Data center PUE and energy efficiency field guide

What PUE is, and the ratio that scores the building

Power Usage Effectiveness, PUE, is the total energy a data center facility draws divided by the energy that reaches the IT equipment. It is a ratio. The floor is 1.0, which would mean every watt into the building went to the servers and nothing to anything else, and no real building hits it. Everything above 1.0 is the overhead: the cooling, the power conversion losses, the lighting, the controls. Lower is better.

A PUE of 2.0 says the building spends a watt on overhead for every watt that reaches a server. A PUE of 1.5 says it spends half a watt of overhead per IT watt. A PUE of 1.2 says the overhead is down to a fifth. The point of the number is to put a single figure on how much of the power bill is doing useful work and how much is the cost of keeping the room cold and the power clean.

PUE came out of The Green Grid in the mid-2000s and is now written into ISO/IEC 30134-2 and the European EN 50600 series. That matters, because it means there is a defined method, not just a marketing ratio. When two operators quote a PUE, the real question is whether they measured it the same way, over the same period, at the same points. The number is only as honest as the boundary it was measured against.

How do you calculate PUE?

PUE is total facility energy divided by IT equipment energy over the same period. Use energy, kilowatt-hours over a span of time, not a single power reading in kilowatts, if you want a number that means anything. A snapshot in watts gives you an instantaneous PUE that swings with the weather and the load. Energy over a year gives you the figure that tracks the bill.

Total facility energy is everything that crosses the meter at the boundary of the data center: the IT load plus the cooling plant, the UPS and transformer losses, the lighting, the fire and security systems, and the building loads that serve the white space. IT equipment energy is the power delivered to the servers, storage, and network gear, measured as close to that gear as the metering allows.

The arithmetic is trivial. The work is in the two inputs. Draw the boundary wrong, leave a load out, or meter the IT side at the wrong point and the ratio is wrong by exactly that error. Getting both numbers from the same period, cleanly metered, is the part that separates a defensible PUE from a number somebody guessed.

Total facility energy

All energy crossing the data center boundary: IT load plus cooling, power losses, lighting, and building systems

IT equipment energy

Energy delivered to servers, storage, and network gear, measured as close to the gear as the metering allows

kWh

Kilowatt-hour, the energy unit PUE is built from; use energy over a period, not a single power reading in kW

What is the difference between PUE and DCiE?

DCiE is the inverse of PUE, written as a percentage. Where PUE is total over IT, Data Center Infrastructure Efficiency is IT over total, times 100. A PUE of 2.0 is a DCiE of 50 percent, meaning half the power reaches the IT gear. A PUE of 1.25 is a DCiE of 80 percent.

The two say the same thing in different directions. PUE counts up from 1.0 and lower is better. DCiE counts toward 100 percent and higher is better. The industry mostly settled on PUE, because a ratio that starts at a hard floor of 1.0 is harder to dress up than a percentage, and because the work to be done is the overhead, which PUE names directly.

You still see DCiE on older reports and in some sustainability filings, so it helps to convert in your head. Flip one to get the other. DCiE is one divided by PUE. If somebody quotes 75 percent efficiency, that is a PUE of about 1.33, and now it is comparable to the rest of your numbers.

Field example: a hall at 1.5 and what that buys

Take a hall that meters 1,000 kW reaching the IT equipment and 1,500 kW crossing the facility boundary, averaged over a year of energy. PUE is 1,500 divided by 1,000, which is 1.5. That extra 500 kW is the overhead: roughly 350 kW of it cooling on a typical air-cooled plant, the rest split across UPS and transformer losses, lighting, and the building systems.

Read what 1.5 costs. For every 1,000 kW of computing, you are paying for 1,500 kW of grid power. At a flat rate, half again the IT energy bill goes to keeping the room running. Drop the PUE to 1.3 and the same 1,000 kW of IT pulls 1,300 kW from the grid, a 200 kW cut that shows up every hour of every day for the life of the plant.

The same hall does not hold 1.5 all year. On a cold night with the economizer carrying the load, the instantaneous PUE might read 1.25. On a hot August afternoon with every chiller running, it can climb past 1.6. The 1.5 is the annual weighted average, and that is the number that matches the bill. The free cooling guide covers how the cold-weather hours pull the annual figure down.

