Learning the Grid: AI and the Future of Infrastructure Management

<p class="dropcap">In the spring of 2024, Pacific Gas &amp; Electric filed an interconnection request for a 480-megawatt data center campus proposed for Contra Costa County, California &mdash; a load figure that, a decade ago, would have described a mid-sized city substation, not a single commercial tenant. The filing sat in a queue alongside 40,000 other projects totaling more than 2,000 gigawatts of requested capacity nationwide, according to Lawrence Berkeley National Laboratory's 2024 grid interconnection study. That backlog, measured in years and billions of dollars of deferred capital, has become the defining constraint of American infrastructure in the AI era. Utilities, developers, and federal regulators are now racing to answer a question that no one adequately planned for: can the grid learn as fast as the machines running on it?</p>

<h3>The Colocation Equation</h3>

<p>The most direct answer the industry has produced is geographic compression &mdash; placing data centers as close as possible to generation sources, eliminating transmission distance and the attendant losses, permitting delays, and right-of-way costs. In Massena, New York, a former aluminum smelter site on the St. Lawrence River has been reborn as a data center corridor, with tenants including Corelink and several hyperscale operators drawn by access to New York Power Authority hydroelectric generation that feeds directly into the local 115-kilovolt network. The arrangement sidesteps the congested Consolidated Edison territory downstate entirely. Records filed with the New York Public Service Commission show the Massena industrial zone absorbed roughly $340 million in new data center investment between 2021 and 2024, with load additions approaching 600 megawatts across multiple parcels.</p>

<p>The colocation logic extends to fossil-fuel retirement sites as well. In Joliet, Illinois, NRG Energy's former coal plant on the Des Plaines River &mdash; a 1,600-acre brownfield with existing 345-kilovolt transmission infrastructure &mdash; has attracted serious acquisition interest from data center developers, according to Cook County property records reviewed by Cornice. The appeal is structural: decommissioned generation sites arrive with switchgear, cooling water rights, and grid interconnection already in place. Adaptive reuse of that infrastructure can shave 18 to 36 months from a typical greenfield interconnection timeline, according to Grid Strategies LLC, a Washington-based transmission consultancy whose 2023 analysis was cited in Federal Energy Regulatory Commission proceedings. That timeline compression translates directly to capital efficiency in an industry where a single month of delayed rack deployment costs hyperscale operators tens of millions in deferred revenue.</p>

<p>Critics of the colocation model argue it concentrates grid risk rather than distributing it. Ari Peskoe, director of the Electricity Law Initiative at Harvard Law School, has noted in public filings that large single-tenant loads can destabilize local distribution networks not designed for the demand profile of AI inference workloads &mdash; which spike sharply during model training runs and then drop to near-idle, creating voltage fluctuations that stress transformers serving adjacent residential and commercial customers. "The grid was not engineered for a tenant class that can double its draw in four minutes," Peskoe wrote in comments submitted to FERC Docket No. RM21-17.</p>

<h3>The Nuclear Recalculation</h3>

<p>No infrastructure story in 2024 moved faster than the rehabilitation of nuclear power as the preferred baseload solution for AI-hungry data centers. In September 2024, Microsoft signed a 20-year power purchase agreement with Constellation Energy to restart Unit 1 of the Three Mile Island nuclear station in Dauphin County, Pennsylvania &mdash; the same plant whose undamaged reactor was shuttered in 2019 on purely economic grounds. The deal, valued at an estimated $1.6 billion over its term according to Constellation's investor filings, will deliver 835 megawatts of carbon-free generation to PJM Interconnection's grid beginning in 2028, with Microsoft holding priority offtake rights under a structure that effectively backstops Constellation's restart capital expenditure.</p>

<p>The Three Mile Island transaction was not an isolated event. In October 2024, Google announced a partnership with Kairos Power to deploy small modular reactors at undisclosed sites, targeting 500 megawatts of SMR capacity by 2035. Amazon Web Services, through its subsidiary Amazon Energy, purchased the Susquehanna Steam Electric Station's entire 2.5-gigawatt output for a 10-year term in a deal that drew scrutiny from Pennsylvania's Public Utility Commission, which questioned whether the arrangement effectively removed baseload capacity from the regional market. The PUC's investigation, opened in November 2024, remains pending. What unites these transactions is the recognition that wind and solar &mdash; however cost-competitive on a levelized basis &mdash; cannot supply the 24/7 firm power that AI training infrastructure demands without storage solutions that do not yet exist at grid scale.</p>

