What a PUE 1.05 Facility Saves Over a PUE 1.4 Facility in a Year

At a constant 1 MW of IT load, a facility running at PUE 1.4 uses 3,066 MWh more electricity per year than one running at PUE 1.05. That gap is one quarter of the PUE 1.4 facility’s annual draw. Priced at the U.S. industrial average EIA reported for June 2026, it is worth $281,152 a year for each megawatt of IT.

The money sits in the overhead, not in the ratio

The useful quantity in any cooling power comparison is overhead energy, PUE minus one, rather than PUE itself. ISO/IEC 30134-2 defines PUE (power usage effectiveness) as total data center energy divided by IT equipment energy, both recorded over a coincident 12-month period.

Source: ISO/IEC 30134-2:2016, “Information technology — Data centres — Key performance indicators — Part 2: Power usage effectiveness (PUE)”

Read that way, 1.05 and 1.4 are not 33 percent apart. Their overhead values are 0.05 and 0.40, an eightfold difference. A pair of ratios that look close to parity hides a cooling and power distribution bill eight times larger on one side.

The standard also fixes where IT energy is metered. Category 1 reads it at the UPS output, Category 2 at the PDU output, and Category 3 at the IT equipment. A Category 1 figure excludes PDU transformer losses from the overhead, so the reported PUE drops while the utility meter reads the same kWh. The 2026 edition of the standard adds updated measurement requirements.

Source: ISO/IEC 30134-2:2016, clause 6.2; ISO catalog entry for ISO/IEC 30134-2:2026

A PUE that arrives without its measurement category is not usable in a cost model. The category decides how much loss has been moved into the denominator.

At 1 MW of IT load, the gap is 3,066 MWh a year

The arithmetic takes two lines. Annual facility energy is

E_facility = PUE × P_IT × t

and the difference between two facilities with the same IT load is

ΔE = (PUE_A − PUE_B) × P_IT × t

The assumptions, with units:

• P_IT = 1,000 kW, held constant for the full year
• t = 8,760 h, a non-leap year
• PUE values are annual averages, as ISO/IEC 30134-2 requires
• Both facilities host identical IT equipment drawing identical power

Substituted values:

• PUE 1.05: 1.05 × 1,000 kW × 8,760 h = 9,198 MWh
• PUE 1.4: 1.4 × 1,000 kW × 8,760 h = 12,264 MWh
• ΔE: 0.35 × 1,000 kW × 8,760 h = 3,066 MWh

The 3,066 MWh gap equals 25.0 percent of the PUE 1.4 facility’s consumption. Extending the comparison to PUE 1.8 more than doubles the gap, to 6,570 MWh, or 41.7 percent of that facility’s draw.

Source: calculation by this publication using the ISO/IEC 30134-2 definition; PUE 1.4 and PUE 1.8 are scenario values chosen to bracket the survey average discussed below

These are scenario results, not measurements. Because ΔE is linear in P_IT, the result scales with average IT power actually drawn. A hall with 10 MW installed that averages 6 MW of IT should multiply by 6, not by 10. Using nameplate capacity here overstates the saving by the full utilization gap.

Published PUE figures are not the same kind of number

The figures usually quoted for immersion and air cooling come from different kinds of evidence, and that difference matters more than the second decimal place.

Google reports a trailing twelve-month fleet PUE of 1.09 for 2025. The company states that the figure comes from continuous measurement across its large-scale campuses and includes all sources of overhead. That makes it a measured, annual, fleet-wide value.

Source: Google Data Centers, “Power usage effectiveness” (2025 fleet report)

Uptime Institute’s 2025 survey puts the global average at 1.54, and Uptime reports little change in average PUE for six consecutive years. The survey asks each operator for the operational PUE of its largest data center. Trade press reported that responses were collected from more than 800 owners and operators in April and May 2025. This is a self-reported value, not a metered audit.

