A new industry framework aims to resolve discrepancies in data center energy consumption estimates.

A perspective article published in Engineering addresses widespread uncertainty in global data center energy accounting, proposing integrated technical and policy solutions to standardize energy-use tracking amid rapidly expanding cloud and large-scale artificial intelligence infrastructure. The article opens by highlighting severe inconsistencies in existing global electricity consumption assessments for data centers, noting that 2020 estimates range from 196 to 1,200 TWh—a disparity exceeding a factor of six. The authors argue this massive variance stems from structural flaws within two mainstream energy calculation paradigms, which together distort carbon accounting, hinder power grid planning, and slow renewable energy integration as computing loads rise. The research categorizes conventional estimation methods into bottom-up and top-down frameworks, each carrying distinct inherent limitations that undermine statistical reliability.

Bottom-up modeling calculates total energy consumption using facility-level parameters such as rack density, server utilization, and power usage effectiveness (PUE). However, it suffers from broad parameter volatility across sites, relies on lab-biased efficiency benchmarks like SPECpower_ssj2008 that fail to reflect real-world heterogeneous workloads, and depends on opaque, self-reported enterprise operational data prone to underreporting without independent verification. Conversely, top-down modeling draws on aggregated national or utility electricity statistics for macro-sector evaluation. Yet it cannot isolate data center loads from mixed-use commercial premises due to the absence of a dedicated industrial classification within official ICT energy datasets, leaving embedded small and medium-sized facilities uncounted in official tallies.

To resolve blind spots surrounding existing unregistered data centers, the article outlines AI-driven identification workflows paired with non-intrusive load monitoring (NILM). Machine learning architectures—including LSTM networks, Transformers, random forests, and support vector machine classifiers—extract unique 24/7 continuous load signatures from aggregated meter readings, effectively distinguishing data center consumption from the variable patterns of offices, retail spaces, and residential buildings. These classification models deliver high accuracy in isolating data center workloads. NILM further disaggregates mixed building electricity signals to pinpoint facility-level power draw without requiring physical submeter retrofits, with periodic manual audit validation used to stabilize model performance under noisy, low-resolution grid data.

For newly constructed computing hubs, the study advocates grid-informed mandatory energy registration as a foundational policy pillar, contrasting implementation conditions between centralized grid jurisdictions like China and decentralized markets across the United States and European Union. Fragmented cross-agency oversight currently delays standardized energy data sharing with grid operators during project approvals. Meanwhile, uneven regional regulatory rules across Europe have created inconsistent disclosure standards, even following the European Commission’s introduction of binding 2024 sustainability reporting mandates for facilities above 500 kW. The article recommends that registration records include standardized technical descriptors, uniform building-use taxonomies, and direct linkage to grid smart meter time-series data. Embedding these classification requirements within routine grid connection applications would streamline data collection without imposing excessive operational burdens on operators.

Beyond accurate energy tallying, unified registration unlocks the temporal and spatial load flexibility inherent to AI and high-performance computing data centers. Batch model training workloads allow for time shifting, while cross-location task migration aligns computing demand with intermittent renewable generation. Formalized energy statistics enable demand-response incentive schemes, reducing operational power costs for facility operators and lowering grid storage and scheduling overheads associated with variable wind and solar supply. The article concludes with cross-stakeholder policy recommendations, calling for aligned interdepartmental reporting standards, targeted financial support for AI monitoring tools and grid infrastructure upgrades, and long-term market incentives to encourage full energy transparency and voluntary electricity market participation. Ultimately, these measures aim to reposition data centers from passive power consumers to active grid stability contributors amid global low-carbon transitions.

The paper "Building Accurate Energy-Use Statistics for Data Centers," is authored by Yong-Zhen Wang, Te Han, Yi-Ming Wei. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.12.014