PUE Measures the Building and Misses the Server Entirely

2026년 9월 11일

ISO/IEC 30134-2 defines power usage effectiveness as total data center energy divided by IT equipment energy. Both terms are meter readings, and the second counts every watt the servers draw, including the watts their fans waste.

ASHRAE TC9.9 puts fan power at 10% to 20% of server power in denser machines, up from as low as 2% in the previous decade. A facility that absorbs that increase reports a lower PUE and a larger electricity bill at once.

PUE is one ratio between two meter readings

Everything contestable about PUE sits in where the two meters are placed. The standard gives it as total data center energy consumption divided by IT equipment energy consumption, both recorded over a coincident period of 12 months, which makes the result always greater than 1. The numerator is read at the utility service entrance and covers power delivery, the cooling system, lighting, security and fire detection. The denominator covers storage, processing and transport equipment.

Three measurement categories are defined by where that IT reading is taken: Category 1 at the UPS output, Category 2 at the PDU output, Category 3 at the IT equipment input. Reporting requires the category as a subscript, so a Category 2 result is written PUE2.

Source: ISO/IEC 30134-2:2016, Clauses 5, 6.1.2, 6.2 and 7.1.1 (first edition, withdrawn; second edition dated 2026-01)

The subscript is the first thing dropped in practice, and the categories do not share a denominator. Category 1 includes the PDU transformer and static switch losses Category 2 excludes, so an operator reporting at the UPS output publishes a lower number than the same hall measured at the PDU output.

The exclusions are printed in the standard, not inferred from it

PUE’s defining document states its own blind spots in a list. Clause 4 says PUE does not take into account the energy efficiency of the IT load, its utilization or its productivity, the efficiency of on-site generation, resources such as space and water, or the reuse of waste by-products such as heat. The same clause says PUE is neither a productivity metric nor a standalone resource efficiency metric, and that it should not be used to compare different data centers.

Source: ISO/IEC 30134-2:2016, Clause 4

Among operators collecting metrics for corporate sustainability purposes in 2025, 74% collect PUE and 18% collect server utilization (n=617).

Source: Uptime Institute, Global Data Center Survey 2025 (July 2025)

The metric that cannot see IT-side waste is tracked by three operators in four; the one that would catch it, by fewer than one in five. Comparing two facilities is the single use the standard rules out, and it is the use PUE is put to in procurement decks.

Server fan power counts as IT load, so waste there pulls PUE down

Fan power inside the server has moved in both directions. ASHRAE TC9.9 records that constant-speed fans once wasted 20% of total server power, and that variable-speed control with tuned algorithms brought that below 2% in many cases during the multicore period.

It is climbing again. The committee reports 10% to 20% as not uncommon for some of the denser servers.

The committee frames the consequence as a capacity problem rather than an efficiency one. Server cooling fans run on the same UPS source as the server, so retaining air cooled IT equipment while fan power rises from 2% to 10% of total server power equates to reducing the data center UPS capacity by 8%.

Source: ASHRAE TC9.9, “Emergence and Expansion of Liquid Cooling in Mainstream Data Centers” (2021)

Fan power is a cooling cost carrying an IT badge, and the badge is what the meter reads. Every watt a fan wastes enlarges the denominator, which is the direction that flatters the ratio.

The arithmetic gives the same servers more power at the same PUE

Run the fan power shift through the governing equation and the ratio moves the wrong way.

PUE = E_DC / E_IT

Assumptions
  compute power (server power excluding fans)   980 kW, held constant
  workload                                      unchanged
  measurement period                            identical for both meters,
                                                so the energy ratio equals
                                                the steady-state power ratio
  facility support energy, base case            500 kW

Case A   fans at 2% of server power
  fan power  = 0.02 x 1,000            =    20 kW
  E_IT       = 980 + 20                = 1,000 kW
  E_DC       = 1,000 + 500             = 1,500 kW
  PUE        = 1,500 / 1,000           = 1.50

Case B   fans at 10% of server power
  P_f        = 0.10 x (980 + P_f)      =   108.9 kW
  E_IT       = 980 + 108.9             = 1,088.9 kW

  B1  facility support held at 500 kW
      E_DC   = 1,088.9 + 500           = 1,588.9 kW
      PUE    = 1,588.9 / 1,088.9       = 1.46

  B2  facility support scaled at the base ratio of 0.5
      E_DC   = 1,088.9 + 544.4         = 1,633.3 kW
      PUE    = 1,633.3 / 1,088.9       = 1.50

Table 1. Server fan power share against reported PUEkW at steady state

Case	IT energy	Facility support	Total	PUE
A. Fans at 2% of server power	1,000.0	500.0	1,500.0	1.50
B1. Fans at 10%, facility support unchanged	1,088.9	500.0	1,588.9	1.46
B2. Fans at 10%, facility support scaled at 0.5	1,088.9	544.4	1,633.3	1.50

* Constructed case, not a measurement; compute power held at 980 kW in all three rows

Source: ISO/IEC 30134-2:2016, Formula (1); fan power shares from ASHRAE TC9.9, “Emergence and Expansion of Liquid Cooling in Mainstream Data Centers” (2021)

Source: formula and boundary from ISO/IEC 30134-2:2016, Formula (1); fan power shares from ASHRAE TC9.9 (2021)

The result is a constructed case, not a measurement and not an upper bound. The fan shares are ASHRAE’s published values; the 500 kW of facility support puts the base case near the industry average, and a different assumption moves the three rows together without changing their order.

Case B1 is the one to read twice. The ratio improves by four hundredths while the hall draws 89 kW more for the same computing work, and a dashboard tracking only PUE reports that as a gain.

