Efficiency Metrics: PUE, WUE, ERF, REF & the Post-PUE Metric Stack

Power Usage Effectiveness is the most quoted number in the industry, and the most broken, in ways that rarely show. It was a brilliant intervention in 2007: a single ratio, total facility energy divided by IT energy, that let an operator see for the first time how many watts went to cooling, conversion and lighting for every watt that reached a server. It drove a fifteen-year efficiency campaign that took hyperscale halls from a PUE near 2.0 to design points around 1.1. But PUE was designed for an air-cooled, conversion-heavy, fixed-IT world, and the AI data center is none of those. Its denominator, "IT energy," is no longer a stable line. Liquid cooling moves fans and pumps across it, and densification makes it swing with utilization. The questions that now gate a project (how much water, how much carbon, how much useful work) sit entirely outside its formula.

This chapter is the canonical home for the efficiency metric stack referenced from Chapter 0.3 and Chapter 5.1. Each metric makes one thing visible and hides another. The chapter defines PUE and shows where it breaks for AI, separates WUE-site from WUE-source and the embedded-water trap between them, and walks ERF/ERE, REF and CUE as the carbon-and-reuse layer. It then adds the work-based metrics (TUE/ITUE, DCeP, tokens-per-joule) that put real output in the denominator, and ends with a measurement plan that survives an audit instead of producing a vanity number for a press release.

Under ISO/IEC 30134-2:2026, PUE is total facility energy divided by IT equipment energy, measured over a full year. An instantaneous or design-point PUE is a weaker claim, so always ask which one a vendor is quoting. A PUE of 1.20 means that for every 1.0 kWh delivered to IT, another 0.20 kWh went to everything else: cooling plant, UPS and transformer losses, switchgear and lighting.

The two 2026 Uptime survey values below weight facilities differently. The respondent mean counts each facility equally, while the capacity-weighted result gives larger facilities proportional weight. Uptime’s August 6, 2026 analysis shows why they differ: efficient large facilities pull the weighted result down, while older facilities still dilute the per-facility mean. Benchmark against the matching population and weighting, then test your project’s own climate and load curve, because neither survey result predicts this facility’s PUE. The 2025 survey average appears below only as history, in the caveat to the respondent-mean key number.

Three structural problems make PUE misleading for AI facilities, and each is a fork an operator must decide how to handle.

1. The liquid-cooling boundary problem. PUE's denominator is the power that crosses into the IT. In an air-cooled hall the line is clean: server fans sit inside it and CRAH fans outside. Direct-to-chip liquid blurs it. Cold plates, in-rack manifolds and coolant pumps can be metered as IT or as facility, depending on where the CDU sits and how the rack is wired. Move a fan's power from the facility side to the IT side and PUE improves while nothing physical got more efficient; you only re-labeled a loss. The error also runs the other way. A genuinely efficient warm-water loop with a big chiller plant can post a higher PUE than a cheaper, hotter air hall, because the liquid plant's parasitic load is honestly accounted while the air hall's server-fan burn hides inside IT. PUE was never meant to compare cooling architectures, and using it that way is the most common metric error in 2026 procurement.

2. The utilization problem. PUE says nothing about whether the IT is doing useful work. A cluster idling at 15% may still report an acceptable PUE while producing little, but low load does not mechanically improve the ratio. With PUE = 1 + overhead/IT, plant losses that are fixed or fall less than proportionally usually make PUE worse as IT load falls, and part-load UPS or chiller efficiency may worsen it further. So publish the load curve and a work-based denominator beside it. Operators who manage on goodput, the useful work delivered within the quality and service limits set for the job (see Chapter 12.2), do not rely on PUE alone for this reason.

3. The everything-else-is-invisible problem. PUE is an energy-overhead ratio. It is silent on water (a 1.10-PUE evaporative hall can drink millions of liters a day), silent on carbon (1.10 PUE on coal is dirtier than 1.4 PUE on hydro), and silent on whether heat is wasted or reused. The metrics below exist to make those silences audible.

Water Usage Effectiveness is liters of water per kWh of IT energy, and which liters is the whole argument. Keep three quantities separate: input (everything crossing the boundary), discharge (blowdown returned to a sewer or watercourse) and consumption (input minus discharge and net storage increase, with evaporation dominant in a tower balance). Evaporate 100 m³ at six cycles of concentration and you draw ~120 m³ of makeup and discharge ~20 m³. Reporting one number for all three quantities hides the difference that the regulator and the basin each care about.

