GPUSmith Article
PUE, WUE, ERE and CUE for AI Facilities: ISO/IEC 30134
How ISO/IEC 30134 and EU rules define PUE, WUE, ERF/ERE and CUE, where to meter them, and why liquid-cooled GPU racks distort PUE. Worked examples.
Summary
- 01Liquid cooling can reduce total facility energy while PUE barely changes, because server fans and cooling equipment affect the IT denominator. Report absolute facility energy alongside PUE.
- 02PUE does not credit exported heat. ERF measures energy reused outside the data center boundary; ERE credits reuse while retaining the IT denominator and should be reported beside PUE.
- 03WUE comparisons need a stated water basis: withdrawal or consumption, and site or source water. Dry cooling can reduce site water while increasing facility energy.
- 04CUE depends on the emission factor as well as PUE. State the carbon accounting basis and whether refrigerant leakage and generator testing are included.
- 05Comparable efficiency reporting requires a consistent boundary, measurement category, period and load factor, with CDU and RDHx energy identified separately.
Inside this article
- 01Why efficiency metrics break for AI racks
- 02The ISO/IEC 30134 family
- 03PUE: definition, categories and the 2026 edition
- 04Where PUE misleads with liquid cooling
- 05ERF and ERE: crediting exported heat
- 06WUE: which water, against which energy
- 07CUE: carbon per unit of IT energy
- 08Metering plan
- 09Benchmarks and comparability
- 10Reporting template for colocation and internal teams
- 11Using PUE in energy and cost models
- 12Sources
Why efficiency metrics break for AI racks
Every metric in this family is a ratio to IT energy or to total facility energy, so every one of them depends on where the line between "IT" and "facility" is drawn. Direct-to-chip liquid cooling moves load across that line. Server fans, which count as IT, shrink. Coolant distribution unit (CDU) pumps, rear-door heat exchanger (RDHx) fans and in-row coolers appear, and depending on which panel feeds them they land on either side.
The consequence: a liquid-cooling retrofit can cut total site energy substantially while PUE barely moves, and two halls with identical plants can report different PUEs because they meter at different points. Definitions below are cited to the standards and regulators that publish them; worked-example figures are hypothetical and labelled as such.
The ISO/IEC 30134 family
ISO/IEC 30134 is the international series of data center KPIs from ISO/IEC JTC 1/SC 39. It "does not specify limits or targets for any KPI" and does not aggregate KPIs into a combined score [1]. Status as listed on iso.org on 2026-10-28:
| Part | KPI | Edition | Status |
|---|---|---|---|
| 30134-2 | Power usage effectiveness (PUE) | 2, 2026-01 | Published; 2016 edition withdrawn (ISO, ISO 2016) |
| 30134-3 | Renewable energy factor (REF) | 2016 + amendment | Published [2] |
| 30134-6 | Energy reuse factor (ERF) | 1, 2021-08 | Stage 90.92, "to be revised" [3] |
| 30134-8 | Carbon usage effectiveness (CUE) | 1, 2022-03 | Published [4] |
| 30134-9 | Water usage effectiveness (WUE) | 1, 2022-03 | Stage 90.92, "to be revised" [5] |
Part 1 sets common requirements; Part 7 covers the cooling efficiency ratio (CER). Full texts are paywalled (30134-2:2026 lists at CHF 159), so the definitions quoted here come from the free informative sections on ISO's Online Browsing Platform. In the EU, the binding rules are in Delegated Regulation (EU) 2024/1364, which applies to sites with at least 500 kW installed IT power demand and references the parallel EN 50600-4-x standards "or equivalent".
PUE: definition, categories and the 2026 edition
ISO/IEC 30134-2:2026 defines PUE as the "calculated ratio of the data centre total energy consumption, within a 12-month continuous period, to information technology equipment energy consumption measured across the same period" [1]. The standard notes the KPI "was originally developed by The Green Grid", whose glossary describes it as "the energy entering the facility divided by the energy used by the IT equipment" [6].
PUE = E_DC / E_IT, both in kWh over the same 12 months; the EU regulation uses the same expression [7]. Heat recovered inside the facility already lowers E_DC, so PUE cannot fall below 1.0 by definition; exported heat needs a separate metric [8].
The 2026 edition's foreword lists the changes from 2016: updated boundary measurement requirements; on-site generation measured in kWh; changed 12-month recording and watt-meter rules; a new requirement for electricity used to export non-electric energy outside the boundary; a new requirement for when E_DC or E_IT cannot be measured as specified; a mixed-use building derivative (mPUE); and more designed-PUE (dPUE) examples [1]. Other derivatives are iPUE (interim, any period other than 12 continuous months) and pPUE (partial, a defined sub-boundary).
