Nuclear Power For AI Data Centers, Microsoft Three Mile Island Deal

Last updated: May 2026 · 15 min read · AI & Energy

The AI Energy Crisis Is Here: Why Big Tech Is Quietly Buying Up America's Nuclear Power in 2026

In the span of eighteen months, the four richest companies on Earth have signed nuclear power agreements worth tens of billions of dollars. They are not doing it for the climate. They are doing it because, without it, ChatGPT goes dark, Gemini stops answering, and the entire generative AI economy hits a literal wall — the wall of the US electrical grid.

If you've been watching your monthly power bill creep up and wondering whether something bigger is going on, you're not imagining it. Something bigger is going on. The same forces driving the AI revolution — the GPUs in Nvidia's most expensive racks, the warehouse-sized data centers spreading across Virginia, Texas, and Ohio — are now driving the largest sustained surge in American electricity demand since the post-war industrial boom.

And the people running these AI companies have done the math. They've looked at the projections. And they've decided, almost in unison, that the only way out is to do something nobody seriously considered five years ago: restart shuttered nuclear plants, build new small modular reactors (SMRs), and lock in 20-year deals for atomic energy. This is the story of the AI energy crisis of 2026 — what's actually happening, why it matters for your wallet, and where the smart money is moving next.

Once seen as relics of the 20th century, nuclear cooling towers are now the most coveted real estate in American tech.

The Numbers Nobody Wants You to See

Let's start with the part of this story you probably haven't heard, because it doesn't fit in a 60-second clip. According to the International Energy Agency's Energy and AI report, global data centers consumed roughly 415 terawatt-hours of electricity in 2024 — about 1.5% of total global electricity. That sounds small until you realize where the curve is heading.

The IEA now projects data center electricity demand will more than double by 2030, reaching roughly 945 TWh worldwide. In the United States — which already accounts for the largest share of that consumption — data center demand is projected to hit somewhere between 300 and 400 TWh per year by 2030. For context, that's roughly equivalent to the entire annual electricity consumption of a country like France.

What's pushing it? Not your Netflix stream. Not your iPhone backup. It's AI. Specifically, large language model training and inference. The Electric Power Research Institute (EPRI) estimated that in 2024, AI workloads already accounted for 10–20% of total data center energy. By the back half of this decade, that share is expected to dominate. A single ChatGPT-style query uses roughly 10 times the electricity of a Google search. Multiply that by billions of queries per day, and you start to see the problem.

The Belfer Center at Harvard put it bluntly in their 2025 analysis: from 2018 to 2023, US data center energy use rose from 76 TWh to a much larger figure, and the next five years will see an even sharper hockey stick. Utilities in Virginia — home to Data Center Alley — are warning that their grids cannot physically deliver the load that hyperscalers are requesting. In Georgia, Texas, and Arizona, queue times for new grid interconnections have stretched from 18 months to nearly five years.

This is what an energy crisis looks like in the modern era. It is not gas lines. It is delayed AI rollouts, scrapped data center groundbreakings, and the quiet, terrified realization inside the C-suite of every major tech company that the bottleneck on the next decade of growth is not chips — it's electrons.

The Microsoft–Three Mile Island Deal That Started It All

In September 2024, Microsoft and Constellation Energy announced something that, on its face, sounded almost like a joke: they were going to restart the Three Mile Island Unit 1 reactor in Pennsylvania — the same site where, in 1979, America's worst-ever commercial nuclear accident happened on the neighboring Unit 2.

The deal is a 20-year power purchase agreement, the largest in Constellation's history. Microsoft will buy 100% of the output. Constellation will invest about $1.6 billion to bring Unit 1 — which was shut down in 2019 for economic reasons — back online. The site is being rebranded the Crane Clean Energy Center. Initially targeted for 2028, the restart is now reportedly being pulled forward to 2027, one year ahead of plan.

It was the shot that started the race. Within twelve months, every major hyperscaler had moved.

The bottleneck is no longer chips. It's electrons.

