Evaluating U.S. Locations for AI Data Centers: A New Framework for Energy Capacity

Energy Constraints as Architectural Bottlenecks: RAND's Site Suitability Framework for AI Data Centers

The analysis addresses a structural question for machine learning infrastructure: where electricity can physically be supplied, not merely how much additional capacity can be added nationally. Sites are stratified into high, moderate, and low energy-potential tiers, with no single location identified as universally optimal.

Framework Design

The methodology decomposes site energy potential into five categories encompassing existing infrastructure, enabling factors, and barriers. The categories collectively capture whether power can be delivered reliably and at scale by the stated horizon. A total of 22 sites were evaluated: 17 identified by the U.S. Department of Energy, two major private-sector sites, and three retired or retiring power plant sites.

Assessment proceeded through a multicriteria decision-making process rather than a single-metric ranking. Sites scoring well across the five categories combine stronger transmission access, available land, favorable regulatory conditions, and proximity to energy resources. High-potential locations cited in the analysis include Stargate, the Pantex Plant, and the Kansas City National Security Campus. These sites generally benefit from pre-existing transmission infrastructure, available land, and favorable regulatory conditions.

Results and Limitations

The moderate-potential tier contains the majority of assessed sites. These locations exhibit promising characteristics alongside one or more constraints — transmission bottlenecks, permitting timelines, or insufficient local generation — that would require targeted intervention to reach frontier-scale AI deployment. The framework does not assign a single composite score; it produces categorical placement intended as a screening tool for planners, policymakers, and developers.

The study explicitly limits its scope to energy suitability. Water availability, cooling topology, networking latency, and compute density are not modeled. The framework is positioned as an energy-potential screen rather than a holistic site recommendation. Researchers note that frontier AI data center site selection will increasingly depend on existing energy infrastructure, local system conditions, and other place-specific constraints — a recognition that grid physics, not chip roadmaps, will determine where scaling is feasible.

Practical Signals for ML Teams

For practitioners planning infrastructure spend or training capacity, the relevant variables are now site-level: transmission headroom, interconnection queue position, and local regulatory friction. The framework offers a reproducible template for filtering candidate locations before deeper engineering analysis. Teams evaluating colocated training facilities or hyperscaler partnerships should track DOE site designations and interconnection queue disclosures as primary signals of where 2030 compute capacity can actually land. Source PDF available via RAND Corporation for full category definitions and site-by-site scoring.