Measurement categories
The 2026 text keeps three categories: Category 1 "basic resolution", Category 2 "intermediate" and Category 3 "advanced" [9]. Meter locations are not in the free preview. The EU regulation ties Category 1 to UPS output and Category 2 to the PDU [7], matching the 2011 multi-agency task force definitions [10]:
| Category | IT meter location | Still counted as "IT" |
|---|---|---|
| 1 | UPS output, 12-month kWh | PDU and transformer losses; rack-mounted non-IT devices such as fans |
| 2 | PDU output | Rack-mounted non-IT devices such as fans |
| 3 | IT equipment input (receptacle) | Only the device, including its internal fans and PSU |
The task force adds that non-IT equipment fed from the UPS or PDU, such as CRAH units or in-row coolers, "must be metered and subtracted" from the IT reading. GPU retrofits miss this often, because CDUs and RDHx fans get fed from the nearest circuit.
The EU rule measures E_DC before the supply transfer switchgear, with back-up generator energy reported separately, and E_IT at Category 1 unless there is no UPS. Category 1 and Category 3 PUEs for the same building differ by construction: PDU losses leave the IT denominator but stay in total energy, so Category 3 reads higher.
Where PUE misleads with liquid cooling
In a Vertiv and NVIDIA study moving a facility from 100% air to 75% direct-to-chip cooling, total data center power fell 10.2% but PUE fell only 3.3%. Server fan power dropped 80%, and because fans sit in the IT denominator, IT power fell 7%. Between two intermediate studies, total power fell 1.8% while PUE stayed at 1.35 [11]. Removing fan kWh shrinks numerator and denominator equally, which pushes PUE up while total energy falls. CDU pumps do the reverse: fed from an IT PDU and not subtracted, they inflate E_IT and flatter PUE.
Worked example (hypothetical)
An air-cooled GPU hall is retrofitted to direct-to-chip. Values are average kW; compute load (GPUs, CPUs, memory, NICs, PSU losses) is held at 920 kW.
| kW | A: air | B1: liquid, CDU on IT PDU | B2: liquid, CDU sub-metered as facility |
|---|---|---|---|
| Compute | 920 | 920 | 920 |
| Server fans | 80 | 20 | 20 |
| CDU pumps | 0 | 15 (in IT) | 15 (in facility) |
| E_IT as metered | 1,000 | 955 | 940 |
| Facility overhead incl. CDU where applicable | 350 | 280 | 295 |
| E_DC | 1,350 | 1,235 | 1,235 |
| PUE | 1,350/1,000 = 1.350 | 1,235/955 = 1.293 | 1,235/940 = 1.314 |
| E_DC / compute | 1.467 | 1.342 | 1.342 |
Total energy falls (1,350 - 1,235)/1,350 = 8.5%; PUE falls 2.7% (B2) or 4.2% (B1), and the 0.021 gap between B1 and B2 is purely a metering choice. The last row is a simplified total-power usage effectiveness (TUE), proposed by Patterson et al. in 2013: total energy divided by energy delivered to computational components, equal to ITUE x PUE [12]. Full TUE also removes PSU and voltage-regulator losses, which this table does not.
Rules for GPU halls:
- Sub-meter CDUs, RDHx fans and in-row coolers and classify them as facility load.
- Publish PUE with its category, plus absolute E_DC in MWh. Falling E_DC with flat PUE is the expected signature of a good liquid retrofit.
- Track server fan power from BMC telemetry; it is the term that hides the benefit.
- Do not rank air-cooled against liquid-cooled halls on PUE; the Vertiv authors say it "shouldn't be used to compare the efficiency of liquid and air-cooling systems".
Plant design and commissioning are covered in GPU data center cooling design and the cold plate commissioning checklist.
A PUE commitment without category and boundary can be met by moving a meter.
ERF and ERE: crediting exported heat
Warm return water from liquid-cooled racks makes heat export more practical. Two metrics credit it.
ERF (ISO/IEC 30134-6:2021) is "the ratio of energy being reused divided by the sum of all energy consumed in a data centre" [3], counting only use "for an alternate purpose outside the data centre boundary" [1]. The EU uses ERF = E_REUSE / E_DC, measured at the hand-off to the other party, minus any reused heat that cools the data center itself [7].
