Any plan that assumes a known liquid-cooling baseline rests on a denominator nobody has counted.
Researched by an agentic pipeline. Reviewed and approved by the editor before publication.
A planner sizing grid connections for dense AI halls needs two numbers. One is how much facility power each unit of IT power costs. The other is how much of the existing estate already runs on direct liquid cooling (DLC). Only the first now has a figure collected by a regulator. The second exists only as answers to surveys of respondents, and the latest publicly available results come from fieldwork completed in March 2024. Any plan that assumes a known liquid-cooling baseline rests on a denominator nobody has counted.
The first reporting period of the EU's mandatory data centre database (May–September 2024) drew reports from 770 data centres with at least 500 kW of installed IT, around 36% of the estimated number in scope1. The Commission's quality review deemed 70.1% of reported data reliable; 681 sites supplied PUE values that passed1. Across those 681, the average PUE was 1.3625, weighted by each facility's total energy consumption. National averages ranged from 1.15 to 1.66, and sites of 500–1,000 kW averaged 1.641.
The basis matters. These are operator-reported figures, screened for plausibility and consistency against expected ranges. They are not third-party metered and not audited. The Commission itself proposes logical checks at data entry, certification of monitoring and reporting, and clearer measurement methodologies for later rounds1. It is the best public PUE dataset Europe has, and roughly two-thirds of in-scope sites are absent from it.
Three other PUE figures circulate alongside it. They have different bases and must not be averaged together.
| Figure | Value | Basis | Publisher | Denominator |
|---|---|---|---|---|
| EU average PUE, 2024 | 1.3625 | Operator-reported to mandatory database; plausibility-checked; energy-weighted | European Commission (regulator) | 681 sites of 500 kW or more with reliable values |
| Industry average PUE, 2026 | 1.52 | Survey self-report, one value per facility | Uptime Institute (research firm) | 644 respondents |
| Capacity-weighted PUE, 2026 | 1.36 | Same survey, weighted by IT capacity; full method gated | Uptime Institute | Not stated publicly |
| Equinix portfolio PUE, 2024 | 1.39 | Operator-reported annual global average; method not stated on the summary page | Equinix (operator; interested party) | Equinix portfolio |
The near-match between the EU's 1.3625 and Uptime's capacity-weighted 1.36 is not independent confirmation. Both weightings pull the average towards large facilities, and both rest on operators' own numbers. Uptime's per-facility 1.52 is higher because many older, smaller sites still count equally23. Uptime says new, larger facilities "routinely report PUEs at 1.3 or lower"3. The operative word is report. Equinix's 1.39, described as a 6% improvement on 2023, is a figure the company has a commercial interest in publishing6.
The most-cited DLC figure comes from Uptime's 2024 Cooling Systems Survey, fielded 8 February to 13 March 2024 among 964 industry respondents, a mix of owner-operators, designers, vendors and consultants4. Its headline: 22% of respondents report their organisation makes some use of DLC4. That is a share of respondents. It is not a share of racks, servers, megawatts or facilities.
The same survey shows how thin that use is. Among DLC users (n=93), 30% say only a few servers use it, 47% some IT racks, 17% make substantial use, and 5% have at least one data centre that predominantly uses it4. Nearly half of DLC users (n=94) report that less than 10% of their organisation's racks use DLC4.
The 22% is a share of survey respondents whose organisations use DLC somewhere. Nearly half of those users have it on fewer than 10% of their racks. No opened source gives the share of capacity that is liquid-cooled.
Nothing primary and public updates this. Uptime's 2026 Cooling Systems Survey (n=1,034, published 12 August 2026) focuses on DLC adoption, but its results sit behind a paywall5. Uptime's 2026 Global Data Center Survey gives no adoption figure and mentions DLC only as a longer-term efficiency lever2. The sections of the Commission report reviewed here contain no cooling-technology share; they note only that liquid-direct cooling uses minimal or no water1.
What fills the gap is weaker still. A Lenovo executive estimated in May 2025 that less than 3% of the world's servers are liquid-cooled7, an estimate from a company that sells liquid-cooled servers. An aggregator states that 19% of data centres had deployed liquid cooling by early 2026, attributing it to Uptime's 2025 survey8; that attribution was not verified against the primary source for this brief. The three figures count respondents' organisations, data centres and servers respectively. They cannot be lined up into a trend.
Cooling choice follows rack density, and facility scale drives reported PUE: in the EU data, sites of 500–1,000 kW average 1.64 against the energy-weighted 1.36251. But no source gives a PUE for liquid-cooled facilities, so the effect of DLC on facility efficiency cannot be read from any dataset. The only efficiency claim in the evidence, a 15% power saving for a self-contained unit sold through Dell from removing fans, is a supplier founder's statement reported in trade press, with no published measurement basis7.
No source reports DLC costs, cancelled deployments or written-down cooling assets. The absence is a gap in the record, not evidence that none occurred.
Established:
Not established:
Treat any liquid-cooling share in a planning document as unverified until two questions are answered: what is the denominator (respondents, facilities, racks, servers or megawatts), and who counted it. If the answer is a survey of respondents, it measures interest and presence, not installed base, and should not drive power or cooling capacity assumptions.
For PUE, use the EU's 1.3625 as the reference for large European sites and Uptime's 1.52 as a global per-facility self-report. Neither replaces a site's own metered figure. The corrected assumption is this: the industry knows roughly how efficient its large buildings report themselves to be. It does not know how much of its capacity is liquid-cooled.
Do not scale the national 52 GW, or Cleanview's 2 GW of gas, into a figure for data-centre storage.
Oversubscription shows demand for allocations; it does not show how much of the allocated time ran jobs.
Treat memory as a priced, volatile input with a public supplier side and a private buyer side, and plan on the private side.