In September 2024, Constellation Energy announced it would restart Three Mile Island Unit 1, the undamaged reactor at the Pennsylvania site made infamous by a 1979 partial meltdown, specifically to supply Microsoft with carbon-free electricity for its AI data centers. The 20-year power purchase agreement was the most visible sign of a scramble now reshaping the American energy landscape: the country’s largest technology companies are turning to nuclear power because artificial intelligence is about to devour the grid.
How much power are we talking about? According to the International Energy Agency’s 2025 analysis of AI-driven energy demand, U.S. data centers are on track to consume roughly 240 additional terawatt-hours of electricity per year by 2030, a 130% increase over 2024 levels. To put that in perspective, 240 TWh is more than the entire annual electricity consumption of Poland.
A demand spike unlike anything grid operators have seen
The United States has not experienced electricity demand growth like this in decades. For most of the 2000s and 2010s, total U.S. power consumption was essentially flat, held in check by efficiency gains and the offshoring of heavy industry. AI has broken that pattern.
The Department of Energy, citing estimates from the Electric Power Research Institute, projects that data centers could account for up to 9% of all U.S. electricity generation by 2030, more than double the roughly 4% share they held in 2023. Globally, data centers already consumed about 415 terawatt-hours in 2024, roughly 1.5% of world electricity use, according to the IEA’s executive summary on AI and energy. But the growth is concentrated: the United States and China account for the vast majority of the projected increase.
What makes this surge different from earlier waves of data center construction is the nature of the workload. Training a single frontier AI model can require tens of thousands of GPUs running continuously for weeks or months, consuming as much electricity as a small town. And inference, the process of actually running AI tools for hundreds of millions of users, is scaling even faster. Each new generation of models demands more compute, and so far, efficiency improvements in chip design have not kept pace with the explosion in usage.
Why Big Tech is betting on nuclear
AI data centers need something that solar panels and wind turbines struggle to provide on their own: power that runs 24 hours a day, 365 days a year, without interruption. Nuclear reactors deliver exactly that. They also produce no carbon emissions during operation, which matters to companies that have made net-zero pledges and face growing scrutiny over the climate footprint of their AI products.
The Microsoft-Constellation deal at Three Mile Island is the highest-profile example, but it is far from the only one. Amazon Web Services purchased a nuclear-powered data center campus directly adjacent to Talen Energy’s Susquehanna Steam Electric Station in Pennsylvania, a deal that drew attention from federal regulators over its potential impact on regional grid reliability. Google signed an agreement with Kairos Power in October 2024 to purchase electricity from small modular reactors the startup plans to build by the end of the decade. Oracle has publicly discussed plans to power new data center campuses with small modular reactors as well.
These moves reflect a broader calculation. Data center operators are scouting locations that combine high-capacity transmission lines, favorable permitting, and proximity to firm, low-carbon power. In northern Virginia’s “Data Center Alley,” which hosts the densest concentration of server farms on Earth, available grid capacity is already strained. Similar bottlenecks are emerging in the Dallas-Fort Worth corridor, central Ohio, and parts of the Pacific Northwest. Nuclear plants, many of which sit on large transmission interconnections built decades ago, offer a ready-made solution, at least on paper.
The gap between announcements and megawatts
There is a meaningful difference between signing a power purchase agreement and actually delivering electrons to a server rack. Restarting Three Mile Island Unit 1, for instance, requires relicensing from the Nuclear Regulatory Commission, physical refurbishment of equipment that has sat idle since 2019, and grid interconnection approvals from PJM, the regional transmission organization. Constellation has targeted 2028 for the restart, but that timeline is not guaranteed.
Small modular reactors face even longer timelines. No commercial SMR is operating in the United States as of mid-2026. NuScale Power’s design received NRC certification in January 2023, but the company’s first planned deployment, at the Idaho National Laboratory site, was canceled later that year due to cost overruns. Kairos Power, X-energy, and other developers are progressing through licensing, but first-of-a-kind construction schedules in the nuclear industry have a long history of slipping.
Grid integration poses its own challenges. Even when a reactor is running, connecting its output to a specific cluster of data centers may require transmission upgrades that take years to permit and build. The Federal Energy Regulatory Commission and regional grid operators are still working through how to handle “co-location” arrangements, where a data center sits behind the meter at a power plant, drawing electricity before it reaches the public grid. Critics, including some utility regulators, argue that these deals could raise costs for other ratepayers or reduce grid reliability.
What efficiency gains can and cannot do
Not everyone believes the demand projections will hold. Advocates of more efficient AI hardware point to advances in chip architecture, liquid cooling systems, and model optimization that could significantly reduce energy consumption per computation. Nvidia’s newest GPU generations, for example, deliver substantially more AI performance per watt than their predecessors. If those gains continue, the demand curve could bend downward.
But history offers a cautionary pattern. In computing, efficiency improvements have consistently been overwhelmed by rising usage, a dynamic sometimes called Jevons’ paradox. Cheaper, faster AI compute tends to unlock new applications and broader adoption, which drives total energy consumption higher even as each individual task becomes more efficient. The IEA’s Base Case projection already accounts for expected efficiency improvements and still arrives at a 130% increase.
The DOE has acknowledged this tension. Its published guidance on clean energy resources for data center demand emphasizes that meeting the projected load growth with low-carbon sources will require action on multiple fronts: new generation capacity, grid modernization, and demand-side efficiency measures. Nuclear is one piece of that puzzle, not the whole answer.
What this means for electricity bills and the climate
For ordinary Americans, the most immediate question is whether AI’s appetite for power will drive up electricity prices. The answer depends heavily on geography. In regions where data centers are clustering fastest, utilities are already filing for rate increases tied to infrastructure upgrades needed to serve new load. Dominion Energy, which serves northern Virginia, has proposed billions of dollars in transmission spending that would ultimately be recovered from all ratepayers, not just the tech companies requesting the capacity.
The climate implications cut both ways. If tech companies follow through on nuclear partnerships and new renewable procurement, the additional load could be met largely with clean energy, limiting the carbon impact. But if demand outpaces clean supply, utilities may keep aging natural gas plants running longer or build new ones to fill the gap. The IEA has flagged this risk explicitly: without deliberate policy and investment, AI-driven electricity growth could slow or reverse progress on grid decarbonization.
Congress has shown bipartisan interest in smoothing the path for nuclear. The ADVANCE Act, signed into law in July 2024, streamlined parts of the NRC’s licensing process and authorized incentives for advanced reactor deployment. Whether those policy changes translate into reactors built on timelines that matter for the 2030 demand surge remains an open question.
Where the story stands in mid-2026
As of June 2026, the core facts are clear: AI is driving the fastest growth in U.S. electricity demand in a generation, and the country’s biggest technology companies have placed significant bets on nuclear power to meet it. The IEA’s projection of a 130% increase in data center electricity use by 2030 is grounded in transparent methodology and corroborated by the DOE’s own estimates. Multiple corporate-nuclear partnerships have been publicly announced, with real money committed.
What remains unresolved is execution. Restarting shuttered reactors, building new ones, upgrading transmission networks, and navigating regulatory and community opposition are all tasks measured in years, not quarters. The gap between the pace of AI deployment and the pace of energy infrastructure construction is the central tension in this story, and it is not close to being resolved. The next 18 months will reveal whether the nuclear bets Big Tech has placed can deliver power fast enough to keep the AI boom from running headlong into the limits of the American grid.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