What counts as IT load and what counts as overhead

The IT load is the gear that does the data center's actual work: servers, storage arrays, network switches, and the supervisory and control gear in the white space. Everything else the building draws is overhead, and the split is the whole basis of PUE.

The overhead breaks into a few buckets. Cooling is the largest on most sites: chillers, pumps, cooling towers, the CRAH and CRAC units, and all the fans. Power chain losses come next. The UPS burns a percentage as heat, the transformers drop a few percent each, and the distribution losses add up across the chain. Then lighting, the fire and security systems, and the building services that keep the space habitable.

The judgment calls live at the edges. Is a network operations center on the same meter part of the IT load or a building load? Does a server's internal fan count as IT or overhead? PUE draws the line at the IT equipment power as delivered, so the server's own fans and power-supply losses sit on the IT side of the ratio. That is one reason a low PUE does not prove the IT itself is efficient. TUE, covered later, exists to close that gap.

Why PUE matters: cost, capacity, and carbon

PUE turns the cost of running the building into one number an owner can act on. Overhead energy is a direct operating cost, and on a large site the difference between a PUE of 1.6 and 1.3 is megawatts of continuous draw and a utility bill measured in millions a year. The metric makes that visible in a way a stack of meter readings does not.

It is also a capacity question. Utility power and the electrical plant are finite. Every watt spent on overhead is a watt that cannot go to a server, so a lower PUE means more of the power you bought and built turns into usable compute. On a constrained site where the next megawatt from the utility is years away, PUE is capacity you already own.

And it feeds the sustainability and reporting side. PUE is the headline efficiency figure in most data center disclosures, energy codes, and corporate carbon accounting. Regulators and large tenants ask for it. A plant that cannot report a credible, annualized PUE is a plant that cannot answer the questions its customers and its jurisdiction are starting to require.

What is a good PUE?

A good PUE depends on the climate and the age of the build, but the rough bands are well established. Legacy halls from the 2000s often sit near or above 2.0. The global average has held around 1.5 to 1.6 for several years, a plateau that has not moved much despite better equipment. New purpose-built facilities commonly target 1.3 or better, and the best hyperscale sites in cool climates run near 1.1, with a few reported close to 1.07 to 1.09. Treat those figures as approximate.

Context beats the headline number. A PUE of 1.4 in a hot, humid climate can be a better engineering result than a 1.2 in a cold, dry one, because the climate does a large part of the work in the second case. The number alone does not tell you who did the better job.

What is a realistic target for a given site falls out of the climate, the design, and the IT density, and the project's energy model is where that target should come from. Chasing a headline number you saw a hyperscaler quote, on a site that cannot physically reach it, spends capital on the wrong problem.

The Green Grid categories and where you meter

Where you put the IT meter changes the PUE, so the standards define measurement categories that say exactly where the IT energy is read. The Green Grid set them out as levels, and ISO/IEC 30134-2 carries them as categories. They run from a basic reading to a precise one, and a PUE is only comparable to another PUE measured the same way.

The categories step the IT measurement point closer to the gear. PUE Category 1, the basic level, reads IT energy at the UPS output. Category 2, the intermediate level, reads it at the PDU output. Category 3, the advanced level, reads it at the IT equipment input, in the rack. Each step down captures more of the downstream losses as overhead instead of IT, so a Category 3 reading on the same hall comes out higher, and more honest, than a Category 1 reading.

This is where the power monitoring guide and the metering hierarchy do the real work. The PUE is only as good as the meters feeding it and the boundary they sit on. Read the metering one-line before you quote a category, because a number measured at the UPS output and a number measured at the rack are not the same number, even on the same floor.

Design PUE, operational PUE, and why you report annual

Design PUE is the figure the energy model predicts for the building. Operational PUE is what the meters report once it is running and loaded. They are rarely the same, and the gap is information. A design PUE assumes a full IT load and a tuned plant. A new hall at 30 percent occupancy will measure a worse PUE than its design number until the load fills in, because the overhead is mostly there whether the racks are full or not.

