What we measure, and what we do not
Imagine three data-center proposals sitting on the same desk.
Each reports a Power Usage Effectiveness of 1.2.
The first number is a design estimate for a facility running at full planned load. The second is an annual average from an operating site. The third covers one technical area inside a larger mixed-use building.
On the page, they look identical.
They are answering different questions.
This is the peculiar power of an infrastructure metric. A number can be calculated correctly and still create a misleading comparison when its boundary, method, and operating conditions disappear.
PUE is the best-known example. It divides total data-center energy by the energy delivered to IT equipment. If servers, storage, and networking consume 100 units while cooling, power conversion, lighting, and other supporting systems consume another 20, the PUE is 1.2.
That tells us something useful: 20 additional units support every 100 used by the IT equipment.
It does not tell us how much useful computing work was completed, how much water the cooling system consumed, where the electricity came from, whether heat was reused, how heavily servers were utilized, or what it took to manufacture and build the infrastructure.
A metric is a carefully chosen window into a system.
The trouble begins when we mistake the window for the whole view.
The boundary decides what the number means
PUE became successful because it made a complicated engineering problem legible. It gave operators a common way to separate electricity used by computing equipment from electricity consumed by the facility around it.
The formula is simple only after the boundaries have been agreed.
Where is total facility energy measured? Which equipment counts as IT? Does the figure cover the entire site or part of a mixed-use building? Is it a design value, a full-load test, an instantaneous reading, or an annual average? How much equipment was installed and active during the period?
The current ISO/IEC 30134-2:2026 standard exists because these questions change the interpretation. It defines measurement categories and reporting rules for PUE, including mixed-use buildings, on-site generation, and energy that is not metered directly.
Operating load can change a ratio even when no equipment becomes more efficient. A facility that uses 20 units of supporting energy and 50 for IT has a PUE of 1.4. If IT load rises to 100 while supporting energy stays at 20, PUE improves to 1.2.
Total energy use has increased. The infrastructure ratio has improved.
Nothing dishonest has happened. Occupancy changed.
Climate, redundancy, facility age, cooling design, and size can change the result too. A resilient facility may keep additional equipment ready or operating. A site in a cool climate may reject heat with less mechanical cooling than the same design in a hot one.
This is why PUE is strongest when a facility compares itself with its own past. Turning it into a league table between unlike facilities asks more of the metric than it can safely provide. The full PUE article examines the formula, measurement categories, load effects, and comparison problem in detail.
The same rule applies beyond PUE.
A maximum specification is not observed demand. A modeled value is not an operating result. A test under stated conditions is not an annual average. A site-specific measurement is not automatically a product-family figure.
The number becomes meaningful only when the reader knows how it was produced.
The work itself often sits outside the metric
The most important thing PUE leaves out is the work completed by the IT equipment.
An idle server still belongs in the denominator. So does an old server completing far less work per watt than newer hardware. So does overprovisioned equipment waiting for demand that may never arrive.
A data center can therefore improve its PUE while leaving substantial inefficiency inside the racks.
This is not a flaw in the calculation. PUE measures facility infrastructure efficiency. The error is allowing it to represent all data-center efficiency.
A work-per-energy measure sounds like the obvious answer, but data centers do many unlike things. A storage system, web service, scientific cluster, and GPU inference platform do not produce one common unit of output. Transactions, bytes stored, simulations completed, images rendered, and tokens generated cannot be added into a universal score.
Within a defined workload, operators can still track useful work per unit of energy over time. That requires stable descriptions of the hardware, software, precision, utilization, service level, and work being measured.
This pattern repeats across infrastructure measurement.
PUE works well when a data center compares compatible periods and boundaries.
Work-per-energy measures work well when an operator compares a clearly defined service or workload with an earlier version of itself.
Metrics become weaker as the comparison moves farther from the conditions that produced them.
The same caution applies to power figures. Connection capacity, peak demand, average load, annual consumption, IT load, and useful output are related but distinct. A megawatt needs the right noun attached to it.
Every new metric makes another boundary visible
The limits of PUE led to more data-center metrics.
Water Usage Effectiveness, or WUE, relates site water input to IT energy. It makes a resource visible that PUE cannot detect.
But WUE introduces its own questions. Does the figure describe water withdrawn or consumed? Does it cover total input, freshwater, or potable water? Does it include water associated with electricity generation? Is the site in a water-rich region or a stressed watershed?
A liter does not have the same consequence everywhere. A low WUE can still accompany a large absolute water demand, while the wider water footprint depends on cooling, electricity, climate, and place.
Carbon metrics face a similar problem. Location-based accounting reflects the electricity supplied through the local grid. Market-based reporting may also account for contractual purchases or certificates. Both can be calculated correctly while describing different relationships between a facility and the power system.
Other metrics track renewable energy, heat reuse, cooling performance, server utilization, or computing work. The expanding ISO data-center performance series makes a point that one universal score cannot: performance has several dimensions.
Those dimensions can pull against one another.
Additional redundancy may increase energy overhead while reducing interruption risk. Evaporative cooling may improve energy efficiency while consuming more water. Running equipment longer can avoid manufacturing another machine, while newer hardware may complete much more work per watt. Heat recovery can make better use of energy leaving a data center, but only where a compatible user exists nearby.
A metric can tell us whether one part of that system improved. It cannot decide which trade-off matters most for the site, workload, customer, or community.
Time creates another boundary. Most operational metrics begin after a data center opens. They say little about the concrete, steel, electronics, cooling equipment, servers, GPUs, transport, maintenance, replacement, and end-of-life treatment required across its life.
A data center contains several physical lifetimes. An operational improvement can be real while remaining only one side of a wider exchange.
A trustworthy number carries its conditions
Good measurement does not require one perfect metric. It requires discipline about what each number can support.
A useful claim should identify the metric and definition, physical boundary, equipment included, measurement period, operating load, relevant climate conditions, and whether the value is modeled, tested, or observed in normal operation.
A comparison needs compatible definitions, periods, boundaries, and assumptions on both sides.
“PUE of 1.2” is less informative than it first appears.
“Annual operating PUE of 1.2 for this facility, measured under the stated category during this reporting period” gives the reader something closer to evidence.
A range may sometimes be more useful than a single value. So may seasonal results, load bands, or separate measurements for different configurations.
The point is not to make every claim unreadable. It is to preserve the information that determines what the number means.
European regulation is moving in this direction. Covered data centers already report energy-performance and water-footprint information under the Energy Efficiency Directive and Commission Delegated Regulation 2024/1364. The Commission is developing a broader rating scheme and minimum performance standards. Its data-center energy-performance page tracks the current status.
More numbers will not automatically create clarity. A label still depends on reliable reporting, comparable boundaries, and an explanation of what remains outside it.
Evidence should follow the physical system
Policloud develops and deploys physical, modular data-center infrastructure for defined sites.
For that kind of infrastructure, one universal efficiency figure is less useful than evidence attached to the configuration and conditions that produced it. Cooling architecture, workload, utilization, climate, measurement boundary, and reporting period can all change the result.
A maximum specification should remain a maximum specification.
A modeled value should remain modeled.
A test result should name the test conditions.
A site result should remain attached to the site.
An operational metric should not become a lifecycle claim.
This has a commercial cost. A single low number is easier to place in a headline.
It creates a larger problem when a technical buyer, regulator, journalist, or local authority asks what the number describes.
The honest distinction is between what has been measured well enough to support a decision and what remains unknown, incomplete, or outside the current boundary.
That is what we measure.
And what we do not.
A trustworthy metric does not make a complicated system look simple.
It tells the reader exactly which part of the system has become visible.