ISO/IEC 30134-9 frames WUE around water used on site. The EU filing differs. Under Delegated Regulation (EU) 2024/1364 as it stands, the numerator is the total water input to the site for every data-center function, metered to the EN 50600-4-9 WUE Category 2 method (Category 1 where Category 2 is not feasible), and the denominator is IT energy. The rating-scheme amendment adopted in September 2026 changes this once it is in force: input is metered to WUE Category 1, and the graded WUE counts freshwater only. If you are placing meters now, separate freshwater from reclaimed, rain and sea water. WUE has moved faster than any other metric from obscure to board-level, because water is now a siting gate (Chapter 3.7) and a social-license risk, not just an operating cost. The fork that matters is WUE-site versus WUE-source, and operators routinely report the flattering one.

WUE-site counts only the water crossing the property boundary: evaporative tower make-up, adiabatic pre-cooling and humidification. A site that meters cooling make-up alone will under-report against the EU input boundary, which also captures water used for power, security and IT functions. LBNL’s December 2024 report modeled US annual-average site WUE at ~0.36 L/kWh in 2023 across all cooling types, with on-site water consumption in the numerator; designs with dry, non-evaporative heat rejection approach zero. Microsoft’s FY2025 cooling-and-humidification WUE is ~0.27 L/kWh for fully owned and controlled sites operating the full fiscal year. Its December 2024 non-evaporative design announcement separately estimates more than 125 million liters of avoided cooling-water consumption per facility per year. Microsoft's first Mount Pleasant, Wisconsin facility has been fully operational since June 2026. Microsoft says more than 90% of its capacity uses closed-loop liquid cooling, and the air-cooled remainder still uses water on the hottest days. No metered WUE for the site has been published. WUE-site lets an arid, power-rich site become viable by designing water out, and it lets a thirsty evaporative hall look responsible if you stop counting at the fence.

WUE-source adds the water embedded in the electricity the facility consumes: the water evaporated at the power plant to make the megawatts. This is where the trap springs. The US-average energy water-intensity factor (EWIF) is roughly 5.1 L/kWh. It ranges from near zero (wind, solar, gas combined-cycle on dry cooling) to roughly 1-3 L/kWh consumed at evaporative-cooled thermoelectric plants, and to tens of L/kWh where reservoir evaporation is allocated to hydro generation, which is what carries the national mean above the thermoelectric range. Once-through plants withdraw tens of L/kWh but consume ~1 L/kWh or less. The 2022 modeled US EWIF in the key numbers below is historical context, not a current site factor, and multiplying it by an unrelated fleet PUE does not establish a typical campus footprint. Instead, multiply each interval’s purchased electricity by a region- and generation-specific consumption factor, sum those liters, and divide by the same IT-energy total used for the direct-water record. A hall using dry, non-evaporative heat rejection can still carry a large WUE-source if it draws coal- or nuclear-heavy thermoelectric power. A site that goes water-free on cooling but sits on a thirsty thermal grid has not eliminated its water footprint; it has relocated it upstream and out of its own report. The water-footprint convention borrowed from greenhouse-gas accounting, which Chapter 15.4 also uses, names the same split: on-site water is scope 1, the water consumed generating purchased electricity is scope 2, and supply-chain water is scope 3 (Li et al., Making AI Less Thirsty). Disclose both the site and the source figure.

How hydro evaporation is allocated can reverse a siting ranking. In NREL's survey of water factors for US generation (Macknick et al., 2012), hydropower's median is about 17 L/kWh, and the underlying studies charge all reservoir evaporation to power, although many reservoirs also serve water supply, flood control and recreation. Under that convention a grid that leans on hydro can carry more embedded water per kWh than a thermal grid. The factor also swings with the reservoir. NREL's state-level estimates (Torcellini et al., 2003) put hydro in Washington and Oregon well below the national hydro average and Arizona's far above it, so apply the regional factor rather than the national median. Split the evaporation among the reservoir's purposes and the hydro figure drops further. Say which convention a figure uses, and never compare WUE-source values built on different ones.

Each metric covers another's silence. PUE is silent on water; WUE fills that gap but is silent on carbon; CUE fills that but is silent on whether the energy did anything; DCeP fills that but is operator-defined and not comparable. No single metric is sufficient, and any one reported alone invites gaming the boundary. The defense is to report a stack (PUE + WUE for site and source + ERF + REF + CUE + a work-based denominator) so that improving one at the expense of another shows up immediately. An operator who lowers PUE by raising water use, or lowers WUE-site by ignoring WUE-source, is caught the moment both lines sit on the same page.

ERF (Energy Reuse Factor) is the share of a facility's energy that is beneficially reused outside its boundary, almost always waste heat exported to a district-heating loop or industrial offtaker (engineered in Chapter 5.9, monetized and regulated in Chapter 15.5). Its sibling ERE (Energy Reuse Effectiveness) folds reuse into the PUE frame: ERE = (Total − Reused) / IT, equivalently ERE = (1 − ERF) × PUE. ERE can legitimately drop below 1.0 because its numerator credits exported useful heat. It is a reuse-adjusted ratio, not a thermodynamic efficiency.

