Since publishing my piece on prioritizing energy accounting in data centers back in June, there’s been a steady wave of media coverage highlighting the rapid growth of datacenters and their impact on energy demand. On August 17, The Economist reported that U.S. household electricity costs have risen by 7% in 2025, partly due to the expansion of datacenters and AI workloads. At the same time, industry stakeholders have released numerous updates on AI’s electricity use. Yet, a recent article underscores a persistent issue: the lack of standardized metrics, methodologies, and best practices for measuring datacenter energy footprints. In this blog, I’ll focus on the need for clear metrics to better understand where electricity is actually going in the datacenter.

The Blind Spot in Our Metrics

Datacenter operators have traditionally used metrics like Power Usage Effectiveness (PUE) to inform infrastructure decisions. PUE—defined as the ratio of total facility power to the power used by IT equipment—was instrumental in highlighting inefficiencies from cooling and power conversion. Over time, PUE values have improved and stabilized, with major players like hyperscalers (global cloud providers) achieving values below 1.2, and multinational colocation providers (which lease server space to enterprises) reaching below 1.4, according to Uptime Institute. In that context, PUE served its purpose well.

With a PUE of 1.2, more than 80% of a datacenter’s electricity reaches the IT equipment. However, PUE says nothing about how efficiently that electricity is used to deliver actual IT services. In fact, the more electricity the equipment consumes for a particular service, the better (i.e., lower) the PUE appears. This limitation has been well-documented, yet the industry continues to treat PUE as a standard metric. As datacenters become major players in global energy consumption and net-zero commitments fade into the background, it’s increasingly urgent to measure how effectively electricity is being used to deliver IT services—not just where it goes.

Underutilized Silicon

Most of the electricity delivered to IT equipment is consumed by silicon components that power computing—CPUs, GPUs, memory, NICs, network switches, flash storage, and hard drives—across both compute servers, networking gear and shared storage clusters. These components draw power even when idle, so electricity is used most efficiently when server utilization is high. Although precise data on average utilization is limited, hyperscalers are generally known to operate servers at high loads (typically over 50%) for first-party workloads, maximizing return on investment.

Unfortunately, server utilization among typical IT operators—such as customers of colocation providers or public cloud users—remains low, often ranging between 12% and 18%. Even sustainability certification standards, like CLC/TS 50600-5-1, set the utilization threshold at just 20%. Cloud providers have also observed inefficiencies, such as up to 25% of memory being stranded when all CPU cores are rented or only half the rented memory being actually used. Similarly, Microsoft has reported GPU utilization below 50% in real-world deep learning deployments. To address this, we need to equip users with better metrics and methodologies to meaningfully increase IT equipment utilization.

Non-IT electricity

There are also sources of electricity consumption that don’t directly contribute to computing, communication or storage, such as power supply units and fans. Fortunately, power supply unit efficiency is regulated by standards, with top-rated units achieving over 90% efficiency. Fan power, however, can range from 5% to 15% of total server power consumption, depending on server load and workload type. The latter becomes especially problematic when using PUE as a metric. Because PUE improves as servers run hotter, higher fan usage—needed to manage that heat—can misleadingly lower the PUE value. Additionally, including fan power within IT equipment skews PUE comparisons between air-cooled and liquid-cooled systems. Finally, at lower PUE values (indicating more efficient facilities), the relative impact of fan power becomes even more pronounced, further distorting the metric.

What Should We Be Measuring?

“To measure is to know” is a quote often attributed to Lord Kelvin. What exactly should we be measuring? Here’s a starting point:

• IT electricity: We need mechanisms to measure the electricity consumed by components directly responsible for computation, communication and storage. Because direct measurement isn’t possible today, component utilization metrics when running workloads—readily available for most hardware—can be combined with typical power breakdowns per component to estimate compute energy use.

• Idle electricity: Measuring power consumed when equipment is idle is critical to understanding the utilization thresholds needed to effectively amortize this overhead.

• Non-IT electricity: Fan power may not be measurable but fan speed is, and fan power is correlated with fan speed allowing operators to attribute fan power to cooling. Power supply unit’s consumption can be estimated based on industry standards for their ratings and the load on the server.

These aren’t abstract ideas. Academic literature has proposed similar models for decades. A few initiatives – e.g., Open Compute Project (OCP) – have started to formalize them. There are also open-source and proprietary tools that enable measuring them.

Ultimately, the goal is to determine how much electricity is required to deliver a specific IT service—for example, answering a query to a large language model. We need proper metrics and methodologies with an apple-to-apple comparison across IT services to shed light on where electricity is going. A logical starting point is understanding how efficiently electricity is being used at the server hardware level.

About the author: Babak Falsafi is a Professor in the School of Computer and Communication Sciences at EPFL (epfl.ch) and the founding President of Swiss Datacenter Efficiency Association (sdea.ch).

Disclaimer: These posts are written by individual contributors to share their thoughts on the Computer Architecture Today blog for the benefit of the community. Any views or opinions represented in this blog are personal, belong solely to the blog author and do not represent those of ACM SIGARCH or its parent organization, ACM.