←25 Sep 2026 · Vinod Bijlani

In AI factories, sustainability and economics are now the same conversation

At 1 GW of IT load, the gap between an average and a highly efficient PUE is around 3,940 GWh and $315M of energy a year.

SustainabilityAI factoriesData centres
Part of the series: My Thoughts on AI factory build-outs →

In AI factories, sustainability and economics are now the same conversation.

At 1 GW of IT load, the difference between an average PUE of 1.54 and a highly efficient 1.09 is enormous: around 3,940 GWh less energy consumed and around $315 million lower annual energy cost at $80 per MWh. Every year. That is before adding another GPU or generating another token.

This is why power usage effectiveness matters. PUE is total facility power divided by IT equipment power, and the closer to 1.0, the better. The global industry average sits around 1.54. Google reports 1.09 fleet-wide. And some individual data centres are getting remarkably close to 1.0.

I looked at some of the facilities pushing the efficiency and sustainability frontier:

  • Google, Lancaster, Ohio: PUE 1.04
  • NREL ESIF, Colorado: PUE around 1.04, with warm-water liquid cooling and heat reuse
  • LinkedIn, Hillsboro, Oregon: published PUE around 1.06, with extensive free-air cooling
  • Verne, Keflavík, Iceland: 100% renewable energy and year-round free cooling
  • Lefdal Mine, Norway: renewable power and cold fjord-water cooling, built inside a former mine

There is a pattern: cooler climates and free cooling, liquid cooling as densities rise, renewable and low-carbon power, lower dependence on freshwater, and heat treated as something to reuse rather than simply reject.

But PUE isn't the whole sustainability story. Carbon matters. Water matters. Heat reuse matters. And as AI factories get larger, another metric becomes increasingly important: useful AI output per MW.

Ultimately the goal isn't simply to build more data centres or achieve the lowest PUE. It is to turn power into useful intelligence more efficiently.

What should be the defining efficiency metric for the next generation of AI factories: PUE, WUE or useful AI output per MW?

What is the most critical parameter for GW-scale data centres? PUE explained, with the world’s best PUE scores
Infographic · tap to enlarge

Invite Vinod to speak, write or advise.

Keynotes, panels and executive briefings on AI factories, sovereign AI and AI economics. Send a message on LinkedIn.