Meta's 6.6 Gigawatt Bet on Vistra, TerraPower, and Oklo

In January 2026, Meta announced what may turn out to be the boldest nuclear strategy of any tech company. The company signed three separate landmark agreements — with Vistra, Bill Gates–backed TerraPower, and Sam Altman–backed Oklo — that together unlock up to 6.6 gigawatts of nuclear capacity by 2035.

The Vistra agreement is a 2.2 GW power purchase deal supporting Meta's growing AI super-cluster footprint. The TerraPower agreement supports the company's pioneering Natrium reactor — a sodium-cooled design that pairs nuclear with a molten salt energy storage system. And the Oklo agreement is, in some ways, the most strategically interesting: a 1.2 GW partnership in Southern Ohio for Oklo's compact "Aurora" microreactor design, with Meta providing prepayment capital to fund early-stage fuel procurement and site work.

What makes Meta's strategy notable is how vertically committed it is. By prepaying for fuel and site development, Meta isn't just buying electrons — it's underwriting an entire next-generation reactor industry. If Oklo's design works at commercial scale, Meta will have a power supply that competitors literally cannot access at any price.

Amazon's SMR Triple-Play

Amazon, never one to be outmaneuvered on infrastructure, has gone in a different direction: Small Modular Reactors, or SMRs. These are factory-built nuclear units, typically rated between 50 and 300 megawatts, that can be deployed in modular fashion rather than as a single mega-plant.

In late 2024 and through 2025, Amazon Web Services signed three major SMR agreements:

1. A partnership with Dominion Energy to explore building SMRs near Dominion's existing North Anna nuclear station in Virginia — the heart of "Data Center Alley."
2. An equity investment in X-energy, a developer of advanced high-temperature gas reactors, with associated deployment agreements in Washington State.
3. A separate partnership with Energy Northwest to develop SMRs in the Pacific Northwest, supporting AWS's growing regional footprint.

AWS CEO Matt Garman has been explicit about why: SMRs offer the kind of predictable, baseload, carbon-free power that data centers need, without the multi-decade build times of traditional gigawatt-scale plants. If they work as advertised — and that "if" is doing some heavy lifting — they could be operational by the early 2030s.

Google's Kairos Power Partnership

Google's first major nuclear partnership, announced in October 2024, is with Kairos Power, an Alameda, California-based SMR developer. The deal supports the deployment of multiple small modular reactors with combined capacity of roughly 500 megawatts. The first units are targeted for the early 2030s, with full deployment by 2035.

The Kairos design is a fluoride salt–cooled, high-temperature reactor — a fundamentally different concept from the light-water reactors that have dominated commercial nuclear since the 1960s. If it succeeds, Google will have helped birth what could become a dominant new reactor architecture.

According to a Deloitte analysis, nuclear energy could meet up to 10 percent of total data center electricity demand by 2035. That's not a footnote. That's the entire annual electricity output of several Western European countries, dedicated essentially to AI.

Why Nuclear and Not Solar or Wind?

This is the question every reader sends me. If the climate is on fire, why aren't these companies doubling down on solar, wind, and battery storage? The honest answer involves three uncomfortable truths.

First: Data centers want 24/7/365 firm power. AI training jobs cannot pause when the sun sets or the wind dies. A modern GPU cluster running a frontier model can have a power profile that's flat to the milliwatt for weeks at a time. Solar and wind, even with battery storage, struggle to deliver that profile economically at the gigawatt scale.

Second: Land and grid constraints. A 1-gigawatt nuclear plant occupies roughly 1 to 2 square miles. The equivalent solar farm — accounting for capacity factor — would require something closer to 50 to 75 square miles, plus enormous battery installations. Finding that land near major fiber routes and existing transmission infrastructure has become essentially impossible.

Third: The "Power Purchase Agreement" math actually works. Restarting a paid-for plant (like Three Mile Island) delivers electricity at a marginal cost that, while not cheap, is at least predictable across 20 years. New solar plus storage, by contrast, carries grid integration risks, transmission build-out costs, and capacity-factor uncertainty that hyperscalers find harder to underwrite over multi-decade contracts.

