Assessing the Energy Potential of Artificial Intelligence Data Center Sites

Artificial intelligence (AI) development is rapidly increasing electricity demand in the United States. Developers and operators of large AI data centers generally prefer grid connections because they offer greater reliability than on-site power generation alone. The challenge is no longer only how much additional power the United States might add by 2030: It is also where that power can realistically be delivered. For frontier AI data centers, site selection will increasingly depend on existing energy infrastructure, local system conditions, and other place-specific constraints.

To help address this challenge, RAND researchers examined two questions: (1) Which sites or regions have favorable or unfavorable characteristics for maximizing energy capacity for AI data centers, and (2) what is the maximum amount of electric power that the grid can deliver at a single site by 2030? To answer these questions, the researchers developed a framework for comparing site energy suitability and applied it to 22 locations: 17 sites identified by the U.S. Department of Energy (DOE), two major privatesector sites, and three retired or retiring power plant sites. The framework is intended to serve as an energypotential screening tool for planners, policymakers, and developers.

The Framework

The framework assesses sites across five categories that together shape how favorable a location is, from an energy perspective, for AI data center development. These categories are intended to capture whether energy can likely be delivered reliably and at scale by 2030 at the site. Because this analysis focuses on site suitability from an energy perspective, the framework emphasizes factors that affect whether power can be supplied, delivered, and supported at scale. Table 1 summarizes the five categories used to assess those energy-related conditions across candidate sites.

These categories are grounded in a broad conceptual view of how site characteristics influence available energy capacity. Figure 1 shows how infrastructure, enabling factors, and barriers combine to shape the five categories used in the assessment.

Infrastructure

• Energy generation
• Energy delivery

Enablers

• Regulatory environment
• Physical resources
• Infrastructure and workforce
• Community and environment
• Regional industry clusters

Barriers

• Delayed permitting
• Inefficient and costly interconnection
• Underutilized transmission capacity
• Lack of incentives to adopt technology for backup generators

Framework Assessment Categories

• Energy supply
• Energy system
• Supporting inputs
• Environmental considerations
• Governance and community considerations

NOTE: This framework consists of the desirable site characteristics, as can be seen in terms of existing infrastructure, enablers, and barriers. The assessment involved a multicriteria decisionmaking process to assess potential sites against five categories that encompass infrastructure, enablers, and barriers. Sites that score well on the framework should have access to multiple enablers and encounter minimal barriers.

Some Sites Stand Out, but No Site Is Perfect

The researchers did not identify one universally best site. Instead, sites fell into broad tiers of high, moderate, and low energy potential.

High-potential sites tended to combine stronger existing infrastructure, favorable supporting inputs, and fewer major bottlenecks. Examples included Stargate, the Pantex Plant, and the Kansas City National Security Campus. These locations generally benefited from some combination of existing transmission access, available land, favorable regulatory conditions, and nearby energy resources.

Many sites fell into the moderate-potential tier. These locations often had promising characteristics but also faced one or more constraints that could limit development unless addressed through targeted action. Examples included Colossus, Oak Ridge National Laboratory, Los Alamos National Laboratory, and retired or retiring plant sites, such as Plant Wansley and the Cumberland Fossil Plant.

Low-potential sites often faced multiple constraints at once, especially around grid conditions, environmental risks, or governance and community factors. More generally, the analysis suggests that site selection for AI infrastructure is best understood as a process of managing trade-offs rather than finding a perfect location.

Applying the framework across the study sites showed that relative site suitability varies considerably and depends on trade-offs across energy, infrastructure, environmental, and governance factors.

Table 2 summarizes how the assessed sites were grouped into high-, moderate-, and low-potential tiers and highlights the trade-offs across the five energy-related assessment categories.

Existing Infrastructure Offers a Major Advantage

A central finding of the research is that using existing infrastructure offers a substantial advantage for meeting the 2030 timeline. Building major new T&D infrastructure is difficult to complete on that schedule because of permitting requirements, regulatory hurdles, supply chain challenges, and construction delays.

This makes previously developed sites especially attractive. Former industrial campuses, federal facilities, and retired or retiring power plant sites can offer existing substations, transmission connections, and local distribution infrastructure that reduce the time and cost needed to support large new loads. In some cases, these sites may also have cleared land, industrial zoning, and legacy interconnection arrangements that make them more viable than undeveloped alternatives.

How Much Power Could One Site Support by 2030?

To estimate the maximum amount of power that might realistically be available at a single site by 2030, the researchers assessed three retired or retiring power plant locations. These sites were selected because they already benefit from substantial T&D infrastructure and offer opportunities for repowering or reuse.

Among them, Rockport Power Plant in Indiana had the highest estimated potential, at about 4.2 gigawatts. This estimate reflects both the site’s strong existing transmission infrastructure and the possibility of using or repurposing power generation assets at the location.

Given current constraints on building major T&D infrastructure before 2030, this estimate likely represents the practical upper bound of grid-connected power that could be made available at a single U.S. site on that timeline. This finding suggests that the U.S. power system could struggle to support the largest projected centralized frontier AI data centers without major infrastructure reuse, significant on-site power generation, or a more distributed compute strategy.

Policy Implications

Expanding AI data center capacity by 2030 will require more than identifying available land and power in the abstract. It will require focusing on the specific conditions that make some sites more feasible than others. RAND researchers suggest several priorities:

• Target infrastructure improvements and investments. Address site-level bottlenecks, such as a lack of substations, T&D upgrades, or interconnection constraints.
• Prioritize sites with reusable infrastructure. Especially important are industrial campuses, former manufacturing facilities, and retired or retiring power plants.
• Account for environmental and reliability risks early. These include water stress, natural disaster exposure, and cooling burdens.
• Build community buy-in. Early and transparent engagement can reduce opposition, litigation risk, and permitting delays.
• Use an energy screening framework, such as the one presented in this brief, for site selection. Policymakers and developers can apply the framework to candidate sites, preserve favorable site characteristics, and focus remediation efforts on addressable weaknesses.

Available for Download

Download PDF

Topics

• Artificial Intelligence
• Energy Distribution
• Energy Resources
• Public Utilities