Source: Uptime Institute press release (July 30, 2025); Uptime Intelligence, “Mapping PUE trends by data center region, age and size” (October 2025); Google Data Centers, citing the Uptime 2025 average; Data Center Knowledge (August 1, 2025)

The immersion evidence is thinner. A 2023 peer-reviewed review in Energy Informatics collected immersion PUE values of 1.02 to 1.04 from the studies it surveyed. Its highest reported immersion value, 1.17, came from an experiment run for maximum cooling capacity rather than efficiency.

The same review cites an Alibaba single-phase immersion deployment at PUE 1.07, a figure it takes from a 2019 Open Compute Project Summit presentation. A 2024 review in Science and Technology for Energy Transition places two-phase immersion at 1.05 to 1.07.

Source: Haghshenas, Setz, Blosch and Aiello, “Enough hot air: the role of immersion cooling,” Energy Informatics 6, 14 (2023); “Immersion cooling technology development status of data center,” Science and Technology for Energy Transition (2024)

Putting a literature immersion value next to a survey average compares a best-documented case with a typical one. PUE 1.05 is a defensible figure for immersion, but most of the evidence behind it is study-scale, and the one deployment figure here reaches this publication secondhand.

Table 1. Published and scenario PUE values converted to annual energyMWh per year, 1,000 kW constant IT load

* Literature, fleet, and survey values are different measurement types; not directly comparable

* 8,760 h; IT power held constant; server fan removal under immersion not included

Source: Haghshenas et al., Energy Informatics 6, 14 (2023); Science and Technology for Energy Transition (2024); Google Data Centers, “Power usage effectiveness” (2025); Uptime Institute Global Data Center Survey 2025, as cited by Google; energy values calculated by this publication

The dollar value depends on the tariff as much as on the cooling

EIA reported average U.S. revenue of 9.17 cents per kWh from industrial customers in June 2026, and 14.19 cents from commercial customers. Applied to the 3,066 MWh gap, those prices give $281,152 and $435,065 per year per megawatt of IT.

Source: U.S. Energy Information Administration, Electricity Monthly Update, data for June 2026 (released August 26, 2026)

EIA calculates these values as average revenue per kWh, which it uses as a proxy for price. One summer month is not an annual contract, and the average does not show demand charges or time-of-use structure. Each 1 cent per kWh changes the value of the PUE 1.05 versus 1.4 gap by $30,660 a year per megawatt, so a local tariff can be substituted directly.

Table 2. Annual gap against PUE 1.05 and its costPer 1,000 kW constant IT load, per year

* 9.17¢ industrial, 14.19¢ commercial: EIA average revenue per kWh, June 2026 only; demand charges excluded

* Each 1¢/kWh moves the PUE 1.4 row by $30,660; fan removal under immersion not included

Source: U.S. Energy Information Administration, Electricity Monthly Update, data for June 2026 (August 26, 2026); gaps and costs calculated by this publication

The table also shows where data center energy savings flatten out. Moving from Google’s reported 1.09 to 1.05 is worth 350.4 MWh per megawatt, 11.4 percent of the gap in the PUE 1.4 case. For an operator already near 1.1, the business case for immersion has to come from rack density rather than from the electricity bill. The density side of that argument is covered in why air cooling reaches a limit per rack.

Source: EIA, Electricity Monthly Update (June 2026 data); calculation by this publication

Conditions the calculation assumes away

The equal-IT-load assumption is the weakest of the four, and it works against immersion. Server fans count as IT load under PUE. Immersion removes them, so IT power itself falls. Haghshenas et al. cite The Green Grid’s estimate that eliminating internal fans cuts IT energy by 5 to 10 percent.

Source: Haghshenas et al., Energy Informatics (2023), citing The Green Grid White Paper 70 (2016)

With f as the fractional IT reduction, the immersion case becomes

E_immersion = 1.05 × 1,000 kW × (1 − f) × 8,760 h

At f = 0.05 that gives 8,738.1 MWh. At f = 0.10 it gives 8,278.2 MWh.

