AI Energy Elasticity and Data Center Energy Consumption Growth Scenarios to 2030

The article introduces the concept of "AI energy elasticity" as the dimensionless sensitivity of data centers' total power consumption to changes in the volume of services based on artificial intelligence algorithms. A formal decomposition is proposed: E = E_IT κ_infr, where the IT component is determined by the training and inference profile, and the infrastructure multiplier is determined by the efficiency of cooling, distribution, and power conversion. Against the backdrop of the accelerated adoption of accelerators with high thermal packages and the transition to direct liquid cooling, three scenarios are assessed for 2030: moderate, baseline, and boosted.

The methodological framework draws on Koomey's law for long-term reduction in energy per operation, systemic assessments of data center energy efficiency (Eric Mazanet, Arman Shehabi), work on the contrast between training and inference energy costs (Emma Strubell et al.), energy reporting protocols in machine learning (Peter Henderson), and recommendations on how to achieve savings through the choice of model, hardware, and deployment region (David Patterson et al.). It is shown that in the short term, energy elasticity remains elevated due to stagnation in the average market energy efficiency ratio and a structural shift towards AI workloads; in the long term, it decreases due to algorithmic and engineering improvements. Practical implications include standardization of measurements, mandatory reporting, spatiotemporal optimization of training, and proactive modernization of network infrastructure.

Regional constraints are separately noted: shortages of connected capacity and water increase the importance of load flexibility and useful heat removal.