Power shortages now threaten to slow artificial intelligence. Hyperscalers have responded with a bold move. They are turning to nuclear energy.
Microsoft struck a deal to restart the Three Mile Island reactor. Google signed on for a fleet of small modular reactors from Kairos Power. Amazon poured money into both existing plants and next-generation designs. These agreements mark more than opportunistic energy buys. They signal a fundamental shift in how the biggest technology companies secure electricity for the coming decade of AI growth.
The numbers tell a stark story. U.S. data center power demand stood at 31 gigawatts in 2025. It climbs to 41 gigawatts this year. By 2027 that figure doubles to 66 gigawatts, according to Goldman Sachs research. Global data center electricity consumption could exceed 1,000 terawatt-hours by 2026. That rivals the output of all nuclear plants worldwide just a few years ago, the International Atomic Energy Agency reports.
But renewables alone won’t cut it. Solar and wind suffer from intermittency. They require massive battery storage to deliver the always-on power AI training demands. Natural gas fills gaps today yet clashes with corporate net-zero pledges. Nuclear offers firm, carbon-free baseload. No wonder the tech sector has committed tens of billions to revive and expand it.
Take Microsoft’s arrangement with Constellation Energy. The 20-year power purchase agreement covers the full 835-megawatt output of the restarted Three Mile Island Unit 1, now called the Crane Clean Energy Center. Constellation plans to invest $1.6 billion. The reactor could come online in 2027, a year earlier than first expected after the regional grid operator accelerated interconnection, Reuters detailed last year. The plant sits in Pennsylvania. Its history weighs heavy. It suffered the worst nuclear accident in U.S. history in 1979. Yet the undamaged Unit 1 closed in 2019 for economic reasons amid cheap natural gas.
Now economics point the other way. AI creates voracious demand. A single frontier model training run may need five gigawatts by 2027, Anthropic has estimated. Microsoft alone projects the U.S. AI sector will require 50 gigawatts of new capacity by 2028. The Three Mile Island restart gives the company dedicated clean power for its data centers. It also helps keep the lights on for surrounding communities.
Google took a different path. In late 2024 it inked the first corporate deal for multiple small modular reactors. The agreement with Kairos Power targets 500 megawatts by 2035, with the initial unit online by 2030. These reactors use fluoride salt-cooled technology. They promise factory-built modules, faster construction and enhanced safety features. The plants will locate in grids serving Google data centers to support the company’s 24/7 carbon-free energy goal. “This deal will enable up to 500 MW of new 24/7 carbon-free power to U.S. electricity grids,” Google stated in its official announcement.
Amazon has pursued both strategies at once. It agreed to buy 1.9 gigawatts from Talen Energy’s Susquehanna nuclear plant through 2042. The company also acquired a data center campus next to the facility for $650 million, creating a behind-the-meter connection that bypasses congested transmission lines. Beyond that, Amazon led a $500 million financing for X-energy, developer of another SMR design, and partnered with Energy Northwest on advanced reactor projects. These moves position Amazon to power AI infrastructure directly from nuclear sources while hedging against grid delays.
Meta joined the fray too. It issued a request for proposals seeking one to four gigawatts of new nuclear generation. Across these four companies, contracts now cover more than 10 gigawatts of potential new or restarted nuclear capacity in the United States. The collective investment exceeds $50 billion when counting direct financing, acquisitions and long-term power deals, recent analyses show.
Yet challenges abound. Construction timelines stretch long. Even restarted plants face regulatory hurdles, supply chain bottlenecks for specialized parts and public skepticism rooted in past accidents. Small modular reactors remain largely unproven at commercial scale. The first U.S. deployments won’t arrive before the end of the decade. In the meantime data center operators scramble for any available power.
The International Energy Agency sees data center demand more than doubling globally to 945 terawatt-hours by 2030. In the U.S. alone it could reach between 325 and 580 terawatt-hours by 2028, up from 176 terawatt-hours in 2023, according to a Lawrence Berkeley National Laboratory study cited in Brookings Institution analysis. That range equals 6.7% to 12% of total U.S. electricity. Uncertainties run high. Faster AI adoption could push numbers toward the upper end. Efficiency gains in chips and algorithms might temper growth.
Energy experts point to a mix of solutions. Renewables will supply roughly half the incremental demand through 2035, the IEA projects in its executive summary on energy and AI. Natural gas handles much of the rest in the near term. Nuclear contributes significantly but deploys slowly. The tech companies’ deals aim to pull nuclear forward. They provide revenue certainty that encourages developers to move faster and investors to fund projects.
Recent developments underscore the momentum. NextEra Energy took full ownership of an Iowa nuclear plant slated for restart to serve Google data centers, traders noted on X this month. Discussions continue around fusion startups like Helion Energy, with whom Microsoft signed a power purchase agreement. These represent longer shots. But they illustrate the breadth of the search for high-density, clean power.
Financial markets have taken notice. Constellation Energy shares surged after the Microsoft deal. Nuclear equipment suppliers and uranium producers gained attention. Real estate around restarted plants sees rising interest for data center co-location, logistics and housing. The energy bottleneck has become a central constraint on AI expansion. Solving it could determine which companies lead the next wave of innovation.
Still, not everyone cheers the nuclear turn. Some environmental groups worry about radioactive waste and accident risks, however low the probability. Others question whether utilities can integrate these large, inflexible power sources without disrupting grids built for variable renewables. Regulators must balance speed with safety. Congress has offered loan guarantees and streamlined permitting. The Department of Energy backed the Three Mile Island restart with a $1 billion loan.
Big tech now acts as an unconventional force in energy policy. Its purchasing power and tolerance for long-term contracts give it leverage traditional utilities lack. Amazon, Google, Microsoft and Meta don’t just consume power. They shape its future supply. Their choices influence everything from uranium mining to reactor design to grid planning.
The bet carries risks. Delays could force reliance on dirtier sources or curtail AI ambitions. Costs may exceed projections. Yet the alternative, stalled data center construction, looks worse for companies racing to deploy ever-larger models. Demand shows no sign of easing. Training runs grow more compute-intensive. Inference scales with user adoption. Every breakthrough seems to multiply electricity needs.
So the nuclear deals keep coming. They reflect a pragmatic recognition. AI’s potential demands energy on a scale few anticipated. Intermittent renewables and flexible gas cannot shoulder the full load alone. Firm, low-carbon power must form the backbone. Nuclear, long sidelined, finds itself back in favor. Not because the technology changed overnight. But because the appetite for computation did.
Whether these projects deliver on time remains the open question. Progress at Three Mile Island will offer the first major test. If the 2027 target holds, it could unlock similar restarts elsewhere. Successful SMR deployments from Kairos or X-energy would open the door to factory-style production and wider adoption. The hyperscalers have placed their chips. The industry now watches to see if the reactors light up in time to keep the AI boom alive.
WebProNews is an iEntry Publication