Data centers are becoming one of the most difficult parts of the global energy system to measure, just as cloud computing, artificial intelligence and high-performance computing are driving demand for electricity to unprecedented levels. A perspective article published in Engineering argues that the world cannot manage the climate and grid consequences of this expansion without first establishing a reliable way to identify, register and track data-center electricity use. The authors, Yong-Zhen Wang, Te Han and Yi-Ming Wei, say that current accounting methods produce estimates so inconsistent that governments, utilities and researchers may be planning for fundamentally different versions of the same digital infrastructure boom.
The scale of the uncertainty is revealed by estimates for global data-center electricity consumption in 2020, which range from approximately 196 to 1,200 terawatt-hours. The difference is greater than sixfold, an extraordinary gap for an industry whose energy demand is increasingly influencing national power strategies. Such uncertainty affects more than academic statistics. It can distort carbon-emissions inventories, complicate electricity-grid expansion, weaken forecasts for renewable-energy demand and make it difficult to determine whether new computing facilities are being supplied by genuinely low-carbon power. As artificial-intelligence systems require increasingly intensive training and deployment, the authors warn that unreliable energy data could become a structural obstacle to effective climate policy.
The article examines the two dominant approaches used to estimate data-center energy consumption: bottom-up and top-down accounting. Bottom-up methods attempt to calculate total demand from technical characteristics such as the number of servers, rack power density, server utilization and power usage effectiveness, or PUE. PUE compares the total energy consumed by a facility with the energy used directly by computing equipment, providing an indicator of how much power is required for cooling, power conversion, lighting and other support systems. Although useful, bottom-up calculations can vary dramatically because these parameters differ from one facility to another and are often unavailable at sufficient detail. A modern hyperscale center, a small enterprise server room and an AI-focused computing campus may have completely different operating profiles, cooling systems and utilization rates.
The reliability of bottom-up estimates is also weakened when laboratory performance metrics are treated as if they represented actual commercial operations. Benchmarks such as SPECpower_ssj2008 can help compare equipment under controlled conditions, but they do not necessarily capture the irregular and heterogeneous workloads found in real data centers. AI model training, cloud services, video processing, storage and conventional enterprise applications can place very different demands on processors, memory, networking and cooling equipment. In addition, the authors point to the limited transparency of self-reported operational data. Companies may calculate energy use according to different boundaries, definitions or reporting schedules, while independent verification is often absent. Even small differences in methodology can produce large errors when aggregated across thousands of facilities.
Top-down accounting approaches the problem from the opposite direction. Instead of estimating each facility, researchers examine national, regional or utility-level electricity statistics and attempt to identify the share associated with data centers and information-communication technologies. This approach can be valuable for understanding broad energy trends, but it struggles to distinguish data centers from other commercial activities. Official energy classifications in many jurisdictions do not provide a dedicated category for every type of computing facility, particularly smaller centers embedded in office buildings, retail complexes, industrial sites or mixed-use properties. Their consumption can therefore disappear inside broader commercial totals. The result is a statistical blind spot in which large, visible campuses may be counted while dispersed or unregistered facilities remain effectively invisible.
To address these gaps, Wang and colleagues propose combining artificial intelligence with non-intrusive load monitoring, or NILM. The central idea is that different types of buildings produce recognizable electricity-use signatures. A data center typically operates continuously, with relatively stable baseline demand and distinctive changes associated with computing activity, cooling systems and backup infrastructure. By contrast, offices, shops and homes generally display more variable schedules and appliance-driven patterns. Machine-learning models can examine aggregated electricity readings and classify these patterns without requiring a separate physical meter for every building. The approach could allow utilities and researchers to identify previously unregistered data centers using information already collected at the grid or building level.
The proposed workflow includes long short-term memory networks, Transformers, deep neural networks, random forests and support vector machines. LSTM networks are designed to recognize relationships across time, making them suitable for detecting recurring daily and weekly load behavior. Transformers can analyze longer-range dependencies and complex interactions within time-series data, while random forests and support vector machines can classify patterns using engineered features such as load variability, persistence, peak timing and ramp rates. NILM then attempts to disaggregate the total electricity signal, estimating how much power is being consumed by individual systems or facilities within a larger aggregate. Because real-world grid data may have low temporal resolution, missing values and measurement noise, the authors recommend periodic manual audits to validate classifications and recalibrate models.
The implications extend beyond measurement. The article proposes that new computing facilities should be subject to mandatory energy registration during the grid-connection and project-approval process. Registration records could include standardized descriptions of building use, computing capacity, technical equipment, cooling systems and expected operating schedules, while also linking facilities to smart-meter time-series data. The authors note that implementation will differ across political and regulatory systems. China’s more centralized grid structure may offer a direct route for coordinated data collection, whereas the United States and European Union face more fragmented markets and overlapping institutional responsibilities. Across Europe, regional differences in regulation can lead to inconsistent reporting, even as the European Commission’s 2024 sustainability requirements introduce binding disclosure obligations for facilities above 500 kilowatts.
A unified registration system could also transform data centers from passive electricity consumers into active participants in grid management. AI and high-performance-computing workloads are not all equally time-sensitive. Batch activities, including some model-training operations, can be shifted to periods when electricity is cheaper or renewable generation is abundant. Workloads may also be moved between facilities in different regions, allowing operators to respond to local variations in wind, solar output or grid congestion. Accurate energy statistics would give utilities the information needed to design demand-response programs and would allow data-center operators to receive incentives for adjusting consumption. In turn, flexible computing demand could reduce the need for costly storage and help grids absorb larger amounts of intermittent renewable power.
The authors ultimately call for common reporting standards across government departments, financial support for AI-based monitoring and upgrades to digital grid infrastructure. They argue that energy transparency should become a long-term market expectation rather than a voluntary exercise dependent on individual companies. Without consistent definitions, independently verifiable data and systematic identification of hidden facilities, the electricity footprint of the digital economy will remain uncertain at precisely the moment when it is expanding fastest. By pairing technical tools such as machine learning and NILM with mandatory registration and coordinated policy, the proposed framework seeks to make data-center growth visible, measurable and compatible with the global transition toward lower-carbon energy.
Subject of Research: Data-center energy consumption accounting, artificial intelligence-based load identification, non-intrusive load monitoring and grid flexibility.
Article Title: “Building Accurate Energy-Use Statistics for Data Centers”
Web References: https://doi.org/10.1016/j.eng.2025.12.014 ; https://www.sciencedirect.com/journal/engineering
References: Yong-Zhen Wang, Te Han and Yi-Ming Wei, Engineering.
Image Credits: Yong-Zhen Wang, Te Han and Yi-Ming Wei
Keywords: Data centers, artificial intelligence, energy consumption, electricity grids, non-intrusive load monitoring, machine learning, renewable energy, demand response, carbon accounting, data-center registration
