The formula
Power usage effectiveness divides everything a facility draws by what its IT equipment draws. IT equipment means servers, accelerators, storage and networking. Everything else (cooling, fans, pumps, lighting and conversion losses) is overhead.
A worked example with round numbers: a site draws 10 MW at the meter, and 8 MW of that reaches its IT equipment. Its PUE is 10 divided by 8, which is 1.25. The other 2 MW is overhead.
The lowest possible value is 1.0, where every watt reaches the IT equipment. No real site gets there, because cooling and conversion always take something.
PUE is usually reported as an annual figure. A single reading on a cool night can look far better than the year, so check the period before comparing two values.
What a good PUE looks like
Good depends on climate, cooling design, load and age. A new site in a cool climate can run far lower than an older site in a hot one, with both operated well.
For context, Uptime Institute's global survey of operators, published in July 2025, put respondents' weighted average annual PUE at 1.54. The same report found that newer and larger facilities tend to report lower values.
The guide to energy efficiency in data centres covers what else an efficient site needs beyond a low ratio.
What PUE cannot see
PUE draws its boundary above the IT load. It grades the building around the load and treats the load itself as a given.
Two sites with the same PUE can produce very different amounts of work per megawatt. One may keep its accelerators on useful work. Another may leave them idling at a high draw. The ratio reads the same for both.
For the load, the matching measure is output per unit of energy. For AI inference that is tokens per watt.
The meter reads less, and PUE can still rise
A site that reduces what its equipment draws reads less at the meter. Its PUE can stay flat or rise, because IT power is the denominator and much of the overhead does not fall in step.
Take the round numbers again, with the overhead holding near 2 MW. Cut the IT load from 8 MW to 7 MW and the meter reads 9 MW instead of 10, a megawatt less. PUE moves from 1.25 to about 1.29, because the overhead stood still.
The ratio grades the overhead, and the overhead held still. If your programme is graded on PUE, agree up front to report a load-side saving in IT power and at the meter.
Two numbers for two questions
Use PUE for the building and a ratio of output to energy for the load. Report both, and never fold one into the other.
When a saving is quoted, ask which side it sits on. A cooling upgrade belongs in PUE. A reduction in what the accelerators draw belongs in the IT power figure, and in the output per unit of energy.
The four places a saving can be measured, from the chip to the bill, are set out in the guide to energy efficiency in data centres.
Questions people ask
- What is a good PUE?
- Lower is better, and 1.0 is the floor. Uptime Institute's July 2025 global survey put respondents' weighted average annual PUE at 1.54, with newer and larger sites tending to report lower values.
- How do you calculate PUE?
- Divide total facility power by IT equipment power over the same period. A site drawing 10 MW at the meter, with 8 MW reaching its IT equipment, has a PUE of 1.25.
- Does reducing IT power improve PUE?
- It cuts what the site draws at the meter. The ratio itself can stay flat or rise, because IT power is its denominator and much of the overhead does not fall in step.
Where ATHLAZ fits
ADAPT reduces what the IT load draws
ADAPT, AI-Driven Adaptive Power Technology, acts on the IT load itself, below the boundary PUE draws. It is a per-GPU software runtime that sets how each accelerator draws power while the work runs.
Its result shows in IT power, and less heat from the load leaves less for the cooling to remove. Measured: up to 21% less GPU die power on NVIDIA H100 NVL, over a continuous 48-hour window, with managed and baseline arms under an equal power cap.