PUE also moves hour to hour with the weather and the load, so a single instantaneous reading is close to meaningless as a scorecard. The metric that means something is annualized PUE: total facility energy over a full year divided by IT energy over the same year. That weighted average folds in the cold nights when the economizer carries the load and the hot afternoons when every chiller runs.

Report the annual number, and say it is annual. A vendor quoting a PUE without a period is quoting the best fifteen minutes of the year. The standards expect a continuous, defined measurement period for exactly this reason, and an owner comparing bids should ask for the period and the category before comparing the ratios.

What drives PUE

Cooling is the largest piece of overhead on most data centers, so it is the largest lever on PUE. The chiller compressor is the single biggest consumer, doing the work of pumping heat from the cold hall out to a warmer outside. Pumps, cooling tower fans, and the air handlers add to it. On a conventional air-cooled mechanical plant, cooling can account for the bulk of the gap between the PUE and 1.0.

The power chain is the second driver. The UPS converts and reconverts power and loses a few percent as heat, more at low load and in older double-conversion units. Transformers drop a few percent at each step down. Distribution losses run through the whole chain. None of it is large on its own, and all of it together is real, continuous overhead.

Air management sits behind the cooling number. Hot and cold air mixing in the hall forces the cooling plant to work harder and supply colder air than it should need to, which drives the cooling energy up. Containment, sealing, and decent airflow management cut that waste, and they are usually the cheapest PUE improvement on an existing floor. The free cooling guide covers how the climate and the economizer carry the cooling load for the hours the weather allows.

How do you improve PUE?

You improve PUE by cutting the overhead, and the cooling plant is where the biggest cuts live. Free cooling, running an airside or waterside economizer when the outdoor conditions allow, takes the compressor out of the loop for those hours and is the largest single move on most sites. The free cooling guide covers how it works and how many hours a given climate buys.

Raising the supply air temperature is the cheapest move that gets skipped. The ASHRAE thermal guidelines give a wider allowable inlet range than most halls actually use, and every degree you can safely raise the cold-aisle setpoint is less work for the cooling plant and more economizer hours. Hot- and cold-aisle containment stops the mixing that forces over-cooling. Sealing the floor and blanking the empty rack spaces cut the same waste.

The power chain has its own moves. A high-efficiency UPS, or one running in an eco or line-interactive mode where the design allows it, cuts the conversion loss. Distributing power at a higher voltage closer to the rack reduces the current and the losses. The order that pays is usually airflow and setpoints first, because they cost little, then the economizer, then the power chain.

Why does PUE get worse at low IT load?

PUE gets worse at low IT load because most of the overhead is fixed and does not shrink when the racks empty out. A chiller plant, the UPS, the lighting, and the fans draw a baseline whether the hall is at 30 percent or 90 percent occupancy. Divide that mostly fixed overhead by a small IT number and the ratio climbs.

This is the part-load penalty, and it bites new halls and lightly loaded ones the hardest. A facility designed for a 1.3 PUE at full load can measure 1.7 or worse in its first year, when the IT load is a fraction of the design and the plant is running near its fixed floor anyway. The PUE is not lying. The plant really is inefficient at that load, because it was sized for a load that has not arrived.

The practical lessons follow from the mechanism. Do not judge a half-full hall by a full-load target. Size and stage the plant so it can turn down: modular chillers, variable-speed everything, UPS that idle efficiently. And when you report PUE on a partly loaded site, report the load with it, because a PUE without a utilization figure hides the one thing that explains it.

The limits of PUE

PUE measures the overhead, and that is all it measures. It says nothing about whether the IT itself is doing useful work. A hall full of idle servers burning power for nothing can post a low PUE, because the wasted IT energy sits in the denominator and makes the ratio look good. A low PUE is not a low energy bill if the computing behind it is wasteful.

It can be gamed, and the gaming follows from the math. Move a loss from the facility side to the IT side and the PUE improves while the total energy is unchanged or worse. Run inefficient power supplies in the servers and the IT number goes up, the ratio goes down, and nothing got better. An operator who wants a flattering number can find ways to get one.

