Beyond PUE: Rethinking data center efficiency in the AI era

A broader view of efficiency as AI workloads proliferate

As AI workloads become better understood and more widespread, operators are moving away from theoretical assumptions and designing for real-world behavior. Historically, infrastructure decisions have been based on rated power, or the maximum a system could draw (though they rarely do). But AI workloads are dynamic, and usage doesn’t always match capacity. That mismatch can distort efficiency metrics and lead to overbuilt or under-optimized systems.

A more practical approach is to consider how workloads actually behave: when they spike, how they scale, and where inefficiencies emerge over time. This mindset extends to cooling strategies, power distribution, and even facility location and configuration. It also reinforces the importance of flexibility, because the pace of change in AI is unlikely to slow anytime soon.

A more complete view of data center efficiency looks beyond power efficiency alone, bringing in several additional considerations:

• How effectively compute power is translated into useful work
• How much water is required to support cooling
• What happens to the heat generated by densely packed hardware
• The carbon impact of energy sources and operations
• How the facility interacts with the electrical grid

Taken together, this creates a more realistic, multifaceted picture of performance that aligns better with the demands placed on AI-era data center infrastructure.

Managing tradeoffs, not just metrics

Data center efficiency isn’t a single objective. It’s a balancing act. Complementary metrics expand on what PUE can tell us and illuminate the ways in which targeting one affects others. Managing tradeoffs effectively requires a system-level perspective, with an eye toward resilience and long-term performance across multiple dimensions, including:

Water usage effectiveness (WUE)

Water usage effectiveness (WUE) highlights the tradeoffs between energy and water use in cooling strategies. As thermal loads increase, cooling choices have a direct impact on resource consumption and scalability. With water availability quickly joining access to power as a critical constraint on data center capacity, WUE is a necessary complement to PUE.

Energy reuse effectiveness (ERE)

Energy reuse effectiveness (ERE) focuses on where the heat generated by data centers goes. It encourages capturing and repurposing it within the facility, campus, or community rather than treating it as waste released into the atmosphere. AI data centers produce higher-quality heat at greater scale than traditional data centers do, making energy reuse more viable.

Compute power efficiency (CPE)

Compute power efficiency (CPE) shifts the conversation toward productivity and business outcomes. Instead of asking how energy is delivered, it asks how much useful compute is produced per unit of energy — tokens per watt. But CPE can sometimes conflict with PUE, highlighting the need to balance multiple objectives and metrics.

Carbon usage effectiveness (CUE)

Carbon usage effectiveness (CUE) brings environmental impact into focus, recognizing that not all energy sources are equal. Two facilities with similar PUE scores can have very different carbon footprints depending on how that energy is generated, e.g., renewables vs. fossil fuels.

Grid-aware efficiency (GAE)

Grid-aware efficiency (GAE) is an emerging set of calculations that consider how data centers interact with the grid itself: when they consume power, how they respond to demand fluctuations, and how they contribute to overall system stability. Shifting workloads, using stored energy, and participating in demand response programs play a role here.

From isolation to integration

Data center operators are well aware of looming constraints on growth, yet they remain locked in a race where speed to deploy often outweighs everything else, including efficiency. But it doesn’t have to be an either/or conversation. Building infrastructure systems that are efficient, adaptable, and aligned to the realities of AI computing can turn efficiency into a competitive advantage, not a constraint.

Adopting a broader efficiency measurement framework and treating it as a system-level equation with multiple interdependences better positions operators to scale sustainably while reducing operational costs through utilization and performance improvements. It also enables them to adjust to evolving regulatory requirements faster and helps ensure their ambitions don’t run roughshod over environmental and community concerns.

PUE remains a useful benchmark, but it’s no longer sufficient on its own. Its overuse can distort decision-making and mask inefficiencies on a number of other fronts. A multifaceted, interconnected framework that considers water, carbon, compute, energy reuse, and grid interaction is essential for navigating the next phase of growth. Learn more in Beyond PUE: Data Center Efficiency in the AI Era.