Watt Lies Beneath, Part 1: a series of articles by Matthias Haymoz, Executive Director, SDEA.
Data centre efficiency figures are quoted like fixed properties of a building. They aren’t. PUE moves with the weather, the load, and the age of the infrastructure – and the number in the brochure was almost always recorded under the conditions most flattering to it.
Western Europe has just come through the hottest June ever recorded – more than 3°C above the seasonal average, with 252 German weather stations setting all-time records – and the heatwaves have kept arriving since. Somewhere on the continent, on the hottest afternoon of this record summer, a server hall is running at the edge of its cooling envelope while the efficiency figure on its website might sit exactly where it has always sat – measured, once, in the mild conditions of an April.
There is precedent for how such afternoons can end. The last time the United Kingdom crossed 40°C, in July 2022, cooling failed at London facilities belonging to two of the world’s largest cloud providers; Google’s incident report described a “simultaneous failure of multiple, redundant cooling systems”. Machines were switched off to protect them. Services went dark. No blame can be placed on the engineers who worked through that day – heat no British building was designed for arrived, and they dealt with it. The detail worth keeping is quieter: every facility that struggled had an efficiency figure on file, and whatever that figure was, it was not measured on an afternoon like that. That is the number’s open secret. The efficiency you were sold was measured on a mild day.
A number with a weather problem
Power usage effectiveness – PUE, the ratio of total facility energy to the energy that actually reaches the IT equipment – is the industry’s favourite number. It appears in brochures, tenders, sustainability reports and planning applications, almost always as a single static figure: our facility operates at a PUE of 1.2. But PUE is not a property of a building. It is a behaviour, and it moves. It moves with the weather. Even the industry’s self-declared disclosures admit as much. Google’s published quarterly figures – like nearly every PUE in circulation, reported by the operator and independently audited by no one – show summer quarters running consistently above winter ones, and sites in the same fleet spread from 1.04 in Ohio to 1.26 in London within a single quarter. A number nobody verifies still cannot keep the weather out of it. Same corporate standard, same engineering culture; different air outside the wall.
It moves with the load. PUE is a ratio, and the overhead in the numerator is substantially fixed: chillers, pumps, transformers and fans draw power whether the servers are busy or idle. The Green Grid – the consortium that created the metric – warns in its own guidance that “a typical data center’s PUE is likely to vary with the levels of its IT load,” and notes the awkward corollary: an operator that virtualises its servers and cuts real energy use can watch its PUE get worse, because the fixed overhead is now spread over less IT work. A facility at 40% load in year one does not perform like the same facility at 85% load in year four. And it moves with age.
The Uptime Institute’s 2025 Global Data Center Survey puts the industry’s average annual PUE at 1.54 – while facilities commissioned in the last five years average 1.48, and the gap widens as plant degrades, filters clog and setpoints drift. The brochure number does not age. The building does. All of that is before asking what PUE leaves out entirely – the half-idle servers, the
water, the carbon. That question gets its own piece later in this series.
The mild-day machine
None of this variability is anyone’s failing. Physics is allowed to be physics. The problem is the practice that has grown up around it: treating a number that moves as if it were fixed, and letting the most flattering reading it ever produced stand for the whole. Because when a figure swings with weather, load and age, whoever chooses the moment of measurement chooses the answer. And the moment that gets chosen – for the launch press release, the sales deck, the planning submission – is rarely the August
afternoon at partial load in year three. It is the design estimate, or the commissioning test: new equipment, ideal conditions, often a simulated load standing in for real customers. The Green Grid’s own examination of the metric is blunt about the shelf life of such numbers: “the initial design of a data center is obsolete the day after the installation and commissioning is complete.” The same document cautioned, from the beginning, against comparing facilities on publicly reported PUE at all – the context
behind each number varies too much for the comparison to mean anything.
Two decades of marketing have been built on that comparison anyway. To be clear about where the criticism lands: the target here is a practice, not the operators – many of whom know their real numbers to three decimals and manage them daily. The practice is quoting a best-case snapshot as if it were the operating reality, letting a design projection from the mild day stand in for the sweaty, half-loaded,
year-round truth. It persists because it has been costless. A claimed 1.2 and a measured 1.5 read identically on a slide.
The mild day is ending
Three things are now converging on that practice – and each is less a threat than an invitation to do better. The first is stagnation that can no longer be presented as progress. Uptime’s survey shows the industry average essentially flat for six consecutive years – after falling steeply from 2.5 in 2007 to 1.65 in 2014. The easy gains are gone. When the real fleet-wide number stops improving, the distance between it and the brochure figures becomes harder to explain.
The second is weather that keeps auditing buildings without an appointment. It is doing so right now, one heatwave at a time, across an entire continent.
The third is regulation that has stopped accepting the snapshot – at least halfway. Under the EU’s data centre reporting scheme, established by the recast Energy Efficiency Directive, large facilities must now report their energy and water indicators from actual operation, every year – not from the design file. The scheme’s soft spot is that the numbers remain self-declared: no independent verification of any KPI is
required, so a reported figure is still a claim rather than a proof. A start, then, not a finish – but the direction is unmistakable. Annual, measured, as-operated reporting is becoming the floor, and a figure that was true once, briefly, in April will not clear it.
Efficiency is a performance, not a rating
There is a straightforward way out of this, and it asks for nothing exotic: report the year – the full twelve months, hot afternoons included, measured on the meters and checked by someone with no stake in the answer – rather than the best quarter or the commissioning test. That is the standard SDEA built its certification on – a year of measured operational data, validated by independent third-party audit – because anything shorter is an anecdote, and anything unaudited is a claim.
An efficiency figure earned that way is less flattering than the brochure number, almost
by definition. It is also the only kind that survives a heatwave. Because a data centre’s efficiency is not a nameplate rating fixed at the factory. It is a performance, sustained – or not – through every afternoon of the year. The brochures have spent twenty years quoting the mild day. The honest figure is the whole, verified year – and increasingly, that is the only figure anyone is going to accept.
__
Leave a Reply