Where the coolant distribution unit sits moves the number more than how hard its pumps run

Liquid cooling components split the loop at a point PUE does not recognize. The Open Compute Project’s requirements document names the parts and fixes the boundary between them: cold plates on the components needing cooling, a rack manifold distributing flow inside the rack, quick disconnect couplings for service, and a coolant distribution unit that isolates the technology cooling system from the facility water system. The CDU can be in-rack, row level or facility level.

Source: Open Compute Project, “ACS Liquid Cooling Cold Plate Requirements Document, Revision 1.0” (October 9, 2019)

An in-rack CDU draws its power inside the rack. Under Category 1 or Category 2 the IT reading is taken at a UPS or PDU output upstream of that rack, so the pump energy falls in the denominator. The same unit at row level on a facility panel falls in the numerator. Category 3 anticipates part of this by measuring at the IT equipment and excluding non-IT loads; Categories 1 and 2 carry no such instruction.

The same document lists CDU power, cooling subsystem fan and pump power per rack, and total power into the compute components as three separate comparison metrics. Hardware documentation already makes the split PUE declines to make. It also puts heat capture by CPU and GPU cold plates at 70% to 75% in hybrid basic equipment, which leaves fans running inside a liquid cooled rack.

A flat industry average is what a boundary metric produces

The headline number has not moved in six years. Uptime Institute reports a weighted average annual PUE of 1.54 for 2025 (n=681), the sixth consecutive year of virtually no change.

Underneath, facilities commissioned within five years averaged 1.48, and those of 20 MW and above averaged 1.44.

The power drawn inside the racks did move over the same stretch. Average modal rack density excluding high-density outliers rose from 6.8 kW in 2024 to 7.5 kW in 2025.

Source: Uptime Institute, Global Data Center Survey 2025 (July 2025)

Table 2. Which side of the boundary each inefficiency falls ondirection of effect

Inefficiency	Where it is metered	Effect on reported PUE
Chiller, tower and pump energy in the facility cooling system	Numerator only	Raises
Supply air temperature set lower than the equipment requires	Numerator only	Raises
Row level or facility level CDU pump power	Numerator only	Raises
Server fan power	Denominator, as IT equipment	Lowers
In-rack CDU pump power under Category 1 or 2	Denominator, as IT equipment	Lowers
Idle or lightly used servers	Denominator, as IT equipment	Lowers
Older, less efficient processors doing the same work	Denominator, as IT equipment	Lowers

* Numerator is total data center energy, which already contains the denominator

Source: ISO/IEC 30134-2:2016, Clauses 4 and 6.2; Open Compute Project, “ACS Liquid Cooling Cold Plate Requirements Document, Revision 1.0” (2019)

PUE tracked the one-time facility wins, containment, set points and economizer hours, then went quiet when those ran out. It has no term for the change actually underway, which is more power drawn per rack for work the ratio never inspects.

Caveats on the sources

The Uptime figures are self-reported for each respondent’s largest facility with no measurement category attached, so they cannot be reconciled with the reporting construct the standard requires. The ASHRAE fan power percentages are a range across unspecified dense servers, not a measurement of one product, and the paper gives no inlet temperature or configuration for either end. The OCP document calls itself a requirements document rather than a specification, and Revision 1.0 predates today’s accelerator racks. The ISO/IEC 30134-2 text cited here is the withdrawn 2016 first edition; the 2026 second edition was not read, and its clause numbering may differ.

Five questions to put to any PUE figure

Ask for the measurement category before accepting any PUE figure. Category 1 and Category 3 values from the same hall are different numbers.

Ask for total energy and IT energy in kWh alongside the ratio. Any operator who calculated PUE holds both.

Ask where the coolant distribution unit takes its power, and whether that circuit sits upstream or downstream of the IT meter.

Ask whether server fan power is separately instrumented. If it is not, no PUE from that site separates compute from air movement inside the chassis.

Treat any cross-site PUE comparison as invalid until boundaries, categories, climate and IT mix all match.

References

ISO/IEC 30134-2:2016, Information technology — Data centres — Key performance indicators — Part 2: Power usage effectiveness (PUE) / ISO and IEC / 2016-04-15 / https://cdn.standards.iteh.ai/samples/63451/87c396de9e76423088236bf87cdf73c1/ISO-IEC-30134-2-2016.pdf
ISO catalog record for ISO/IEC 30134-2:2016, status withdrawn, superseded by ISO/IEC 30134-2:2026 / ISO / 2026 / https://www.iso.org/standard/63451.html
Emergence and Expansion of Liquid Cooling in Mainstream Data Centers / ASHRAE Technical Committee 9.9 / 2021 / https://www.vertiv.com/49db1c/globalassets/documents/white-papers/ashrae_tc0909_emergence_and_expansion_of_liquid_cooling_in_mainstream_data_centers_5_may_2021_332771_0.pdf
ACS Liquid Cooling Cold Plate Requirements Document, Revision 1.0 / Open Compute Project Foundation / 2019-10-09 / https://www.opencompute.org/documents/ocp-acs-liquid-cooling-cold-plate-requirements-pdf
Uptime Institute Global Data Center Survey 2025 / Uptime Institute Intelligence / 2025-07 / https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2025.Annual.Survey.Report.pdf

Heonsu Nam

Heonsu Nam runs Data Center Technology Lab. His day job is running a small career-education business in South Korea; this site is where he studies data center cooling in public. He is not a mechanical engineer, so every article starts from the primary documents (ASHRAE, Open Compute Project, vendor data sheets) and shows the arithmetic so readers can check it. Spot an error? Please use the Contact page.