Global Data Center PUE Stalls at 1.54

Global average data center power usage effectiveness (PUE) remained stagnant at 1.54 for the sixth consecutive year in 2025, according to Uptime Institute 15th Annual Global Data Center Survey. A PUE of 1.54 means 54 cents of every dollar spent powering compute equipment is spent on overhead: cooling fans, chillers, power distribution losses, and lighting. The stagnation persists even as AI training racks now routinely exceed 100 kilowatts per cabinet, a level that makes traditional air cooling architecturally insufficient and forces operators to redesign facilities they built for 5 to 10 kilowatts per rack.

The Hyperscale Gap

The 1.54 global average conceals a deep divide. Hyperscale operators, including Google, Meta, Microsoft, and Amazon, report average PUE of 1.10 to 1.15, achieved through purpose-built cooling infrastructure, free cooling in northern climates, and direct liquid cooling on high-density compute nodes. Co-location and enterprise facilities average 1.58 to 1.80, constrained by legacy air-cooled raised-floor designs that cannot be retrofitted for liquid cooling without major capital expenditure and potential service disruption. The Lawrence Berkeley National Laboratory 2024 U.S. Data Center Energy Usage Report projected total data center electricity consumption will reach 325 to 580 terawatt-hours annually by 2028, up from 176 terawatt-hours in 2023. The wide range in that projection reflects uncertainty about whether the industry can close the efficiency gap or whether AI workloads will compound existing inefficiencies at scale.

Liquid Cooling Accelerates Despite Legacy Drag

Vertiv reported that its liquid cooling revenue more than doubled in the first quarter of 2025, with the company projecting a 40 percent compound annual growth rate through 2028 as hyperscalers specify direct liquid cooling for all new AI compute pods. New GPU architectures including NVIDIA GB200 NVL72 draw up to 120 kilowatts per rack and are incompatible with air cooling at full utilization. ASHRAE Technical Committee 9.9, which publishes thermal guidelines for data center equipment, updated its standards for liquid cooled equipment to address rack densities above 40 kilowatts, acknowledging that the previous air-cooling-centric framework no longer reflects the industry trajectory. The Uptime Institute survey found that operators intending to deploy AI workloads in existing facilities cited power density and cooling capacity as the primary constraints, ahead of power availability and physical space.

Why Efficiency Stalls Despite Investment

The 1.54 global PUE average has not improved materially since 2019, despite hundreds of billions of dollars invested in new data center capacity. Uptime Institute attributes the stagnation to three factors. First, the majority of operating floor space was built before 2015 to air-cooled specifications with PUE targets of 1.5 to 2.0, and retirement of this infrastructure lags behind construction of more efficient facilities. Second, data centers in geographic markets without access to free cooling, including large portions of Asia, the Middle East, and the U.S. Sun Belt, cannot achieve northern European-style PUE regardless of equipment investment. Third, power conversion losses in the distribution path from grid to chip, including transformers, UPS systems, and rack power distribution units, account for 8 to 12 percent of total facility power and are difficult to improve below physical limits without switching to direct current distribution architectures that most operators have not adopted.

Critical Perspective

A facility operating at 1.54 PUE means 54 percent of incoming power feeds overhead systems, each of which contributes to harmonic and reactive-power impact at the service entrance. Cooling efficiency improvements cut per-facility peak demand, though rising rack densities at hyperscale sites often offset those savings at the distribution feeder level. Google 1.10 PUE is not comparable to a co-location facility 1.65 PUE, because Google designed its facilities around its workload and procurement; legacy enterprise sites cannot be retrofitted to that envelope without near-greenfield investment.

Why It Matters

If AI rack density continues to climb without a corresponding step-change in cooling efficiency, the LBNL 580 TWh upper-bound projection for 2028 becomes the operating case rather than the worst case. That translates to a near-tripling of US data center electricity demand in five years, with regional concentration in Virginia, Ohio, Texas, and the Pacific Northwest. The grid implications (interconnection queue length, transmission build-out, baseload retirement timing) are downstream of whether the global PUE number moves below 1.40 in this decade or stays at 1.54 indefinitely.

Related Coverage

• Data Center PUE Gains Stall as Liquid Cooling Adoption Hits 19% — Data center PUE gains stall as liquid cooling adoption hits 19%
• Liquid Cooling Cuts Data Center Power Use by 10% — Liquid cooling cuts data center power use by 10%
• Vertiv Launches MegaMod HDX: Prefabricated 10 MW Liquid Cooling Module — Vertiv launches prefabricated liquid cooling module for data centers
• IREN Builds $300M Liquid-Cooled AI Data Center at Childress, TX — AI data center uses liquid cooling for improved efficiency