None of this means nuclear is the "right" answer for everyone, or that it's risk-free. It just means that — for the unique workload of giant AI data centers — the spreadsheet keeps pointing back to atoms.

A modern AI training cluster can demand electricity equivalent to a small city, around the clock.

What This Means for Your Electric Bill

Here's where the story gets uncomfortable. Power purchase agreements between hyperscalers and nuclear operators are typically structured so that the hyperscaler buys most or all of the output of a specific plant. In Microsoft's case at Three Mile Island, the company is buying 100% of Unit 1's output. None of that power goes to homes or businesses in Pennsylvania.

That doesn't directly raise residential bills, but indirectly, it absolutely can. When a region's largest dedicated baseload generator is committed to a single corporate customer, other utilities have to source replacement power from elsewhere — often at higher cost. Meanwhile, transmission upgrades needed to serve massive new data center loads are frequently socialized across all ratepayers, even though only one customer is creating the demand.

Multiple analyses in 2025 and early 2026 have flagged exactly this concern. In Virginia, regulators have begun examining whether Dominion's residential customers are effectively subsidizing the grid build-out needed to serve hyperscaler data centers. In Ohio, similar questions are being asked. The political conversation in 2026 around AI is, increasingly, also a conversation about who pays for the power AI consumes.

The Investment Angle: Stocks, ETFs, and Sectors to Watch

I am not a financial advisor, and nothing in this section is investment advice. Always do your own research and talk to a licensed professional before making decisions with your money. But it would be silly to write a piece about the AI energy crisis without noting that this is one of the biggest capital-allocation stories of the decade.

Three categories are worth understanding:

1. Existing nuclear operators. Companies like Constellation Energy (the Microsoft partner) and Vistra (the Meta partner) operate large fleets of existing US nuclear plants. Their long-running asset bases have suddenly become enormously more valuable because hyperscalers are paying premium, multi-decade prices for output that, five years ago, was at risk of early retirement.

2. Small Modular Reactor developers. Public and pre-public SMR companies — including Oklo, NuScale, X-energy, Kairos Power, and TerraPower — represent the highest-risk, highest-potential category. None of them are operating commercial reactors yet. Many will fail or merge. The ones that succeed could become enormously consequential. This is venture-style risk inside the public market.

3. The "picks and shovels" supply chain. Companies producing enriched uranium, reactor components, nuclear fuel services, and specialized engineering — names like Cameco, Centrus Energy, BWX Technologies, and various utility-scale construction firms — stand to benefit regardless of which specific reactor technology wins.

Many investors are also looking at uranium spot prices and uranium-mining ETFs. Uranium has been on a multi-year run, partly because of this AI-driven demand thesis. Whether that continues depends on supply response, geopolitics (Russia and Kazakhstan remain key suppliers), and the pace of actual reactor deployment.

The Risks Nobody Is Talking About Loud Enough

Let's be balanced. There are real risks to the nuclear-AI thesis that deserve careful thought.

Construction delays. The Vogtle Units 3 and 4 in Georgia — America's most recent new nuclear builds — came in roughly seven years late and approximately $17 billion over budget. SMR economics depend on factory-style modular production that has never been demonstrated at scale in the US. There is a real chance many SMR projects slip from "early 2030s" to "late 2030s or beyond."

Regulatory pace. The Nuclear Regulatory Commission has been working to streamline approvals for advanced reactors, and the 2024 ADVANCE Act in Congress helped. But the historical NRC approval timeline for new reactor designs has been measured in decades, not years. If that pace doesn't accelerate meaningfully, the commercial timelines start to look heroic.

Fuel supply. Many advanced reactor designs require HALEU (high-assay low-enriched uranium), a fuel type that, until recently, was almost exclusively produced in Russia. Domestic supply chains are being built, but they are not yet at the scale needed to fuel a fleet of new reactors.

The "AI demand pause" scenario. If AI capital spending plateaus — whether because of model efficiency gains, regulatory limits, or just a softening of corporate AI budgets — the demand projections that underpin these 20-year nuclear PPAs could look very different. We've already seen 2025 efficiency gains (smaller, faster models) start to bend the curve somewhat.