<p>The federal government has moved to accelerate the nuclear recalculation. The Department of Energy's Loan Programs Office, under Director Jigar Shah, committed $1.52 billion in conditional loan guarantees to restart Holtec International's Palisades Nuclear Plant in Van Buren County, Michigan in 2023 &mdash; the first time in American history that a fully decommissioned commercial reactor was approved for restart. Construction crews returned to the site in early 2024. The Palisades project is explicitly framed in DOE filings as a "clean energy anchor" for regional industrial and data center development, language that signals the federal government's willingness to use nuclear restart as an infrastructure policy instrument rather than purely an energy policy one.</p>

<h3>Surplus as Asset: The Community Power Inversion</h3>

<p>The conventional narrative frames data centers as parasitic loads &mdash; extracting power from communities while returning little beyond tax revenue and a handful of jobs. The emerging reality in several markets is more complicated. In Cheyenne, Wyoming, Microsoft's 400-acre data center campus, which draws from a dedicated 300-megawatt substation built in partnership with Rocky Mountain Power, has generated enough contracted generation capacity that the utility has been able to defer planned rate increases for residential customers in Laramie County, according to Wyoming Public Service Commission rate filings from 2023. The mechanism is counterintuitive: a large anchor tenant with a flat, predictable load profile improves the economics of generation investment, reducing per-unit fixed costs across the rate base.</p>

<p>A more deliberate version of this dynamic is being engineered in Quincy, Washington, where Grant County Public Utility District has structured long-term data center contracts &mdash; with tenants including Sabey Data Centers and a Microsoft campus exceeding 1.2 million square feet &mdash; to cross-subsidize agricultural irrigation pumping rates for the Columbia Basin farming community. Grant County PUD's 2023 annual report shows the district's average commercial rate of 2.1 cents per kilowatt-hour, among the lowest in the nation, is sustained in part by the margin earned on data center contracts priced at a premium to that base rate. The arrangement is not charity; it is load-balancing economics. Data centers consume power most intensively during off-peak hours, smoothing the duck curve that plagues utilities with high solar penetration.</p>

<p>Whether this surplus-as-asset model scales beyond markets with abundant hydroelectric resources remains genuinely uncertain. In Northern Virginia's Loudoun County &mdash; which hosts more data center square footage than any other jurisdiction on earth, with approximately 35 million square feet of raised-floor space as of 2024 &mdash; Dominion Energy has been forced to propose $7.6 billion in new transmission infrastructure over a 15-year capital plan, costs that will be socialized across Virginia ratepayers regardless of whether those ratepayers benefit from the data centers generating the demand. The Virginia State Corporation Commission approved Dominion's 2023 Integrated Resource Plan with explicit reservations about cost allocation, a tension that state legislators have yet to resolve.</p>

<h3>Portable Power and the Edge Frontier</h3>

<p>At the furthest remove from the hyperscale campus model sits a class of infrastructure that received almost no serious attention until the Department of Defense began procuring it in volume: mobile nuclear microreactors. In 2022, the Defense Advanced Research Projects Agency awarded contracts under its Pele program to BWX Technologies and X-energy to develop truck-transportable reactors in the 1-to-5-megawatt range, designed to be airlifted to forward operating bases and remote installations. In a demonstration that crossed from military into civilian infrastructure discourse, the Air Force Materiel Command conducted a logistics exercise in 2023 airlifting a diesel generator of comparable physical dimensions to validate transport protocols &mdash; a proof-of-concept that the Pele program's contractors cited in subsequent briefings to the Senate Armed Services Committee.</p>

<p>The civilian implications are not hypothetical. The Alaska Energy Authority has formally engaged with the Nuclear Regulatory Commission regarding microreactor deployment in rural Alaskan communities currently served by diesel generation at costs exceeding $0.75 per kilowatt-hour &mdash; a price point at which almost any alternative technology achieves immediate payback. The NRC's draft regulatory framework for mobile reactors, published for comment in late 2023, anticipates deployment scenarios including remote mining operations, disaster recovery staging areas, and &mdash; critically &mdash; edge data center nodes in locations where grid interconnection is economically or physically impractical.</p>

<p>The convergence of AI's computational demands with the physical realities of American infrastructure &mdash; aging transmission, constrained interconnection queues, and communities that cannot wait decades for grid expansion &mdash; is forcing a genuine rethinking of how power and processing are co-located, financed, and governed. The answers emerging from Massena and Quincy and Dauphin County are provisional, market-driven, and often legally contested. What they are not is theoretical. The grid is already learning, unevenly and under pressure, and the institutions charged with managing it &mdash; FERC, state utility commissions, the DOE Loan Programs Office &mdash; are writing the rules in real time, against a demand curve that has not paused to let them catch up.</p>