The Hidden Energy Cost of AI: Analysing the Relationship Between AI Workloads and Computing Energy Consumption

Artificial intelligence (AI) systems can require substantial computing resources, yet reported energy estimates are not always directly comparable because studies differ in measurement methods, workload definitions, system boundaries, and reporting practices. This structured narrative review examines how AI workloads relate to computing energy consumption and synthesises eight related research-gap themes: inconsistent measurement and uncertainty; limited comparability between training and inference; incomplete workload and hardware reporting; insufficient joint consideration of energy and model performance; incomplete lifecycle accounting; limited comparability across benchmarks and environments; incomplete accounting for embodied hardware impacts and non-energy environmental burdens; and insufficient treatment of electricity carbon intensity across time and location. The review draws on selected published studies, technical documentation, and benchmark resources. It is not presented as a completed systematic review: a full database search, screening log, and study-level extraction have not yet been completed, so no exhaustive-coverage claim or pooled numerical result is made. The synthesis highlights the importance of clearly defined functional units, measurement boundaries, hardware and software descriptions, performance context, uncertainty, and lifecycle scope. It proposes an expanded reporting checklist and a reproducible review protocol to support more transparent comparisons in future research.