The AI Boom Is Bringing Nuclear Power Back — And Investors Are Betting Billions on It
AI Is Running Out of Power — So Big Tech Is Bringing Nuclear Energy Back
Big Tech Turns to Nuclear
Artificial intelligence has created an unexpected winner far beyond semiconductors and software: nuclear power. As technology companies race to build enormous data centres capable of training and operating increasingly powerful AI systems, electricity is becoming one of the industry's most important constraints — and some of the world's largest companies are turning towards reactors that only a few years ago appeared to belong to an ageing industry.
Microsoft is supporting the restart of a retired American nuclear reactor. Google has agreed to buy electricity from a planned fleet of advanced reactors. Amazon is backing projects that could eventually exceed five gigawatts, while Meta says agreements involving existing and future reactors could support as much as 6.6GW of nuclear capacity by 2035. At the same time, hundreds of millions of dollars are flowing into advanced-reactor developers and enormous energy transactions are being justified partly by the electricity demands of the emerging AI economy.
AI Has Turned Electricity Into a Strategic Resource
The first phase of the generative-AI boom was dominated by chips. Nvidia GPUs became extraordinarily valuable because companies desperate to train larger models needed enormous quantities of specialised computing hardware, but the physical bottleneck is now expanding beyond processors.
Every accelerator installed in a data centre needs electricity. It also needs cooling, networking equipment, substations, transmission infrastructure and a power system capable of operating around the clock. The International Energy Agency expects global data-centre electricity consumption to more than double to around 945 terawatt-hours by 2030, slightly more than Japan consumes today. AI is expected to be the largest driver of that increase.
The situation is especially significant in the United States. The IEA projects data centres will account for almost half of the growth in American electricity demand through 2030, creating an unusual collision between an industry capable of deploying billions of dollars of computing equipment quickly and an electricity system where new generation and transmission can take years to approve and construct.
That is why power and cooling are becoming decisive constraints on AI growth. The battle to build the best artificial intelligence increasingly depends on something much older and less glamorous: securing enough electricity to keep the machines running.
Microsoft Is Bringing a Retired Nuclear Reactor Back
Few deals illustrate the reversal in nuclear power's fortunes better than Microsoft's agreement with Constellation Energy.
Three Mile Island Unit 1 in Pennsylvania stopped operating in 2019 because of economics rather than a technical failure. It is separate from Unit 2, which was permanently closed after the famous 1979 accident. Under a 20-year power purchase agreement with Microsoft, Constellation is now working to restart Unit 1 under a new name: the Crane Clean Energy Center.
The plant is expected to return approximately 835MW of generation to the grid. Constellation originally estimated that roughly $1.6 billion of capital expenditure would be required for the restart, while the US Department of Energy subsequently approved a loan guarantee of up to $1 billion to support the project. Regulatory approvals, including nuclear safety and environmental reviews, are still required before electricity can begin flowing.
Microsoft's involvement changes the economics. Instead of reopening a reactor and hoping wholesale electricity prices remain attractive for decades, Constellation has a major technology company prepared to underpin the project through a long-duration agreement.
That model could become enormously important.
Google Wants an Entire Fleet of Advanced Reactors
Google has taken a different route. Rather than reviving an existing conventional reactor, it signed an agreement with Kairos Power to support multiple advanced nuclear reactors.
The initial plan aims for the first reactor to come online around 2030, followed by additional deployments through 2035. Google described the arrangement when announced in 2024 as the world's first corporate agreement to purchase electricity from multiple small modular reactors.
The concept has since moved beyond a corporate announcement. In 2026, Kairos said construction had begun on its Hermes 2 demonstration plant in Oak Ridge, Tennessee. The project is intended to supply up to 50MW to the Tennessee Valley Authority grid and represents the first delivery under the company's broader development agreement with Google.
Small modular reactors are attractive partly because their developers hope standardised components and repeatable designs can overcome one of conventional nuclear power's greatest weaknesses: enormous bespoke construction projects that can become expensive and slow.
Whether they actually achieve that at commercial scale remains one of the biggest unanswered questions in the nuclear revival.
