Tech Giants Turn to Nuclear Revival to Feed AI’s Insatiable Power Hunger

Electricity demand from data centers is exploding. The International Energy Agency projects global consumption will roughly double from 485 terawatt-hours in 2025 to 950 TWh in 2030. AI-focused facilities drive much of that surge, with their usage tripling over the same period. And the numbers keep climbing. By some estimates, data centers could exceed 1,000 TWh by 2026 alone.

That’s more electricity than Japan consumes in a year. Short sentences hit hard here. The scale surprises even seasoned energy analysts.

Tech companies see the writing on the wall. Microsoft, Google and Amazon have struck deals to restart shuttered nuclear plants, back small modular reactor startups and secure long-term supplies of carbon-free baseload power. They aren’t waiting for the grid to catch up. They are rewriting the rules.

The International Energy Agency spelled it out in its latest assessment. Electricity demand from data centers grew 17 percent in 2025. AI-specific loads jumped 50 percent. Model providers reported triple the active users and five times the revenue. Growth shows no sign of slowing.

But bottlenecks exist. Supply chain constraints, regulatory hurdles and construction timelines limit how fast new capacity comes online. Natural gas still fills much of the gap today. Renewables add volume yet fall short on the constant, high-density power AI training and inference require. Nuclear offers a different path. It runs 24 hours a day. Capacity factors exceed 90 percent. Emissions stay near zero.

Microsoft moved first and fastest. The company signed a 20-year agreement with Constellation Energy to restart the Three Mile Island reactor in Pennsylvania. That deal, valued around $16 billion according to reports, targets commercial operation by 2028. It marks the first time a technology firm has financed the restart of a nuclear facility specifically for data center needs. Constellation received a $1 billion Department of Energy loan to speed the process.

Amazon took a direct approach. It bought a data center campus adjacent to the Susquehanna nuclear plant from Talen Energy for $650 million. Then it expanded the relationship with a 1.9-gigawatt power purchase agreement running through 2042. The company also poured more than $500 million into various nuclear projects across Virginia and Washington state. These moves let Amazon tap power behind the meter. They sidestep some transmission congestion that slows other developments.

Google chose the future. It agreed to buy power from up to seven small modular reactors developed by Kairos Power. The first units could come online by 2030, eventually delivering 500 megawatts. Google also signed deals with the Tennessee Valley Authority to support reactors in Tennessee and Alabama. The technology remains unproven at commercial scale. Yet the contracts signal confidence that SMRs can deliver cheaper, faster builds than traditional large reactors.

Meta joined the fray. The social media giant issued requests for proposals seeking 1 to 4 gigawatts of new nuclear capacity. Other hyperscalers followed similar paths. In total, big tech has contracted for more than 10 gigawatts of potential new nuclear supply in the United States over the past year. Goldman Sachs Research estimates 85 to 90 gigawatts of additional nuclear capacity would be needed globally to cover all data center demand growth by 2030. Less than 10 percent of that will likely materialize in time.

So utilities scramble. Georgia Power revised its demand forecasts upward by a factor of 17 in some cases, largely due to AI-driven projects. Data centers are projected to claim 11.6 gigawatts of new North American power this year. That represents one-third of all new generation added across the continent. The numbers don’t lie. Demand outruns supply.

France offers a glimpse of what targeted nuclear policy can achieve. Data center operator Data4 signed a 12-year contract with state-owned utility EDF for 40 megawatts of nuclear power priced at production cost rather than volatile wholesale rates. Nuclear supplies about 70 percent of France’s electricity. The country markets that reliability as an advantage in the AI race.

In the United States the picture mixes promise with delay. The Palisades plant in Michigan received approval to restart. NextEra Energy explores bringing back the Duane Arnold facility in Iowa. Yet new large reactor construction still takes a decade or more. Small modular reactors promise shorter timelines and factory-built modules. Hyperscalers bet heavily on that promise.

