Power Grid Infrastructure for AI Data Centers

Abstract

The paper examines how the rapid buildout of large AI data centers is reshaping power-grid planning and operation in the United States. It treats the AI compute boom as an unprecedented, geographically concentrated load-growth shock and asks what specifically bottlenecks the grid's ability to absorb it — generation capacity, transmission capacity, or the interconnection process itself.

Key Contributions

• Analyzes interconnection-queue backlogs directly from FERC filings, finding thousands of MW of proposed capacity awaiting approval and connection timelines now extending 5-10+ years in the most congested regions.
• Quantifies regional transmission deficits: existing transmission infrastructure cannot accommodate projected AI compute expansion, with the most severe constraints in regions that already carry heavy industrial load.
• Produces capacity-gap estimates of 50-150+ GW regionally by 2030-2035 — i.e., the shortfall between what AI data center developers want to interconnect and what the grid can deliver, region by region.
• Frames peak-demand scenarios in which simultaneous AI facility requests within a single region exceed the grid's available capacity outright, not just its queue-processing speed.

Methodology

Interconnection-queue data pulled from FERC filings; regional transmission organization (RTO) capacity assessments; growth projections built from publicly announced AI facility development plans; comparative infrastructure modeling across North American grid regions.

Results

1. Grid bottlenecks precede generation constraints — transmission and interconnection process capacity are the near-term binding constraint, ahead of whether enough megawatts of generation exist system-wide.
2. Regional disparities are acute — some regions face interconnection delays 2-3x longer than others, meaning the constraint is not uniform and site selection materially changes a developer's time-to-power.
3. Coordinated, cross-entity infrastructure investment is required — the paper argues single-utility or single-developer fixes are insufficient at the scale AI load growth demands.

Limitations

• Projections are conditional on assumed AI adoption/compute-growth rates, which the authors flag as inherently volatile.
• Policy responses (permitting reform, interconnection-process reform) remain uncertain and are not modeled as a resolving variable.
• Regional variance in the underlying data limits how far the capacity-gap figures generalize across grid regions with different topologies and regulatory regimes.

Full Content

This paper is squarely a grid-planning/power-systems analysis of the AI data center buildout, not a chemistry or hardware paper — it treats AI compute growth as an electrical-load-forecasting and transmission-planning problem. Its central empirical claim is that the binding constraint on data center growth through 2030-2035 is not "not enough power plants" but "not enough transmission and interconnection-processing capacity to deliver power that likely does exist, to where it's needed, fast enough." The 50-150+ GW regional gap figure and the 5-10+ year interconnection-queue timelines are the load-bearing numbers; both are drawn from FERC-filing-level data rather than industry survey sentiment, which gives them more evidentiary weight than vendor-survey-based estimates of the same phenomenon (compare the Bloom Energy 2026 Data Center Power Report's developer/utility time-to-power gap, ingested the same day, which corroborates this from the demand side).

Source: Power Grid Infrastructure for AI Data Centers (arXiv:2606.00941) by Sajadi, Za'ter, Vabson, Baker, Hodge. Submitted 2026-05-31.