For years, efficiency was measured one way. It worked until AI changed the equation. Traditional efficiency metrics don’t tell the full story. Liquid cooling is becoming the default for AI infrastructure, but the way we measure efficiency requires tracking data center efficiency metrics beyond Power Usage Effectiveness (PUE), which only captures energy overhead. AI demands a different question: How much compute does that energy actually produce?
Goldman Sachs estimates that 76% AI servers deployed by the end of 2026 will be liquid-cooled. As this shift from air cooling occurs, operators will turn to Power Compute Effectiveness (PCE) and Water Usage Effectiveness (WUE) as important AI data center KPIs in addition to PUE. This is necessary because PUE only compares total facility power to IT power, without accounting for compute outcomes.
What is replacing PUE in data centers?
Even when data centers operate at PUE levels of 1.4 to 1.8, systems can be inefficient if they limit rack density, increase airflow, and constrain compute performance. PUE tracks energy overhead, but that isn’t enough for power-intensive AI workloads, which require metrics like PCE to determine data center energy efficiency in converting power to AI tokens.
What is Power Compute Effectiveness?
In AI data centers, efficiency is no longer measured by energy use alone, but by what that energy produces. Power Compute Effectiveness (PCE) defines this shift by quantifying how much of your available power is actually doing useful computational work:
A higher PCE means more of your provisioned power is actively driving compute, not stranded, underutilized, or lost in system inefficiencies.
Why do these AI data center metrics matter?
With PCE as the foundation, AI performance is evaluated through what power actually delivers—not just what it consumes. Two outcome-driven metrics define this shift:
Tokens per watt — how much AI work is generated per unit of energy
Cost per token — how efficiently that energy converts into usable output or revenue
As AI workloads scale, power systems are hitting real constraints. Capacity is finite, costs are rising, and inefficiencies compound quickly. PCE brings these dynamics into focus:
Higher tokens per watt = more revenue-generating compute delivered within fixed power limits
Lower cost per token = predictable, scalable costs as AI workloads grow
In the AI era, advantage doesn’t go to who has the most power, but to who converts power into compute the most effectively.
How does liquid cooling improve efficiency and performance?
In AI environments, efficiency is not determined by individual components, but by how cooling, power, and compute operate as a coordinated system. Liquid cooling helps improve PUE, delivering scores of 1.05 to 1.15. Leading operators like Amazon Web Services report fleet-wide PUE near 1.15, demonstrating what optimized infrastructure can achieve.
Beyond PUE, liquid cooling enables higher rack densities without thermal constraints, less reliance on energy-intensive airflow systems, and more stable GPU performance with less throttling. From a system design perspective, liquid cooling enables a higher Delta T (ΔT), removing more heat, reducing pump energy, and improving heat transfer efficiency. This results in lower PUE thanks to reduced waste and higher PCE because of improved output.
How is WUE measured in liquid-cooled data centers?
As liquid cooling becomes integral to data center operations, operators are placing greater emphasis on tracking their water use, driven by sustainability goals, evolving government regulations, and increased public scrutiny. WUE measurements compare a facility’s consumption with its IT equipment energy usage.
Single-phase direct-to-chip (DTC) liquid cooling influences WUE by changing how heat is removed—and giving operators more control over where water is used. By capturing heat directly at the processor within a closed-loop coolant circuit, DTC reduces the need to move large volumes of air at the rack level.
This shifts water use away from the IT environment and into the facility heat-rejection layer—where it can be actively managed based on climate, infrastructure, and sustainability goals. In many deployments, this enables higher rack densities (>100 kW) while aligning water usage more closely with operational and environmental priorities.
Sustainability is about the effective use of resources
As AI deployments scale, sustainability is no longer defined by reducing energy use—it’s defined by how efficiently that energy is converted into compute. Metrics like tokens per watt and cost per token are becoming more meaningful, shifting the focus from consumption to output.
Liquid cooling enables this shift by reducing cooling-related energy demand, supporting higher compute density without increasing footprint, and enabling warm-water operation for heat reuse in the right environments. But these gains only materialize when cooling is designed as part of the system. Treating cooling as a standalone product introduces inefficiencies across flow, pressure, and heat transfer—limiting what each watt can ultimately deliver.
AI efficiency depends on performance per watt
With AI operations, efficiency isn’t just about how much energy you consume. Rather, it’s about how much compute you produce from it, making output a more meaningful measure of performance. Schneider Electric can help data centers maximize efficiency. As AI workloads scale, the focus shifts to newer metrics that measure how much useful output is generated per unit of energy consumed. Improving these metrics requires more than incremental gains. It requires infrastructure designed to sustain performance under continuous, high-density load. The winners in AI will be the ones who get the most compute from every watt. Learn more about how to harness the energy and water efficiencies of Direct Liquid Cooling.
Add a comment