Amazon and Meta Are Going Even Bigger
Amazon has also moved directly into advanced nuclear.
X-energy says its partnership with Amazon includes the option to deploy more than 5GW of new nuclear capacity by 2039, designed partly to support AWS data-centre demand and wider grid reliability. The company is developing its Xe-100 reactor, with individual units designed for roughly 80MW and multi-reactor plants capable of reaching hundreds of megawatts.
Capital is following the opportunity. X-energy closed an upsized $700 million financing round as it worked to advance its reactor technology, while its commercial pipeline has continued expanding. By 2026 the company said it was working across more than 11GW of potential nuclear capacity involving projects and partnerships in the United States and Britain.
Meta has pushed the scale further. In January 2026 it announced agreements involving Vistra, TerraPower and Oklo, building on an earlier agreement with Constellation. Meta said the combined projects could support up to 6.6GW of new and existing nuclear power by 2035.
Part of the strategy involves preserving existing nuclear generation. Another part involves supporting advanced reactor developers whose technologies have yet to prove themselves through widespread commercial operation.
That distinction matters. Big Tech is simultaneously extending the life of nuclear technology that already works and placing much riskier bets on the reactors it hopes will work next.
Why Nuclear Suddenly Makes Sense for Big Tech
Nuclear energy has several characteristics that match the unusual requirements of AI infrastructure.
Data centres operate continuously. A company spending tens of billions of dollars on GPUs does not want those machines waiting for favourable weather conditions before they can operate, and although renewables can provide enormous amounts of low-cost electricity, matching supply continuously requires grids, storage and complementary generation.
Nuclear plants can produce very large quantities of electricity with extremely low operational carbon emissions while running at high capacity factors. For technology companies simultaneously promising enormous AI expansion and ambitious emissions reductions, that combination is increasingly valuable.
It does not mean nuclear will power the entire AI revolution. The IEA expects renewables to supply roughly half of the increase in global data-centre electricity demand through 2035, with natural gas also playing a substantial role. Nuclear is expected to contribute additional generation alongside them rather than replace them.
The change is therefore not that Silicon Valley has discovered a single perfect energy source. It is that electricity demand has become so large that the companies building AI infrastructure increasingly need almost everything that can be deployed at scale.
The Nuclear Revival Has Become National Strategy
The shift is also moving beyond private companies.
In May 2025, President Donald Trump signed executive orders aimed at accelerating advanced nuclear development. One specifically linked reliable, high-density power with AI infrastructure and directed the Department of Energy to facilitate privately financed advanced reactors at federal sites, with an ambition for an initial reactor to operate within 30 months of the order.
That places nuclear power inside a much larger competition over artificial intelligence, industrial capacity and national security.
If advanced AI eventually becomes critical to military planning, scientific research, cybersecurity, manufacturing and economic productivity, then the electricity infrastructure supporting AI becomes strategically important as well. The distinction between technology policy and energy policy starts to disappear.
The AI power race is therefore becoming an industrial race, with countries increasingly forced to ask not simply who possesses the best AI models, but who can build enough physical infrastructure to operate them at enormous scale.
Investors Are Betting on More Than Reactors
The investment opportunity extends well beyond companies actually designing nuclear plants.
A sustained nuclear expansion requires uranium mining, enrichment, specialised fuel, engineering companies, reactor components, construction capacity, turbines, cooling equipment, electrical infrastructure and skilled workers. Existing nuclear operators may also become more valuable if long-duration contracts with technology companies allow old plants to remain open or increase output.
Constellation has become one of the clearest corporate examples of that shift. Its acquisition of Calpine, completed in January 2026 after a deal originally announced with a roughly $16.4 billion equity purchase price plus assumed debt, created a huge electricity producer combining nuclear generation with natural gas, geothermal and other assets. Constellation has explicitly framed the enlarged company around growing electricity demand from data centres and the wider data economy.
That is significant because it reveals what investors are really buying.
The nuclear thesis is increasingly part of an electricity-scarcity thesis.
