Mineral demand from AI data centers: Infrastructure intensity, processing bottlenecks, and supply competition

Abstract

This paper demonstrates that the Artificial Intelligence (AI) buildout's demand for critical materials is driven primarily by bulk infrastructure requirements, using a bottom-up model to quantify this effect across 20 materials through 2035. The framework distinguishes compute, thermal management, and power infrastructure; separates training, inference, and legacy fleets; and incorporates technology-adoption dynamics for batteries, storage media, and fleet composition. The results show that copper dominates total modeled mass, with grid transmission and distribution accounting for the largest share of demand. Separately, grain-oriented electrical steel (GOES) emerges as a particularly constrained enabling material. Beyond tonnage pressure in copper and GOES, the analysis identifies processing-stage vulnerability as the main competitive risk for several lower-volume materials, including gallium, germanium, graphite, lithium, cobalt, and rare earths. The paper concludes that by-product recovery from existing U.S. mining operations offers one near-term pathway for reducing exposure to concentrated supply chains.