Every vendor selling AI into energy quotes a percentage. Forty per cent is the most popular. It has been repeated so often that it now arrives as common knowledge, detached from what it originally described.
It is worth knowing what the original claim actually was, because the gap between that and how it gets used in a sales deck is the whole problem.
What the famous number described
The best-known version comes from DeepMind's work on Google's own data centres. What was reported was a reduction of roughly 40% in the energy used for cooling — not total facility energy — in Google's own data centres, measured against their own prior baseline.
Read that back slowly, because three qualifications are doing enormous work:
Cooling, not everything. Cooling is a large slice of a data centre's consumption but it is a slice. The headline number is not a 40% cut in the building's energy.
Google's data centres. Among the most heavily instrumented buildings on earth, with years of clean sensor data, a dedicated engineering team, and equipment that responds precisely to control signals. Almost nothing else looks like that.
Against their own baseline. Which was already unusually efficient. The improvement is real and it is relative to a starting point most facilities would consider aspirational.
None of this makes the work less impressive. It makes it less transferable, which is a different thing — and it is the thing the number loses on its way into a pitch.
What happens to it in a sales deck
The qualifications fall off. "40% cooling energy reduction in hyperscale data centres with dense instrumentation" becomes "AI cuts energy by 40%", and that sentence is then pointed at a factory, a hospital or an office estate that has none of the preconditions.
The move is not usually dishonest in intent. It is the same failure that produces "AI-powered" as a feature claim: a real result, stripped of the conditions that produced it, restated as a general property of the technology.
Why a savings percentage before a baseline is meaningless
Here is the part a buyer can use immediately.
A saving is a comparison. To claim one you need a defensible before: what the site consumed, under what conditions, at what production volume, in what weather. Energy consumption moves with all of those. A mild winter will "save" you more than most interventions.
Which means anyone quoting you a percentage before they have seen your data is quoting you a number about somebody else's building. They cannot know yours. There is no version of this where the figure is honest — it is either a guess or a borrowed headline.
The discipline that handles this properly has a name: measurement and verification. It exists because separating a genuine intervention effect from weather, occupancy and production variation is difficult, formal work — and it is the step that vendors leading with a percentage have skipped.
What to ask instead
Four questions that sort a serious supplier from a confident one:
What baseline are you comparing against, and how will you establish it here? "We'll measure for a period first" is the right answer. A number offered before this is offered without evidence.
How do you separate your effect from the weather and production volume? If there is no answer, any future saving is unattributable — you will not be able to tell whether it worked.
What exactly does your system control? Software that changes setpoints on plant is a different risk category from software that reports on it. Both are legitimate; conflating them is not.
Show me the record. What is retained, for how long, and could it survive someone external asking how a figure was produced?
Where we sit, plainly
We build the measurement, the record and the reporting: live dashboards from your meters and systems, alerts when consumption misbehaves, and reporting that generates itself with an audit trail behind it. We will commit to that in a fixed scope.
We do not promise a savings percentage, because we would have to invent it. Engineered interventions with verified savings are energy engineering with formal measurement and verification, and we work alongside that rather than pretending to be it.
The uncomfortable implication for our own market: if a supplier's opening move is a percentage, they have told you something about their evidence standards before you have asked a single question. This is the layer we build, and a Reality Check is a week and a fixed fee to establish what your data can actually support.
This is the part we do — the crossing from a demo to a system that survives production.