Against 12,264 MWh for the PUE 1.4 facility, the gap widens to between 3,525.9 and 3,985.8 MWh. PUE flatters air cooling on this point: the fan energy an air-cooled server spends sits in the denominator, where it lowers the ratio instead of registering as overhead.

Season moves the ratio even inside one fleet. Google’s fleet reported a quarterly PUE of 1.11 in Q3 2025 and 1.08 in Q4 2025, while the trailing twelve-month value held at 1.09.

Climate moves it further. In Q2 2026, Google reported a quarterly PUE of 1.28 for its Mesa, Arizona campus and 1.04 for its Lancaster, Ohio campus. Multiplying a single quarter’s PUE by 8,760 hours produces an annual figure the standard does not recognize.

Source: Google Data Centers, “Power usage effectiveness” (quarterly tables, 2025 and 2026)

None of the published figures above states the IT utilization at which it was recorded. Transferring any of them to a facility with a different average IT load requires engineering review.

A 0.35 PUE gap is a per-megawatt budget line, not a cooling verdict

The arithmetic supports one rule firmly. At equal IT load, each 0.01 of annual PUE is worth 87.6 MWh a year per megawatt of average IT power. The 1.05 versus 1.4 comparison is that rule applied 35 times, and the fan effect makes the true gap larger than 3,066 MWh.

What the numbers do not support is a claim that a given facility will reach 1.05. The evidence for that value is mostly study-scale, while the 1.09 and 1.54 figures come from a measured fleet and a large self-reported survey. Design decisions built on either end of the comparison need engineering review.

Before a PUE figure enters a cost model, check five things:

• Whether it is an annual value or a single quarter
• Which ISO/IEC 30134-2 measurement category it uses
• The average IT power behind it, not the installed capacity
• Whether server fans sit inside the IT figure
• The tariff structure, including demand charges

References

ISO/IEC 30134-2:2016, Information technology — Data centres — Key performance indicators — Part 2: Power usage effectiveness (PUE) / ISO/IEC / 2016 / https://docbox.etsi.org/stf/Archive/STF516_M462_EnergyEfficiency/STFworkarea/WG1/Documents/Foundation/ISO_IEC%2030134-2(2016).pdf ISO/IEC 30134-2:2026, catalog entry / ISO / 2026 / https://www.iso.org/standard/30134-2 Power usage effectiveness / Google Data Centers / 2025 fleet report, quarterly tables through Q2 2026 / https://datacenters.google/efficiency/ Uptime’s 15th Annual Global Data Center Survey Results (press release) / Uptime Institute / July 30, 2025 / https://uptimeinstitute.com/about-ui/press-releases/uptimes-15th-annual-global-data-center-survey-results-shows-both-commitment-and-hesitancy Mapping PUE trends by data center region, age and size / Uptime Intelligence / October 16, 2025 / https://intelligence.uptimeinstitute.com/resource/mapping-pue-trends-data-center-region-age-and-size Uptime Institute: Data Center Industry Faces Management Crisis Amid AI Transformation / Data Center Knowledge / August 1, 2025 / https://www.datacenterknowledge.com/energy-power-supply/uptime-institute-data-center-industry-faces-management-crisis-amid-ai-transformation Enough hot air: the role of immersion cooling (Haghshenas, Setz, Blosch, Aiello), Energy Informatics 6, 14 / Springer Nature / August 9, 2023 / https://link.springer.com/article/10.1186/s42162-023-00269-0 Immersion cooling technology development status of data center / Science and Technology for Energy Transition / 2024 / https://www.stet-review.org/articles/stet/full_html/2024/01/stet20240005/stet20240005.html Liquid Cooling Technology Update, White Paper 70 (Neudorfer, Ellsworth, Kulkarni, Zien) / The Green Grid / 2016 / Electricity Monthly Update, End Use: June 2026 / U.S. Energy Information Administration / August 26, 2026 / https://www.eia.gov/electricity/monthly/update/end-use.php