Microsoft has locked in a 20-year power purchase agreement with Constellation Energy tied to the restart of Three Mile Island Unit 1, a deal now tracked in federal regulatory filings as AI-driven electricity demand pushes major tech companies toward nuclear energy contracts. The arrangement, documented in a U.S. Nuclear Regulatory Commission filing, reflects a broader pattern: hyperscalers are bypassing traditional grid procurement and signing long-term nuclear deals to guarantee the carbon-free baseload power their expanding data centers require. The International Energy Agency has separately published analysis showing that energy demand from AI is growing faster than total electricity demand under several scenarios, raising hard questions about whether existing grid infrastructure can keep pace.
AI power demand is outpacing grid expansion
The gap between what AI data centers need and what the grid can deliver is widening. Training large language models and running inference workloads at scale requires steady, high-capacity power, and the construction timelines for new transmission lines and generation assets are measured in years, not months. The IEA’s AI demand analysis frames data-center electricity consumption as a fast-growing share of total demand, with multiple scenarios showing AI loads climbing well beyond what incremental renewable additions can absorb on their own.
That mismatch helps explain why the largest U.S. cloud and AI companies are turning to nuclear power. Solar and wind projects face long interconnection queues, and battery storage alone cannot yet replicate the round-the-clock output of a nuclear plant. A nuclear reactor delivers consistent megawatts regardless of weather or time of day, making it a natural fit for data centers that run at near-full load continuously. The commercial logic is straightforward: if a hyperscaler needs guaranteed power for a training cluster scheduled to come online by 2027 or 2028, a signed nuclear PPA offers more scheduling certainty than waiting for a renewable project to clear regulatory and grid-connection hurdles.
This dynamic points to a testable idea. The volume of new nuclear power purchase agreements signed by the five largest U.S. hyperscalers may end up correlating more closely with the pace of AI training-cluster deployments than with total renewable interconnection capacity added over the same period. Renewable capacity is growing, but the queue backlogs and permitting delays mean that much of it arrives too late to match the aggressive buildout schedules tech companies are setting for their AI infrastructure.
Constellation, Microsoft, and the Three Mile Island restart
The most concrete example of this trend sits in central Pennsylvania. Constellation Energy has stated its intent to restore Three Mile Island Unit 1 to service, a plan now referenced in an NRC filing. That same document notes Microsoft’s 20-year power purchase agreement with Constellation, placing the deal squarely within the NRC’s regulatory engagement process. The filing confirms that the federal nuclear regulator is actively tracking the restart effort and the commercial arrangement behind it.
Three Mile Island Unit 1 operated for decades before shutting down, and its potential return represents a significant addition of carbon-free generation capacity. For Microsoft, the deal secures a dedicated source of baseload electricity that can feed data centers without adding carbon emissions or depending on variable renewable output. For Constellation, the agreement provides long-term revenue certainty that helps justify the capital investment required to bring a dormant reactor back online and maintain it over the contract term.
The structure of the deal also signals a shift in how power procurement works for the largest electricity consumers. Rather than buying power on wholesale markets or relying on utility-scale contracts negotiated through regulated processes, Microsoft is contracting directly with a generator for a specific asset over a two-decade horizon. That kind of bilateral arrangement was once unusual outside heavy industry. Its spread into the tech sector reflects how large AI workloads have turned data-center operators into some of the biggest single buyers of electricity in the country.
Grid planning cannot keep up with corporate procurement speed
Regional grid operators and state utility commissions are built to plan on multi-year cycles. They forecast load growth, approve new generation, and schedule transmission upgrades through processes that typically span five to ten years from proposal to energization. AI data-center demand does not follow that timeline. A hyperscaler can commit capital to a new training facility, break ground, and begin drawing power within two to three years, often before the surrounding grid has been upgraded to handle the added load.
The result is a growing tension between corporate procurement speed and regulated planning cycles. When a company like Microsoft signs a 20-year PPA for a specific nuclear plant, it effectively removes that generation capacity from the broader market, tightening supply for other customers in the same region. Grid planners must then account for both the new data-center load and the reduced availability of the generation that was contracted away. If several hyperscalers execute similar deals simultaneously, the cumulative effect on regional power markets could be substantial.
Transmission constraints compound the problem. Even when new generation is available, moving power from a reactor site to a data-center campus requires adequate transmission capacity. Building new high-voltage lines involves years of permitting, environmental review, and right-of-way acquisition. The IEA’s scenario work suggests that without faster grid buildout, the electricity system will struggle to absorb AI-driven demand growth regardless of how much new generation comes online.
Open questions about reactor restarts and new nuclear
The Three Mile Island agreement also raises questions about how far reactor restarts can scale as a strategy for meeting AI demand. Restarting a shuttered unit is not the same as building a new plant from scratch, but it still requires extensive safety reviews, equipment upgrades, and regulatory approvals. Communities around legacy sites may have mixed views about bringing reactors back into service, and regulators must weigh both safety and reliability as they evaluate restart proposals.
There are also policy implications. If more hyperscalers pursue similar contracts, regulators may face pressure to clarify how long-term PPAs interact with wholesale market rules and reliability standards. Questions about who bears the cost of grid upgrades needed to move power from contracted reactors to new data centers are likely to sharpen. In some regions, that could reopen debates about cost allocation between large industrial customers and residential or small-business ratepayers.
For now, Microsoft’s deal with Constellation stands as a template others can study. It illustrates how a single AI-focused company can underwrite the economics of a major nuclear asset while locking in carbon-free power for decades. It also underscores how rapidly AI is reshaping electricity markets: planning processes designed for gradual, predictable load growth are being stress-tested by a wave of concentrated, time-sensitive demand. Whether grid institutions can adapt as quickly as corporate procurement teams remains an open question, and the answer will help determine how sustainable the AI boom proves to be from an energy perspective.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