The Brookings Institution examined the mix in April 2026. Natural gas provided 40 percent of data center electricity in 2024. Renewables followed at 24 percent, nuclear at 20 percent and coal at 15 percent. Projections show natural gas retaining the largest share through 2030. Nuclear gains ground after that, especially once SMRs enter service. Renewables meet nearly half the added demand in the next five years. Fossil fuels still cover much of the incremental load because of existing plant economics.

Critics point to safety concerns and waste issues. Supporters highlight decades of reliable operation and the absence of viable alternatives for always-on power. Goldman Sachs analysts argue nuclear stands as the preferred baseload choice. Construction difficulties push companies toward natural gas and renewables for the short term. The tension plays out in boardrooms and regulatory hearings alike.

Recent developments add momentum. In early June 2025 Amazon expanded its Talen Energy agreement. Google deepened ties with Kairos. Meta’s RFP process continues. These aren’t isolated bets. They form a pattern. Tech executives concluded that waiting for grid upgrades or hoping renewables scale fast enough carries too much risk. Direct control over power generation becomes a competitive necessity.

The Goldman Sachs report from January 2025 captured the stakes. Efficiency gains in data center hardware have begun to slow. Power density in GPU racks keeps rising. A single next-generation AI supercluster can draw over a gigawatt. That’s the output of a full-scale nuclear reactor. Or enough electricity for hundreds of thousands of homes.

Utility Dive explored the grid implications in March 2026. Renewables and batteries cannot alone satisfy the 24/7 requirements of hyperscale AI workloads. Nuclear scales with the need. It avoids the intermittency problem. Communities dependent on variable sources risk higher costs and reliability shortfalls.

Yet challenges remain. Regulatory approval for restarts and new builds still takes years. Supply chains for specialized components stay tight. Public perception of nuclear power varies by region. Tech companies spend heavily on community outreach and safety demonstrations to ease those worries.

IEA analysts see nuclear playing a larger role after 2030. Small modular reactors should begin contributing then. Hyperscalers rank among the biggest corporate backers of SMR development. Their commitments help de-risk the technology for other customers. By 2035 the combination of added nuclear output and continued renewable growth could finally reduce coal use for data center operations.

Pew Research Center put U.S. figures in perspective last October. American data centers consumed 183 TWh in 2024. That equals more than 4 percent of national electricity use. Equivalent to the entire power demand of Pakistan. Projections show that load more than doubling to 426 TWh by 2030. The growth trajectory looks set.

Executives speak carefully in public. They emphasize carbon-free goals alongside reliability. Few admit outright that intermittent renewables alone won’t cut it for AI at scale. The deals tell the real story. Nuclear is back on the table because the math demands it.

Construction timelines still loom large. Even optimistic SMR forecasts put meaningful capacity online only in the early 2030s. In the meantime natural gas plants fill the breach. Some utilities extend the life of coal facilities. The interim solution carries emissions costs that clash with corporate net-zero pledges. The contradiction creates awkward conversations with investors and regulators.

But the direction feels clear. AI won’t pause for the energy industry to sort itself out. Model sizes grow. Inference demand multiplies. Training runs stretch longer. Each advance consumes more power than the last. Companies that secure stable supplies first gain an edge. Those that don’t risk curtailments, higher prices or outright inability to expand.

The revival of nuclear for data centers marks a quiet but profound shift. Decades after Three Mile Island and Chernobyl slowed the industry, commercial demand from Silicon Valley breathes new life into it. Old plants get new leases. Startups attract billion-dollar commitments. Policy makers in Washington and state capitals reconsider long-stalled projects.

Whether small modular reactors deliver on their cost and schedule promises will decide much of the next decade. Early contracts from Google, Amazon and others provide crucial validation. Success could open the door for widespread adoption beyond tech. Failure might push the sector back toward gas turbines and hope for battery breakthroughs.

For now the bets are placed. Microsoft restarts a plant synonymous with nuclear anxiety. Amazon co-locates data halls next to fission reactors. Google orders fleets of miniature reactors not yet built. The power hunger of artificial intelligence has forced a reckoning. And the answer, at least for these companies, comes with a nuclear signature.