ERF is the metric most prone to intent inflation. A signed MOU with a municipality is not delivered heat, and ISO/IEC 30134-6 and EN 50600-4-6 require measured offtake at the boundary, not a press release. In Germany, operative EnEfG § 11(2) (read with the data-center scope in § 3(24)) requires new in-scope data centers entering operation from 1 July 2026, 2027 and 2028 to reach energy-reuse factors of at least 10%, 15% and 20%, respectively; the statute makes the 2027 and 2028 figures planned values. The requirement dates from the EnEfG of 13 November 2023 (BGBl. I No. 309), so verify the project threshold and commissioning date against the current statutory text. Operators currently have two years to reach the annual average. A government bill (Bundestag Drs. 21/8027; committee hearing on 5 October 2026) would leave the cohorts alone but give operators four years instead of two. It would also exempt a site where a cost-benefit analysis shows that a heat-network connection is not technically and economically reasonable. The bill is still pending, so check its status before you fix a commissioning plan.

REF (Renewable Energy Factor) is the renewable share of consumption (ISO/IEC 30134-3). The decision buried inside it is the accounting boundary. An unbundled-REC annual-matching REF of 100% can coexist with a facility that runs on a fossil grid every night, while a 24/7 carbon-free-energy (CFE) score exposes same-hour supply gaps. Eligible annual attributes, hourly matching and causal avoided emissions are separate tests. Chapter 15.3 covers the gap between annual matching and 24/7 CFE; here the point is only that REF without a temporal qualifier hides its own assumptions.

CUE (Carbon Usage Effectiveness) is total operational CO2e per unit of IT energy (ISO/IEC 30134-8). The shortcut CUE = CEF × PUE, where CEF is the carbon emission factor in kgCO2e/kWh, holds only when every kWh inside the boundary is purchased grid electricity. A campus running behind-the-meter gas, fuel cells or its own generation has to compute operational emissions source by source and divide by IT energy, and a zero-carbon grid factor does not make a gas-fired hall's CUE zero. CUE finally makes a low-PUE-on-dirty-power facility look as bad as it is. The fork is location-based versus market-based carbon: location-based reflects the physical grid you actually burned, while market-based credits contracts and instruments. Both are legitimate under the GHG Protocol, but they answer different questions, and an operator reporting only the flattering one is choosing what to hide. CUE is also silent on embodied carbon (the concrete, steel and silicon), which can dominate the lifecycle on low-carbon power and short refresh cycles and is treated in Chapter 15.6.

Every metric above is an infrastructure-efficiency ratio: it asks how cleanly you delivered energy to IT, never whether the IT produced anything. For AI, where a cluster's value is tokens generated or models trained, this is the deepest blind spot.

TUE and ITUE (proposed by Patterson at Intel with the LBNL Energy-Efficient HPC Working Group) push the boundary inward. ITUE is PUE's logic applied inside the server: total IT energy divided by the energy that actually reaches compute silicon (CPU, GPU, memory, storage), which separates out server fans, PSU conversion and VRM losses. TUE = ITUE × PUE then compares an air-cooled and a liquid-cooled facility without letting a boundary shift hide a loss, so ask for it whenever cooling modality is in question. It measures the overhead around the compute, not whether that compute did anything useful. Its cost is instrumentation: it needs per-server or per-component telemetry, which pays for itself only if it exposes losses the PUE boundary hides.

DCeP (Data Center Energy Productivity) is the boldest move: useful work divided by total facility energy. The Green Grid deliberately left "useful work" operator-defined: searches served, transactions cleared, or for AI, tokens generated per MWh and training-FLOP-hours per MWh. DCeP exposes under-utilization when accepted work is defined consistently. A half-idle cluster consumes energy while producing little, though counting redundant tokens as useful work can still flatter its score. Because operator-defined work is not comparable across operators, DCeP is an internal optimization metric, not a benchmark.

For 2026 AI operators the practical instantiation is joules-per-token (energy in the denominator's place, or equivalently tokens-per-joule with output in the numerator), alongside $-per-million-tokens-at-fixed-carbon. Together they fuse model efficiency, hardware efficiency and facility efficiency into one number the business cares about. Mind the dimensions: tokens-per-watt is not tokens-per-joule unless a time basis is supplied, and tokens-per-second-per-watt is. This metric aligns the efficiency team with the goodput team, since both teams now chase useful output per unit of scarce power.

The recast Energy Efficiency Directive (EU) 2023/1791, Article 12, and operative Delegated Regulation (EU) 2024/1364, Articles 3 and 5, establish annual reporting to the European database for in-scope data centers, including the defined PUE, WUE, ERF and REF inputs. Operators submit individual-facility information. The relevant Member State and the Commission can access it and must keep individual database records confidential, and the database's authorized public output is aggregated at Member-State and Union level. Article 12(1) separately requires Annex VII information to be made public subject to trade-secret and confidentiality protections, and national law may add operator-publication duties.

