AI Data Center Efficiency: Looking Beyond Power Usage Effectiveness

Jae-cheol Lyu is President and CEO of LG Electronics, a global innovator in technology and consumer electronics.

AI’s advancement is reshaping data center requirements, from the scale of computation they must support to the way facilities are designed and operated. As AI training and inference require more high-performance GPUs, data centers face rising rack power density and heat generation, along with changing demands for power, water and space.

Power usage effectiveness (PUE) has long been the most widely used metric for evaluating data center energy efficiency, but it has limitations across different climates and resource conditions. The Uptime Institute noted that PUE is well suited to tracking changes in a facility’s efficiency over time but that factors such as climate, resilience and water use need to be evaluated separately.

To properly assess data center efficiency in the AI era, we therefore need to move one step beyond PUE. We also need to look at three things: how flexibly a system can keep pace with rapidly evolving hardware, how efficiently it can move heat generated from the chip to beyond the facility and how reliably it can operate under local power, water and climate conditions.

Building For New Requirements

According to the same Uptime Institute survey, the share of operators running high-density racks of 30 kilowatts or more rose from 19% in 2025 to 24% in 2026, while roughly one in seven now operates racks of 50 kilowatts or more. Growing AI and GPU workloads are driving higher power densities, increasing both electricity demand and the amount of heat that cooling systems must remove.

For cooling providers, this changes the emphasis from fixed-specification equipment toward ongoing engineering support. Systems need to accommodate evolving AI hardware and site conditions, with real-world performance validation informing subsequent improvements.​