And PUE is silent on water and carbon. A plant can cut its PUE by leaning on evaporative cooling that drinks water, which a WUE would catch and a PUE never will. A plant on a dirty grid can have a fine PUE and a heavy carbon footprint. PUE is one axis of efficiency. It was never meant to be the whole picture, and treating it as the whole picture is the most common error made with it.

Partial PUE (pPUE) for a zone

Partial PUE, pPUE, applies the same ratio to a defined slice of the facility instead of the whole building. You draw a boundary around a subsystem or a zone, total the energy inside it, divide by the IT energy inside it, and you have a pPUE for that piece. It is the tool for measuring a cooling system, a single hall, or a containment zone on its own.

The common use is cooling. A cooling pPUE tells you how much energy the cooling system spends per IT watt inside its boundary, which lets you compare cooling approaches directly. Air cooling typically runs a cooling pPUE in the range of 1.15 to 1.35, while direct liquid cooling can reach 1.02 to 1.08, because it sheds the fan energy and much of the mechanical cooling. That gap is why the high-density section leans on liquid.

The catch is the boundary. A pPUE only means something if you state exactly what is inside it, because two people can draw the line around a cooling system differently and get different numbers. Use pPUE to compare like with like inside one facility. Do not quote a cooling pPUE as if it were the building's PUE, because it is missing everything outside its boundary.

Related metrics: WUE, CUE, ERE, and TUE

PUE has a family of companion metrics, each built to cover something PUE ignores. They share the same denominator, IT energy, so they read the same way.

WUE, Water Usage Effectiveness, is the water a data center uses divided by the IT energy, in liters per kilowatt-hour. It exists because the cheapest way to cut PUE is often evaporative cooling, which trades electricity for water, and WUE is what stops that trade from hiding. Industry water use varies widely, with efficient sites well under a liter per kilowatt-hour and others several times that.

CUE, Carbon Usage Effectiveness, is the carbon emitted per unit of IT energy, in kilograms of CO2 per kilowatt-hour. It folds in the grid's carbon intensity, so a plant on clean power scores well even at a mediocre PUE, and a plant on a dirty grid scores poorly even at a good one. ERE, Energy Reuse Effectiveness, credits energy sent off-site for reuse, such as waste heat piped to a district loop, and can fall below 1.0 because the reused energy is subtracted out. TUE, Total-power Usage Effectiveness, pushes the boundary inside the server to capture the IT's own fan and power-supply losses, which is the piece PUE counts as IT. Use the metric that answers your question, and report PUE with at least a WUE on any site that uses water for cooling.

High density, AI, and liquid cooling

High-density AI racks change the PUE picture, because they concentrate heat that air cooling cannot move efficiently. Above roughly 30 kW a rack, air cooling pays a steep energy penalty in fan power and over-cooling, and the cooling overhead climbs. Liquid cooling, bringing the coolant to the chip, moves that heat with far less energy, and it is the reason dense AI halls can hold a low PUE at densities air could never serve.

The reported results are real. Direct-to-chip liquid cooling has let operators push past 100 kW a rack while dropping PUE toward 1.1, because the fan load falls away and the mechanical cooling does less work. The cooling pPUE for liquid lands near 1.02 to 1.08 against 1.15 to 1.35 for air, and on a high-density floor that gap dominates the building number.

The caution is that PUE alone flatters liquid cooling. A liquid plant can cut PUE while raising water use, so WUE has to ride alongside it. And TUE matters more on dense gear, because at very high power the server's own fans and supplies are a real slice that PUE hides on the IT side. On AI builds, PUE is the start of the conversation, not the end of it.

Can you compare PUE across data centers?

Comparing PUE across data centers is only fair when the climate, the measurement category, and the period match, which they usually do not. A site in a cool, dry climate has a large free-cooling advantage that has nothing to do with how well it was engineered. Putting its 1.15 next to a hot-climate hall's 1.4 and calling the first one better is comparing the weather, not the design.

