Author: Faruk Ucur, Vice President, Engineering CHVAC EMEA, Trane Technologies
Across the globe, AI adoption is accelerating at unprecedented speed. Hyperscaler,colocation operators, enterprise data centers, and emerging AI factories are rapidly scalingtheir IT infrastructure to support increasingly complex models and massive datasets. Thissurge is driving increased power demand, driving cooling requirements to new extremes.
Traditionally, the data center industry is using Power Usage Effectiveness (PUE) and WaterUsage Effectiveness (WUE) as primary benchmarks for resource efficiency. These metricshave served the industry well for years – simple, intuitive indicators of how effectively afacility uses energy and water.
But today’s AI workloads, advanced chips and liquid-cooling architectures are changing theequation. Next-generation chips designed for AI and accelerated computing operate undervery different thermal profiles and performance characteristics than traditional CPUs. Theygenerate far higher heat loads, rely more heavily on liquid cooling, and deliver dramaticallydifferent compute output depending on operating temperature.
As compute power becomes the dominant driver of energy consumption for datacenters, additional efficiency metrics have started to emerge which aim to holisticallyrepresent efficiency, sustainability, and operational performance.
Why PUE isn’t enough to measure efficiency of AI data centers
PUE Ignores Compute Power
A facility may report an excellent PUE yet deliver relatively low compute output. As globalcompute demand is expected to triple – rising from 82 GW in 2025 to more than 220 GW by2030 according to McKinsey & Company – organizations need metrics that factor in usefulwork, not just energy in vs. energy out.
Climate and Cooling Variability Distort Measurements
AI data centers in hotter climates or regions with water scarcity face environmental realitiesthat inherently impact both PUE and WUE. A site with highly optimized cooling may stillappear “inefficient” due to geography rather than engineering.
AI Workloads Change the Relationship Between Cooling, Temperature, and Output
A chip may run “adequately” at higher temperatures – improving PUE by reducingmechanical cooling – but deliver superior compute output at lower temperatures. In this casePUE worsens, compute performance increases and total system productivity rises. Thiscontradiction highlights why today’s metrics fall short for AI-driven environments.
To evaluate efficiency in modern facilities – especially those supporting AI,hyperscale, and high-density colocation applications – alternative metrics are beingconsidered which combine energy and water metrics with compute-centric KPIs.
Below is an example of some advanced metrics which aim to provide a more complete andscalable view of performance:
- Performance Per Watt (PPW) measures compute output per unit of energy. Aholistic way to align IT performance with facility efficiency.
- Data Center Energy Productivity (DCeP) evaluates how much “useful work” (taskscompleted, data processed, etc.) is produced for each unit of energy consumed.
- Compute Power Efficiency (CPE) expresses the computational power delivered perunit of energy. Ideal for AI-heavy operations.
- Energy-Proportional Computing (EPC) assesses how efficiently a system adjustspower use relative to workload. The closer energy tracks to compute demand, thebetter.
- Green Efficiency (GE) links compute output to carbon impact, enablingsustainability-aligned decision Performance Per Liter (PPL) aligns water consumption with compute output -critical for regions facing water scarcity or high regulatory scrutiny.
Why These Metrics Will Matter
These alternative measures could allow stakeholders to evaluate not only the facility’sresource efficiency but also the value generated from each kilowatt, liter, and kilowatt ofcooling capacity. They reshape how operators assess design decisions, cooling architecture,chip temperatures, and sustainability strategies
However, this expanded toolkit also brings greater complexity. The convergence ofhigh-density cooling, electrification, water strategy, advanced controls, modeling, andregional regulatory variance demands a more integrated approach.
A Holistic Partner for an AI-Driven Future
To succeed, organizations need an engineering and thermal-management partner deeplyexperienced in system-level design, carbon reduction, performance optimization, andscalable solutions. Early collaboration is critical – from concept and design tointegration and ongoing operations.
Ideally the right partner should bring expertise in understanding the entire data centerecosystem from roof to chip. Offer consultative approach like modeling, testing, intelligentcontrols, energy recovery, thermal storage, real-time analytics, and predictive maintenance.This will help operators maximize compute potential while aligning with local communityneeds, environmental expectations, and sustainability priorities.
