical and desert) regions, and helped only over the longer term by direct liquid cooling running at higher temperatures.

Rack density

• The average modal (most common) rack density passed 11 kW for the first time (n = 679), up from 9 kW in 2025. The jump comes partly from "a small number of new high-density facilities with typical racks above 30 kW"; without them the average is 7.8 kW, up from 7.5 kW.
• Peak density: 24% of respondents now have some racks at 30 kW or more, up from 19% in 2025 (n = 676); about one in seven have some at 50 kW or more, including some sites with racks above 100 kW. The majority (76%) still have no rack at 30 kW or above.
• 4–5 kW is still the most common rack in deployment, and has been for more than a decade.
• As a driver of the highest-density racks, AI inference has matched AI training for the first time: 21% of respondents name each.

Other

• Cost is the top management concern: 79% say management is at least somewhat concerned. Concern about capacity forecasting, power availability and supply-chain disruption is rising.
• More than half of operators now track water use; carbon-emissions tracking is still a minority activity.

Why it matters here — the token-to-task path

PUE is the multiplier between the electricity a site buys and the electricity that reaches a chip. Every served token carries it. At the industry average of 1.52, about 0.52 kWh of overhead is paid for every kWh of computing — against 0.09 at Google's fleet (1.09) and 0.17 at Microsoft's (1.17). For the same power bill, a site at 1.52 delivers roughly 28% less computing than one at 1.09 (1.09 ÷ 1.52 ≈ 0.72; our arithmetic, not Uptime's). Where power is the binding constraint, that is 28% fewer tokens from the same grid connection.

The survey's value is that it is the fleet-wide average, not a best case. The rung's other PUE figures come from two hyperscalers describing their own newest, owned sites. This is the number those should be read against, and it says most of the industry is not getting more efficient quickly.

It also puts a population behind the rung's rack-density page: the 135 kW GB300 NVL72 rack is the top of a distribution whose typical value is 11 kW and whose most common design is still 4–5 kW.

Limitations

• Self-reported, unaudited, self-selected. The operators who answer an Uptime survey may not represent the industry, and there is no way here to check any single answer.
• Only the executive summary was read. The full report (32 pages) is gated; regional splits and any breakdown by operator type in it are not available to this KB.
• "Average PUE" is an average of respondents' reported site averages, not an energy-weighted fleet figure; a small legacy site and a 300 MW campus each count once.
• The respondents' sites are mostly not AI-dedicated. Uptime says purpose-built AI facilities are still few and their influence on the benchmark is "tempered".

Source: Uptime Institute Global Data Center Survey Results 2026, Uptime Institute. Executive summary read from the publicly served PDF; press release read at uptimeinstitute.com.