The category matters as much as the climate. A PUE measured at the UPS output, Category 1, will always look better than the same hall measured at the rack, Category 3, because more of the loss is counted as IT. If two operators do not state their category and their measurement period, their numbers are not comparable, and any ranking built on them is noise.

Benchmark against yourself first. Your own plant, same boundary, same category, tracked over time, tells you whether a change helped. Cross-facility comparison is for rough banding and for spotting a plant that is clearly off the pace, not for declaring a winner between two sites in different climates measured different ways.

What to document

Report a PUE with no method behind it and no one can reproduce or check it, which makes the figure worthless the moment a tenant or auditor asks how you arrived at it. The record is what lets a reviewer, a tenant, or an auditor check the figure a year later, and what lets you compare this quarter to last.

Capture the measurement category, the boundary you drew, the measurement period, the total facility energy and the IT energy with the meter points they came from, the IT utilization over the period, and the climate context. Note whether the figure is design or operational, and annual or instantaneous. If you report a pPUE, state exactly what is inside its boundary. The power monitoring guide covers the meter accuracy and reconciliation that the PUE inherits from the metering hierarchy.

Common mistakes

• Quoting an instantaneous or best-day PUE as if it were the annual figure.
• Drawing the boundary wrong, so a facility load lands on the IT side or an IT load falls outside the meter.
• Comparing PUE across sites in different climates without saying so.
• Reporting a PUE with no measurement category, so a UPS-output number gets compared to a rack number.
• Chasing a lower PUE while the IT runs idle servers that waste the energy the ratio rewards.
• Cutting PUE with evaporative cooling and never tracking the water with a WUE.
• Judging a partly loaded hall against a full-load PUE target and missing the part-load penalty.
• Treating PUE as the whole efficiency story and ignoring carbon and water.

Field checklist

Standards and references

The defining standard for PUE is ISO/IEC 30134-2, which formalizes the metric, the measurement categories, and the requirement for a defined, continuous measurement period. The European EN 50600-4-2 carries the same metric in that series. The Green Grid originated PUE and DCiE and published the measurement levels and usage guidelines the ISO categories follow.

The ASHRAE TC 9.9 thermal guidelines set the inlet temperature and humidity envelope that decides how warm you can run the hall, which drives the cooling energy and the economizer hours behind the PUE. The related water and carbon metrics, WUE and CUE, and the energy-reuse metric ERE come from The Green Grid as well. Exact category definitions and edition details shift as the standards are revised, so confirm them against the current published edition and the project's chosen reporting framework before you put a category and a number on a report.

Where a jurisdiction or an energy code mandates a PUE method or a maximum, that adopted requirement controls over any rule of thumb. The project specification and the reporting framework the owner has committed to decide which method you measure and report against.

Units, terms, and conversions

PUE is a dimensionless ratio, so the inputs have to be in the same units, energy over the same period, for it to mean anything. The companion metrics carry units, and the same idea shows up under a few names across reports.

PUE is total facility energy over IT energy. DCiE is its inverse as a percentage. WUE is liters of water per kilowatt-hour of IT energy. CUE is kilograms of CO2 per kilowatt-hour of IT energy. Energy is in kilowatt-hours or megawatt-hours over the measurement period, and a common error is mixing a power reading in kilowatts with an energy figure in kilowatt-hours.

PUE

Power Usage Effectiveness, total facility energy divided by IT equipment energy, 1.0 or higher

DCiE

Data Center Infrastructure Efficiency, the inverse of PUE expressed as a percentage

pPUE

Partial PUE, the same ratio for a defined subsystem or zone within a stated boundary

WUE

Water Usage Effectiveness, water used per unit of IT energy, in liters per kilowatt-hour

CUE

Carbon Usage Effectiveness, carbon emitted per unit of IT energy, in kg CO2 per kilowatt-hour

ERE

Energy Reuse Effectiveness, PUE adjusted for energy exported for reuse, which can fall below 1.0

TUE

Total-power Usage Effectiveness, extends the boundary into the IT to capture server fan and supply losses

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• What is free cooling in a data center?
• What is an EPMS in a data center?
• What is the difference between cold and hot commissioning of a BESS?
• What kind of batteries do data centers use?
• What is a rack BBU?