The green credentials of data centers don't add up

1.09. That's the score Google publishes for the energy efficiency of its data centers. A near-perfect number. The only catch: no one outside Google can verify it.

That number is called PUE, for Power Usage Effectiveness. Together with its cousin WUE for water, it serves as the green business card of every cloud giant: environmental reports, press releases, "sustainability" pages. And they share one awkward trait. They are either unverifiable, cherry-picked, or shifted off the books somewhere else.

The outlet Next.ink did the work of recomputing them from public data. We checked and expanded the analysis with independent sources. Here is what these numbers actually say.

Two ratios, one marketing reflex

PUE measures the ratio between a data center's total electricity and the power its servers actually use. A PUE of 1.0 would be the holy grail: not a single watt lost to cooling, conversion or lighting. The global average hovers around 1.56 according to the Uptime Institute, and it has been flat for five years. Hyperscalers, meanwhile, claim 1.2 or less on their best sites.

WUE measures the liters of water used for cooling, divided by the servers' electricity. The lower, the better, in theory.

Why the sudden appetite for transparency? Since 2024, Europe has required data centers above 500 kW to publish these figures. The AI boom makes their consumption visible, and therefore political. And a good ratio is a sales argument that costs almost nothing to display.

PUE is taken on trust

Here is the core flaw. Computing a PUE requires knowing the servers' own consumption. But only the operator measures that. From the outside, all you see is the site's total electricity bill. No way back to the ratio.

In other words, when Google announces 1.09, you take its word for it. It's a bit like a restaurant posting its own hygiene rating with no inspector checking behind it. The grade might be accurate, but it carries no more weight than a statement.

WUE is the more interesting case, because you can sometimes recompute it: a site's water and electricity volumes often surface in public filings. That is exactly what Next.ink did. Google does not publish its WUE, and when you reconstruct it, the result is very poor. Scaleway, by contrast, reports a figure consistent with the calculation. The ratio exists, but whether it matches measurable reality depends on the operator.

"Zero water": the accounting sleight of hand

In late 2024, Microsoft announced a new data center design that uses zero water for cooling. On paper, the WUE drops to zero. Perfect green performance.

Except water does not vanish from the books by magic. Cooling without water means cooling with electricity, using mechanical systems that draw more power. And generating that power consumes water. A lot of it. In the United States, it takes about 2 gallons of water to produce 1 kWh of electricity, versus 0.48 gallon used directly by an average data center for cooling. Most of a data center's water consumption is therefore indirect, pushed back to the power plant.

The "zero water" design moves roughly 125 million liters a year onto electricity generation. The water disappears from the data center's WUE because it is now billed to the plant, outside the measured boundary. Presented as an absolute zero, the move hides a transfer rather than a saving.

The average that erases the worst sites

Another blind spot: the fleet average. Google reports a single global figure for all its data centers, handy for smoothing over the hard cases.

The details of its 2024 report are telling. The Council Bluffs site in Iowa consumed nearly 4 billion liters of water on its own. The one in Pflugerville, Texas, used 38,000. A ratio of 1 to 100,000 between two sites in the same group, which the average files into the same box.

OVHcloud offers the same lesson from another angle. Its WUE is 0.20 in France and 1.31 in the United States. Same operator, same method, but climate and energy mix change everything. The star pupil depends on where you look.

Not greenwashing, but not proof either

These metrics have cut waste. An average PUE drifting toward 1.5, hyperscalers at 1.2, those are real engineering gains over fifteen years. PUE forced the industry to hunt down lost watts, and it worked.

The problem is what people make it say. A PUE tells you nothing about the energy source: a coal-powered data center with a perfect PUE is still dirty. It also ignores the local climate, and a site in a cold country inherits a good score for free.

Without third-party verification, without geographic context, these ratios remain self-declarations. The European directive making reporting mandatory since 2024 is a first guardrail. The question is who will actually grade the papers. For now, the most-touted green number in the sector is also the one that nobody outside the company displaying it can recompute.