If AI infrastructure expands at the extraordinary scale currently contemplated, companies controlling reliable generation, grid access and developable energy projects could become almost as strategically important to the AI economy as some of the companies producing the computing hardware itself.
The Problem: Nuclear Cannot Be Built at AI Speed
There is still a huge gap between announcing nuclear projects and generating electricity.
Large conventional reactors have historically suffered from long construction schedules and cost overruns in several Western countries. New advanced designs promise improvements, but many have not yet demonstrated that they can be manufactured repeatedly, licensed quickly and built economically at commercial scale.
Even the IEA's base-case outlook does not expect the first small modular reactors serving this broader electricity-demand wave to arrive until around 2030.
Fuel supply is another constraint. Several advanced reactor concepts require high-assay low-enriched uranium, or HALEU, which has a much more limited Western supply chain than conventional reactor fuel. Regulatory processes, manufacturing capability and nuclear-skilled labour also cannot be expanded overnight.
This creates a strange mismatch.
AI investment can move extraordinarily fast. A technology company can order tens of thousands of accelerators and commit billions of dollars to a data-centre campus within a relatively short period.
A nuclear industry cannot simply multiply itself on the same timetable.
The Biggest Opportunity May Be Keeping Existing Reactors Alive
That makes existing nuclear plants particularly valuable.
Restarting or extending a reactor that has already been built can potentially add dependable electricity far sooner than constructing an entirely new fleet. Microsoft's agreement with Constellation and Meta's support for existing Vistra reactors demonstrate why the first phase of the nuclear comeback may depend as much on preserving and upgrading America's current fleet as on futuristic SMRs.
Constellation has also secured extended operating licences for its Clinton and Dresden nuclear stations and said it is investing more than $370 million in relicensing and upgrades. Clinton can now operate through 2047, while Dresden's units have licences extending into 2049 and 2051.
When electricity demand was stagnant, ageing reactors could struggle economically against cheaper alternatives.
When the customer is an AI company desperate to secure dependable electricity for decades, the same reactor can suddenly become a strategic asset.
AI Could Change the Economics of Nuclear Power
This may ultimately be the most important part of the story.
For decades, one of nuclear power's fundamental problems in Western electricity markets was financial. Reactors required enormous upfront investment, lengthy development and confidence that electricity revenues decades into the future would justify the risk.
Hyperscalers change that equation.
Microsoft, Google, Amazon and Meta are among the largest and most creditworthy companies on Earth. They are also facing an electricity requirement that could persist for decades if AI computing continues expanding.
Long-term agreements can therefore transfer some of the demand risk away from the nuclear developer. A reactor no longer has to be financed purely on an uncertain prediction of future wholesale electricity prices if a giant technology company is willing to contract for its output.
AI is effectively creating a new class of industrial electricity customer with enormous balance sheets, long planning horizons and an extraordinary appetite for dependable power.
That is almost exactly the type of customer nuclear energy has been waiting for.
The Nuclear Boom Is Real — But the Winners Are Not Yet Clear
The revival should not be confused with certainty.
Some advanced-reactor companies will almost certainly fail. Construction costs could prove higher than promised. Regulatory deadlines may slip. Fuel supplies could remain constrained. AI hardware could become dramatically more energy-efficient, while renewables, storage, geothermal power or natural gas could capture more of the demand than nuclear advocates expect.
There is also a broader question hanging over the entire investment boom: whether future AI revenues will become large enough to justify the staggering quantities of infrastructure currently being planned.
But something fundamental has already changed.
A reactor that once looked like an expensive relic can now look like an AI infrastructure asset. Technology companies that previously spent their capital on software, chips and data centres are increasingly helping determine which power stations remain open and which next-generation reactors get financed.
The physical limits of AI are becoming impossible to ignore. The industry may eventually discover that the most valuable resource in the artificial-intelligence race is not another algorithm or even another GPU.
It may simply be electricity — enormous quantities of dependable electricity, available every hour for decades.
If that proves true, AI will have done something few expected when ChatGPT triggered the modern technology boom: it will have helped bring the nuclear age back.
