Power usage effectiveness
Power usage effectiveness (PUE) is a data centre's total energy use divided by the energy its IT equipment uses. A PUE of 1.0 would mean every watt goes to computing; higher values mean more is spent on cooling and power losses.
Also called: PUE
By Divya Sharma · Checked Nov 20, 2026 · How we review
The AI models you call run on chips such as a GPU, in buildings that also has to be cooled and fed with power. PUE measures that overhead. Divide the electricity the whole site draws by the electricity that reaches the servers, storage and network gear. A PUE of 1.5 means half a watt of overhead for each watt of computing.
Typical values
Google reports a 2025 fleet-wide average PUE of 1.09 in its environmental report, and cites an Uptime Institute survey average of 1.54 for the same year. For 100 MW of IT load running all year, that gap is about 394,000 MWh of extra electricity. Real values depend on climate, cooling design and how full the site is.
What it does not measure
PUE says nothing about how useful the computing is. A site full of idle servers can have an excellent PUE. It also ignores water: evaporative cooling often lowers PUE while consuming more water, which is why researchers track a separate water usage effectiveness (WUE) in litres per kWh. When you compare providers, such as an GPU cloud against self-hosting, PUE tells you about the building, not the price per token.
Related terms
- GPU
- A GPU (graphics processing unit) is a processor with many small cores built to run many calculations in parallel. Its speed at matrix maths makes it the standard hardware for training and running AI models.
- GPU cloud
- A GPU cloud is a service that rents out GPU machines by the hour or second for AI training, fine-tuning and inference. Specialist providers, sometimes called neoclouds, focus on GPUs and often charge less than the large general clouds.
- Self-hosting (AI models)
- Self-hosting an AI model means running it on hardware you control, such as your own servers, rented GPUs or a laptop, instead of calling a provider's API. You choose the model and keep the data, but you run the infrastructure.
- Inference
- Inference is running a trained model to produce output from new input, such as an LLM generating a reply to a prompt. It is the stage users pay for every time they call a model, as opposed to training.
- Serverless GPU
- Serverless GPU is a cloud model where you deploy code or a model and the platform starts GPU workers only when requests arrive, billing for the time they run and scaling to zero when idle.