ERE is The Green Grid's metric (white paper #29, member-gated); NREL's 2011 report, citing it, keeps the IT denominator: ERE = (Total energy - Reuse energy) / IT energy, which can fall below 1.0. NREL's Research Support Facility data center reported a PUE of 1.16 and an approximate ERE of 0.91 [8].
Substituting E_REUSE = ERF x E_DC (our algebra): ERE = (1 - ERF) x PUE.
Scenario B2 with a hypothetical 2,000 MWh/year heat-export contract:
- E_DC = 1,235 kW x 8,760 h = 10,818.6 MWh; E_IT = 940 x 8,760 = 8,234.4 MWh
- ERF = 2,000 / 10,818.6 = 0.185
- ERE = (10,818.6 - 2,000) / 8,234.4 = 1.071; check: (1 - 0.185) x 1.314 = 1.071
Report ERF for ISO and EU purposes; if you quote ERE, quote PUE beside it, since ERE alone hides whether the facility got more efficient.
WUE: which water, against which energy
ISO/IEC 30134-9:2022 defines WUE as the "ratio of the data centre water consumption divided by the energy consumed by IT equipment", with water in m3 and energy in MWh [1]. m3/MWh equals L/kWh numerically (1,000 L per 1,000 kWh), the unit most operators publish. The Green Grid introduced WUE in March 2011 for water "used on-site in operating a data center, including water used for humidification and water evaporated on-site for energy production or cooling" [13].
ISO's symbols separate water input, water returned, water usage and reused water, and define an energy water intensity factor (EWIF, m3/MWh) for water used to produce energy. That is the site-versus-source split: site water crosses your meter; source water was consumed by generators making your electricity. The combining formulas are in the paid clauses and were not reviewed.
The EU rule is simpler: WUE = W_IN / E_IT, where W_IN is all water entering the boundary for data center functions (EN 50600-4-9 Category 2, else Category 1), E_IT in MWh, with potable input reported separately [7]. W_IN is an input quantity, not net consumption.
Published figures mix bases. Microsoft reported a fiscal-2024 global average "withdrawal WUE of 0.30 L/kWh" and says its closed-loop chip-level design will avoid "more than 125 million liters of water per year per datacenter", while replacing evaporative systems with mechanical cooling "will increase our power usage effectiveness (PUE)" [14]. Lower site water for higher energy is the central WUE trade.
Worked example (hypothetical): evaporative versus dry heat rejection
Scenario B2, E_IT = 8,234.4 MWh/year:
| Evaporative | Dry coolers | |
|---|---|---|
| Site W_IN | 12,000 m3 | 500 m3 (domestic) |
| E_DC | 10,818.6 MWh | 1,275 kW x 8,760 = 11,169.0 MWh |
| Site WUE | 12,000 / 8,234.4 = 1.457 L/kWh | 500 / 8,234.4 = 0.061 L/kWh |
| Source water at EWIF f | 10,818.6 f | 11,169.0 f |
Combined water is equal when 12,000 + 10,818.6 f = 500 + 11,169.0 f, so 11,500 = 350.4 f and f = 32.8 m3/MWh. Below that grid water intensity, dry cooling uses less total water in this example. The break-even moves with the dry-cooling energy penalty and the local grid, so a site WUE without PUE and EWIF beside it is incomplete. For volumes, see how much water data centers use.
- Site water input: 12,000 m3.
- Site WUE: 1.457 L/kWh.
- Facility energy: 10,818.6 MWh.
- Site water input: 500 m3, domestic.
- Site WUE: 0.061 L/kWh.
- Facility energy: 11,169.0 MWh.
Both use the same IT energy. Below a grid water intensity of 32.8 m3/MWh, dry cooling uses less total water in this hypothetical example. The break-even depends on the energy penalty and local grid.
CUE: carbon per unit of IT energy
ISO/IEC 30134-8:2022 defines CUE as the "ratio of the data centre annual CO2 emissions and IT equipment energy demand", emissions in kg and energy in kWh, so kg CO2e/kWh [1]. Its symbol list includes emissions from on-site generation, generator testing and refrigerant leakage, and it considers seven greenhouse gases including HFCs and SF6.
For an all-electric site with grid factor f_e and no other emissions, CUE = f_e x E_DC / E_IT = f_e x PUE (our algebra). At a hypothetical 0.35 kg CO2e/kWh in scenario B2: 0.35 x 10,818,600 kWh = 3,786,510 kg; CUE = 3,786,510 / 8,234,400 = 0.460 kg CO2e/kWh (= 0.35 x 1.314). Refrigerant leakage and generator tests then add to the numerator independently of PUE.
