Our World in Data Quantifies AI Energy Use

For practitioners, electricity use is not reducible to a per-token or per-query estimate. Editorial analysis: comparable AI infrastructure assessments need to distinguish IT load from facility overhead, account for the location and time of consumption, and separate training from ongoing inference. These boundaries affect capacity planning, carbon accounting, and the interpretation of efficiency claims.

Our World in Data reports that data centers used around 485 TWh of electricity last year, citing the International Energy Agency, equivalent to roughly 1.5% of global electricity generation. The article characterizes demand as geographically concentrated. Its accounting includes server electricity and supporting facility loads such as cooling and lighting, while excluding the electricity used by end-user devices and excluding cryptocurrency mining.

The article notes that technology companies rarely publish model-training energy data. Based on the estimates available, Our World in Data writes that AI electricity demand is likely dominated by inference rather than training.

According to the IEA's Energy and AI analysis, servers account for around 60% of electricity demand in modern data centers on average, although the share varies by facility type. The IEA assigns around 5% to storage systems and up to 5% to networking equipment, while cooling and environmental controls range from about 7% of consumption at efficient hyperscale facilities to more than 30% at less-efficient enterprise data centers.

The IEA describes data centers as facilities containing servers, storage, networking equipment, and supporting systems. It identifies CPUs and specialized accelerators such as GPUs as server components relevant to electricity demand. UPS batteries and backup generators contribute to reliability requirements, though the IEA notes that they are rarely used.