None of these risks invalidate the thesis. They just mean the path is bumpier than the press releases suggest.

What Investors and Curious Readers Should Actually Do

If you take one practical thing away from this piece, let it be this: the AI energy crisis is not a future problem. It's a current bottleneck reshaping multi-trillion-dollar corporate strategies right now, in 2026.

For investors, the broad lesson is to think about AI exposure as a stack: there's the chips layer (Nvidia, AMD, TSMC), the model layer (OpenAI, Anthropic, Google), the application layer (everything from Cursor to your favorite AI tool), and — increasingly recognized — the infrastructure layer: power, cooling, water, transmission, and nuclear fuel. Most public-market attention has gone to the first three. The fourth is where many of 2026's most consequential corporate deals are happening.

For policy-curious readers, watch three things: (1) how state utility regulators handle the "who pays" question for data-center-driven grid investments; (2) how quickly the NRC actually licenses advanced reactor designs; and (3) whether congressional support for nuclear remains bipartisan as the topic gets politicized.

For everyone else — the people just trying to make sense of why their tech company is suddenly in the nuclear business — the simplest framing is this: the AI revolution has run into a fundamental physical constraint, and the response is not just bigger chips or smarter algorithms. It is rebuilding the American energy backbone. That is a story that will run for the next twenty years.

Frequently Asked Questions

How much electricity does an AI data center actually use?
A modern hyperscale AI training campus can demand anywhere from a few hundred megawatts to over a gigawatt continuously. A single 100 MW data center already consumes roughly the same electricity as a town of 80,000 people. Larger AI campuses can rival the entire load of a small city.

Why are Microsoft, Amazon, Google, and Meta all turning to nuclear in 2026?
Because their AI workloads need carbon-free, 24/7, baseload power at gigawatt scale, on contracts that can be locked in for 20 years. Nuclear is currently the only mature technology that meets all four criteria. Wind, solar, and storage can meet some, but not yet all, of those needs at the price points hyperscalers are willing to pay.

What is a small modular reactor (SMR) and is it safe?
An SMR is a factory-built nuclear unit, typically rated under 300 MW, designed to be transported and assembled on site. Most modern SMR designs use passive safety systems — meaning they shut down safely even without active intervention or external power — which is generally considered a significant safety improvement over older reactor architectures. They are not yet operating commercially in the US.

Will the AI energy crisis raise my electric bill?
Possibly, depending on where you live. If your local utility is investing heavily in transmission or generation to serve large data centers, some of those costs may be socialized across all ratepayers. Several state utility regulators are actively studying this question in 2026.

Is this a good time to invest in nuclear stocks?
This is not financial advice — please consult a licensed professional. That said, nuclear-adjacent equities have been on a strong multi-year run, driven by exactly the demand dynamics described in this article. The thesis is multi-decade, but valuations have already priced in significant optimism. Risk management matters as much as the thesis.

The Bottom Line

The largest, richest, most data-driven companies in the world have looked at the AI roadmap and concluded that without a massive expansion of nuclear power, the entire generative-AI buildout hits a ceiling within five to ten years. Their response — restarting Three Mile Island, signing 6.6 GW of Meta deals, betting on Oklo, Kairos, X-energy, NuScale, and TerraPower — is one of the largest sustained corporate investments in nuclear energy that the world has ever seen.

Whether you find that comforting or alarming, it is happening. The American power grid of 2035 will look very different from the one we have today, and the AI tools you're using right now are a major reason why.

The chip war defined 2023 and 2024. The model wars defined 2025. The power wars are defining 2026. Pay attention.

---

Disclaimer: This article is for informational and educational purposes only. It does not constitute financial, investment, legal, or tax advice. Always consult a qualified, licensed professional before making investment decisions. Nuclear energy, like all forms of energy generation, carries unique risks; readers are encouraged to research independently before forming views. All statistics and corporate announcements referenced here are publicly reported as of May 2026.