The emission factor dominates. The GHG Protocol Scope 2 Guidance requires companies operating in markets with contractual instruments to report scope 2 "in two ways", location-based and market-based [15]; a revision consultation ran from 2025-10-20 to 2026-01-31 [16]. Whether 30134-8 prescribes either method is in the paid text and was not verified, so state the basis next to every CUE. The EU regulation does not use CUE; it reports REF = E_RES-TOT / E_DC, summing guarantees of origin, PPAs and on-site renewables.
Metering plan
| Quantity | Meter point | Notes for liquid-cooled halls |
|---|---|---|
| E_DC | Every supply before transfer switchgear, plus generators | Generator kWh separately |
| E_IT (Cat. 1) | Every UPS output feeding IT | Subtract CDUs, RDHx, in-row units on UPS circuits |
| E_IT (Cat. 2/3) | PDU outputs or rack PDU receptacles | Enables per-hall pPUE |
| CDU / RDHx | Dedicated sub-meters | Facility load |
| Thermal energy | Flow and delta-T on secondary and heat-export loops | Needed for pPUE and ERF |
| E_REUSE | Hand-off to heat customer | Net of heat reused on site |
| W_IN, potable | Every water supply at the boundary | Separate cooling make-up from domestic |
| Carbon inputs | Bills, contracts, refrigerant logs, generator hours | Location- and market-based |
The EU regulation also requires annual average waste heat temperature, measured where heated fluid enters the heat exchangers at the computer-room boundary [7]; in direct-to-chip halls, plan that sensor at commissioning. A site in operation for less than a year reports "only for the period the data centre has been in operation" (Art. 3), which makes the value an interim figure; label it as one. A new GPU hall at 30% load will show a poor iPUE because fixed overhead is spread over little IT energy.
Benchmarks and comparability
Uptime Institute's 2025 survey reports "average PUE levels show little change for the sixth consecutive year", few facilities above 30 kW per rack, and no improvement in sustainability-metric reporting [17]. Survey averages mix categories, climates and load factors; they are context, not a target.
Before comparing two PUEs, check category, a full 12-month basis, treatment of CDU and RDHx loads, and load factor. For WUE, check withdrawal versus consumption and whether source water is included. For CUE, check the emission-factor method and refrigerant treatment.
Europe is forcing a common basis. Regulation 2024/1364 required reports by 15 September 2024, then by 15 May each year from 2025, covering PUE, WUE, ERF and REF. On 21 September 2026 the Commission adopted a delegated regulation, C(2026) 3472, creating a mandatory A-to-G label graded on PUE and WUE for sites of at least 500 kW, subject to a two-month (extendable) scrutiny period, with first labels expected in 2027 [18][19]. As of 2026-10-28 it was still under scrutiny and not yet in the Official Journal; take grade thresholds from the published text.
- 01Align the PUE basis
Check category, a full annual period and load factor before comparing PUEs.
- 02Separate cooling loads
Sub-meter CDUs, RDHx fans and in-row coolers and classify them as facility load.
- 03Identify the water basis
Check withdrawal versus consumption and whether source water is included.
- 04Identify the carbon basis
Check the emission-factor method and refrigerant treatment before comparing CUE.
Reporting template for colocation and internal teams
| Field | Content |
|---|---|
| Boundary | Diagram; mixed-use status (mPUE) |
| PUE | Annual value, category, period dates |
| E_DC, E_IT | MWh, generator MWh separately |
| Cooling on IT circuits | CDUs, RDHx, in-row units and whether subtracted |
| pPUE | For your hall or suite, if sub-metered |
| WUE | L/kWh, withdrawal or consumption, potable share, EWIF if source water included |
| ERF | E_REUSE MWh and hand-off point, or none |
| CUE | Location- and market-based, refrigerant in or out |
| Gaps | Estimated or unmetered quantities |
Example contract clause: "Provider shall report annual PUE per ISO/IEC 30134-2:2026 (state category), with coolant distribution and rear-door heat exchanger energy metered separately and excluded from IT energy, and shall provide monthly E_DC and E_IT in kWh." A PUE commitment without category and boundary can be met by moving a meter.
Using PUE in energy and cost models
Facility MWh = IT MWh x PUE only holds when both share a period and a boundary; the GPU cluster energy calculator follows that structure. Three cautions:
- Boundary. IT energy built from measured node power is effectively Category 3; multiplying it by a Category 1 PUE understates PDU losses. Ask for the matching category or a pPUE.
