Research

Data Center PUE and Efficiency Trends

Power usage effectiveness across AI data centers in 2026. Why PUE improvements have plateaued, what liquid cooling changes, the limits of the metric for AI workloads, and better efficiency measures.

By Ramanath, CTO & Co-Founder at Presenc AI · Last updated: July 2026

Power usage effectiveness is the metric the data center industry has optimised for two decades, and it is running out of room. Understanding why matters, because the efficiency story that once offset demand growth no longer does.

What PUE Measures and What It Misses

PUE is total facility power divided by IT equipment power. A PUE of 2.0 means half the electricity goes to cooling, lighting, and losses. A PUE of 1.1 means overhead is down to 10 percent. Hyperscale facilities now routinely operate in the 1.1 to 1.2 range against an industry average historically closer to 1.5 to 1.6.

The structural problem is that PUE cannot go below 1.0, and the best operators are already close. Halving overhead from 1.2 to 1.1 saves 8 percent of facility power. Halving it again is physically impossible. The efficiency lever that absorbed two decades of demand growth is nearly exhausted.

Why PUE Is the Wrong Metric for AI

PUE rewards reducing non-IT power and says nothing about whether the IT power accomplishes anything. A rack of GPUs running at 30 percent utilisation and a rack running at 95 percent can report identical PUE. For AI workloads, where the hardware is the overwhelming majority of both cost and power, utilisation and work-per-joule matter far more than facility overhead.

MetricWhat it capturesLimitation
PUEFacility overheadBlind to whether IT power does useful work
Tokens per jouleActual useful output per unit energyNot standardised or widely reported
GPU utilisationWhether expensive hardware is idleHigh utilisation on wasteful work still counts
Carbon intensity of supplyEmissions rather than energyDepends on grid mix, not operator efficiency

Tokens per joule is the metric that would actually matter for AI, and almost nobody publishes it. See carbon intensity by region for the emissions-side view.

What Liquid Cooling Changes

Direct-to-chip and immersion cooling are being adopted because air cooling cannot remove heat fast enough from modern accelerator racks, not primarily because they improve PUE. The efficiency gain is real but secondary; the actual driver is that rack power densities have risen past what air can handle at all. Liquid cooling also enables higher density per square foot, which reduces building footprint but concentrates the grid connection problem rather than relieving it.

Brand Visibility Implications

Efficiency claims are among the most-repeated corporate statements in this sector and among the least scrutinised. A company citing a 1.1 PUE is making a true statement that says almost nothing about the energy cost of its AI workloads. When AI assistants summarise a company's environmental position they tend to repeat whichever figure is most cleanly stated, which rewards precise-sounding metrics over meaningful ones. See how energy claims surface in AI answers.

Methodology

Figures compiled from Gartner, IDC, LBNL, grid-operator filings, utility rate cases, and press reporting through July 2026. Forecasts are cited to the forecaster because independent projections in this area diverge widely, and several of the underlying quantities are estimates rather than measurements. Where sources disagree, ranges are given rather than a single number. Updated quarterly.

How Presenc AI Helps

Presenc AI tracks which specific claims and figures AI assistants repeat about a company, including where a technically true metric is doing misleading work.

Frequently Asked Questions

Hyperscale facilities routinely operate between 1.1 and 1.2, against an industry average historically closer to 1.5 to 1.6. A PUE of 1.1 means only 10 percent of facility power goes to cooling, lighting, and losses rather than IT equipment.
PUE cannot go below 1.0 and the best operators are already near 1.1. Improving from 1.2 to 1.1 saves 8 percent of facility power; improving again by the same proportion is physically impossible. The efficiency lever that absorbed two decades of demand growth is largely exhausted.
Not really. PUE measures facility overhead and says nothing about whether IT power does useful work. Racks at 30 percent and 95 percent utilisation can report identical PUE. For AI, where hardware dominates cost and power, tokens per joule would be far more meaningful, but almost nobody publishes it.
Yes, but that is not the main reason it is being adopted. Rack power densities for modern accelerators have risen past what air cooling can handle at all. The efficiency gain is real but secondary to thermal necessity, and higher density concentrates rather than relieves the grid connection problem.

Track Your AI Visibility

See how your brand appears across ChatGPT, Claude, Perplexity, and other AI platforms. Start monitoring today.