Rare-Earth Exposure and Bottlenecks in AI Data Centers: Cost and Schedule Risk

<p><span>AI data centers (AIDCs) are often described as electricity-hungry assets. This study treats AI DCs as gigawatt-scale capital builds and focuses on how a small set of REE-bound subcomponents enter the campus cost stack yet exert outsized leverage on delivery timelines. Generic criticality rankings and sector-level demand narratives rarely locate this dependence inside the campus cost stack, where procurement decisions and lead times are actually managed.</span></p> <p><span>An auditable bridge is developed from build economics to element-level exposure. A per-GW AI-campus CAPEX baseline is decomposed into a transparent taxonomy of buckets and subcomponents, then mapped to REE-bound components using documented physical mechanisms. Bounded cost-intensity parameters and a 0–2 schedule-at-risk score are used to discipline uncertainty and separate cost exposure from schedule vulnerability. These inputs are consolidated into a traceable Rare-earth Exposure (REX) metric and extended with forward-looking scenarios that jointly evolve CAPEX mix and intensity parameters under plausible architecture shifts and substitution pathways.</span></p> <p><span>In the baseline (mid) case, total mapped exposure is ≈ US$158.0M/GW, concentrated in IT equipment (≈ US$149.2M/GW; 94.4%) with a smaller cooling contribution (≈ US$8.8M/GW; 5.6%). Servers and accelerators account for ≈ US$138.8M/GW (87.8%), dominated by Nd (≈ US$83.3M/GW) and Pr (≈ US$34.7M/GW), with heavy-REE add-ons (Dy ≈ US$16.7M/GW; Tb ≈ US$4.2M/GW). Normalized, Nd exposure is 17.49 bp of total CAPEX. Cost share and delivery risk need not align: schedule-weighted exposure can be driven by long-lead, qualification-sensitive components. Scenario results show that design evolution reallocates exposure across elements and channels, revealing both mitigating and amplifying pathways that static assessments miss.</span></p>