- Load. Use a PUE measured or designed at the load you will run, not at full design load.
- Cooling changes. After a liquid retrofit, model E_DC from the new plant, fan and pump loads, as in the worked example, rather than applying the old PUE to an IT figure that no longer contains the same fans.
GPU Smith works on these boundary and metering questions in private GPU facility planning.
Sources
- ISO, ISO/IEC 30134-2:2026 (PUE) and OBP informative sections, accessed 2026-10-28.
- ISO, ISO/IEC 30134-2:2016 (withdrawn), accessed 2026-10-28.
- VDE Verlag, ISO/IEC 30134-2:2026 preview, accessed 2026-10-28.
- ISO, ISO/IEC 30134-1/-2/-3 package, accessed 2026-10-28.
- ISO, ISO/IEC 30134-6:2021 (ERF) and OBP, accessed 2026-10-28.
- ISO, ISO/IEC 30134-8:2022 (CUE) and OBP, accessed 2026-10-28.
- ISO, ISO/IEC 30134-9:2022 (WUE) and OBP, accessed 2026-10-28.
- European Commission, Delegated Regulation (EU) 2024/1364, 14 March 2024.
- Linklaters, EU Commission adopts data centre sustainability rating scheme, 2026-10-14; European Commission, Delegated Regulation C(2026) 3472 publication page, accessed 2026-10-28.
- Data Center Metrics Task Force, Recommendations for Measuring and Reporting Overall Data Center Efficiency, Version 2, 2011-05-17.
- The Green Grid, Legacy glossary: PUE, accessed 2026-10-28.
- Sheppy et al., NREL, Reducing Data Center Loads for a Large-Scale, Low-Energy Office Building, December 2011.
- Patterson et al., TUE, a new energy-efficiency metric applied at ORNL's Jaguar, ISC 2013.
- Vertiv, Understanding the limitations of PUE in evaluating liquid cooling efficiency, 2023-03-03.
- Trellis, Green Grid rolls out data center metrics for water, energy reuse, 2011-03-08.
- Microsoft, Next-generation datacenters consume zero water for cooling, 2024-12-09.
- GHG Protocol, Scope 2 Guidance (2015) and Scope 2 page, accessed 2026-10-28.
- Uptime Institute, Global Data Center Survey 2025, 2025-07-29.
External Sources (19)
About
GPUSmith
Plan and build private AI compute with GPU Smith. We connect workload requirements with GPU hardware, networking, deployment and operating decisions so your infrastructure fits the work it must perform.
GPU Smith provides independent engineering and research for private AI infrastructure. We help technical buyers and operators reason about GPU workload sizing, hardware selection, procurement, deployment and inference operations, from the compute system to the networking, power and cooling requirements around it.
Start with the workload
Useful infrastructure decisions begin with the models, data, throughput, latency and operating constraints a team actually has. GPU Smith helps connect those requirements with system design and deployment choices. The goal is a privately controlled AI environment whose capacity and operational demands are understood before a purchasing decision.
Hardware and supplier research
Our public reference library covers GPUs, complete systems and networking components. The server-vendor directory and manufacturing research help teams investigate suppliers, compare options and follow sources behind technical claims. We distinguish manufacturer specifications from measured benchmarks and advertised capabilities from tested configurations. Reference pages are research resources, not a live inventory listing or a binding equipment quote.
Deployment and operations
GPU Smith's engineering scope includes infrastructure integration, networking, power, cooling and the practical operation of inference workloads. Our published articles explain the tradeoffs behind deployment, procurement and ongoing operations so teams can ask better questions and document decisions.
Work with GPU Smith
Explore the hardware library, GPU references, networking references and vendor research. Contact GPU Smith with your workloads and deployment constraints to discuss a project and confirm current scope, pricing and availability.
Inclusion of a manufacturer or product in our research does not imply a partnership, certification or endorsement.
Disclaimer
This document is provided for informational purposes only. No representations or warranties are made regarding the accuracy, completeness, or reliability of its contents. Any use of this information is at your own risk. GPUSmith shall not be liable for any damages arising from the use of this document. This content was generated with assistance from artificial intelligence tools, which may contain errors or inaccuracies. Readers should verify critical information independently. All product names, trademarks, and registered trademarks mentioned are property of their respective owners and are used for identification purposes only. Use of these names does not imply endorsement. This document does not constitute professional or legal advice. For specific guidance related to your needs, please consult qualified professionals.