Alongside these filings, the Commission adopted delegated regulation C(2026) 3472 on 21 September 2026. It creates an EU rating scheme built on the same data, and it applies only once Parliament and the Council have finished scrutiny and the Official Journal has published it. Parliament extended the scrutiny period on 6 October 2026, so the two institutions have until 21 January 2027 to object. From 15 August 2027 the European database would then publish an annual label for each reporting data center. PUE is graded from A (1.15 or lower) down to G (above 1.9), and WUE from A (0.1 or lower) to G (above 1.0), with the WUE bands reading as liters per kWh. Each label also shows grid-flexibility provision, whether the site is "waste heat reuse ready," and an energy-supply breakdown that includes a new Low-Emission Energy Factor (LEEF), which counts nuclear as well as renewable supply. Operators below 500 kW, and sites not yet operating, can opt in, and a site not yet operating reports design values. Binding minimum performance standards follow on a separate track: the Commission takes consultation responses until 14 December 2026 and plans a legislative proposal for the second quarter of 2027.

In Germany, PUE is already a legal limit under the EnEfG. Under §11, a data center entering operation from 1 July 2026 must reach an annual-average PUE of 1.2 or better within two years. Existing sites face a limit of 1.5 from 1 July 2027, tightening to 1.3 from 1 July 2030. The same section sets the ERF cohorts, and the Act requires a national register. A government bill (Bundestag Drs. 21/8027) would keep 1.2 for new sites, though it would give them four years instead of two to reach it, and would loosen the existing-site limits to 1.6 and 1.4. The bill had its committee hearing on 5 October 2026 but is still pending, so confirm which pair applies before you plan against either. The register went public in August 2026. The Federal Environment Agency's September 2026 analysis of its 319 sites, after setting aside obviously faulty entries, found about 70% already at or below 1.5 and about 40% at or below 1.3.

The standards backbone is the ISO/IEC 30134 KPI series and EN 50600 / ISO/IEC 22237. The European PUE standard, EN 50600-4-2, is the 2016 text with a 2019 amendment, while ISO/IEC 30134-2 was reissued in 2026, so a European filing should state which edition its PUE method follows. Keep five boundaries apart: what you submit, what regulators can see, what national law and the EED make you publish, what the EU publishes in aggregate and, after the rating act takes effect, each facility's public label. From then on anyone can read a facility's label, but the rest of its database record remains confidential. The wider disclosure stack is in Chapter 15.7.

Trace. IT energy = 6,000 h × 5.0 MW + 2,760 h × 10 MW = 57,600 MWh. Mode A facility energy = 30,000 MWh × 1.20 + 27,600 MWh × 1.10 = 66,360 MWh, so its annual PUE is 66,360/57,600 ≈ 1.15. Mode B uses 30,000 × 1.10 + 27,600 × 1.18 = 65,568 MWh, giving 1.14. Select B on energy: about 800 MWh/year less for the same service. The unweighted bin means (1.15 for A, 1.14 for B) happen to rank the same way here but cannot size the saving. Energy weights PUE, not the number of bins or their hours alone.

Flip. A spends 0.10 more per cool-bin IT MWh but saves 0.08 per warm-bin IT MWh. With warm-bin IT-energy share x, A minus B = 0.10(1 − x) − 0.08x, and the two are equal at x = 0.10/0.18 ≈ 56%. The assumed share is 27,600/57,600 ≈ 48%, so B wins; above 56%, A wins. Any quality or thermal-failure test that B fails also rejects B, whatever its energy result. Annual metering follows ISO/IEC 30134-2. Chapter 5.8 owns the plant and weather curves, and Chapter 15.2 turns the measured saving into an investment decision.

The metric you optimize is a statement about what you are willing to make invisible. Optimize PUE alone and you will be rewarded for moving losses inside the IT boundary while poor useful-work productivity stays invisible. Low load commonly worsens the ratio, and the ratio still ignores every liter of water and gram of carbon. The 2026 practice is to commission against a stack: PUE for energy overhead, TUE when cooling modality is in question, WUE-site and WUE-source together, ERF/ERE for reuse, REF and CUE for carbon (each with a temporal and location-versus-market qualifier), and a work-based denominator, tokens-per-MWh or DCeP, so the whole apparatus is accountable to output. Each metric covers another's silence. Reported together on consistent boundaries and periods, a gain bought at another metric's expense shows up on the same page. In a power-bound era, choose the stack that exposes useful work, energy and water on their own denominators. A lower overhead ratio that costs accepted tokens or exhausts the basin’s water allowance is no efficiency win.