Why Big Tech Is Betting on Nuclear Power for AI Data Centers

01 The Electricity Bill Behind the AI Boom

Training and running large AI models is an industrial-scale electricity problem. A modern accelerator rack draws tens of kilowatts, and a large training cluster can pull hundreds of megawatts continuously, roughly the load of a mid-sized city. Inference is compounding the demand: every chat query, image generation and agent step spends real watt-hours in a data center, and usage is growing faster than training workloads because millions of people now touch these systems daily.

Quantifying the total is genuinely difficult because companies treat consumption figures as confidential. The most widely cited institutional estimate comes from the International Energy Agency, which put global data center electricity use at roughly 415 TWh in 2024, around 1.5 percent of world demand, and projected it could roughly double to about 945 TWh by 2030. These are estimates, not meter readings, and the IEA itself attaches wide uncertainty bands to the projection. Still, even the low end of the range implies the largest sustained jump in a single sector's power demand in decades.

Utilities, which spent years watching flat demand, are suddenly revising load forecasts upward, and grid operators in several US regions have flagged that large interconnection requests from data centers now stretch into the gigawatt range. That mismatch between fast-moving compute roadmaps and slow-moving infrastructure is the backdrop for everything else in this story.

02 The Deals: Three Different Bets on the Same Answer

The most concrete deal came in September 2024, when Microsoft signed a 20-year power purchase agreement with Constellation Energy to restart Unit 1 of Three Mile Island, the Pennsylvania reactor that shut down in 2019 for economic reasons. The unit's roughly 835 megawatts are slated to flow to the grid with Microsoft effectively buying the clean-energy attributes to match its Mid-Atlantic data center load. Restarting an existing plant sidesteps the decade-plus timeline of new construction, though it still required Nuclear Regulatory Commission review, fuel procurement and turbine work before the unit could return to service.

Amazon took a different route: it agreed to acquire a data center campus adjacent to the Susquehanna nuclear plant in Pennsylvania from Talen Energy, with the site designed to draw power directly from the plant in a configuration known as behind-the-meter service. The arrangement attracted regulatory attention because the grid operator and intervenors argued that letting a large customer bypass transmission charges could shift costs to other ratepayers, a dispute that has shaped how later deals are structured. Google, meanwhile, became the first hyperscaler to order small modular reactors outright, signing an agreement with Kairos Power targeting an initial 500 megawatts of SMR capacity, with a first unit planned in the early 2030s.

Read together, the three deals form a portfolio approach: restart proven capacity fast, co-locate load next to existing plants, and place long-lead orders on next-generation technology. Meta has explored similar territory, and the underlying logic is identical for all of them: these companies are not buying electricity as a commodity, they are buying certainty.

03 Why Nuclear Fits the Problem

Three physical properties make nuclear unusually well matched to AI workloads. The first is baseload reliability: a reactor runs around the clock at high capacity factor, typically above 90 percent for the US fleet, which mirrors the way training clusters and always-on inference fleets consume power. Solar and wind, whatever their cost advantages, are intermittent, and batteries cover hours rather than weeks. For a company whose million-dollar training run fails if the power blinks, availability is the product.

The second is energy density. A conventional light-water reactor enriches uranium to a few percent of the fissile isotope U-235, and that low-enriched fuel carries on the order of 120,000 times more usable energy per unit mass than coal, a comparison documented on Wikipedia's nuclear power pages and in standard references. Density matters less for fuel logistics than for footprint: a 900 MW reactor campus occupies a few square kilometers, where the equivalent wind capacity would sprawl across hundreds of them, a real constraint in the land-constrained regions near major fiber routes and metro areas.

The third is carbon accounting. Microsoft, Google and Amazon have all made public commitments to net-zero emissions, and grid purchases, even at renewable-heavy hours, cannot credibly cover 24/7 compute growth in coal-heavy regions. Nuclear is the only dispatchable, carbon-free generation technology with a multi-gigawatt track record already operating. These companies are buying something they cannot buy anywhere else at scale: firm clean power.

04 What the Companies Are Actually Paying

Nuclear power is not cheap electricity. US nuclear plants have among the highest levelized costs of any mainstream generation source, driven by enormous up-front capital requirements, staffing, security and waste-handling obligations. The specific price terms of the Microsoft-Constellation deal have not been disclosed, but reporting by CNBC and trade outlets consistently describes these AI-driven PPAs as struck at a premium above prevailing wholesale rates, with the buyer paying for decades of guaranteed availability rather than for the cheapest megawatt-hour.

The economics nonetheless pencil out from the buyer's side under one assumption: that AI revenue will keep scaling. A hyperscaler with operating margins in the tens of billions can afford to overpay for power certainty if the compute it feeds generates returns several times the premium. That is a bet on the durability of AI demand, and it is worth being explicit that it is an economic bet, not an engineering conclusion. If AI revenue growth slows materially, the same contracts become a fixed cost burden.

For plant owners, the deals reverse a bleak decade. Reactors that closed early, including Three Mile Island Unit 1, did so because cheap natural gas and flat demand made them uncompetitive. A buyer with a 20-year horizon and a premium tolerance changes the entire financing picture, which is why Constellation's leadership described the Microsoft agreement as a template rather than a one-off, and why other operators have since marketed restart and uprate opportunities directly to the tech sector.

05 The Obstacles Between Contract and Current

Signing a PPA is the easy part. Restarting a shuttered reactor is a category of project the US nuclear industry has rarely executed, and while the regulatory framework exists, fuel reordering, staffing back up, and requalifying systems that sat idle for years carry real cost and schedule risk. Grid interconnection is a second choke point: the Federal Energy Regulatory Commission's intervention in the Amazon-Talen configuration showed that co-locating load at a plant is legally contested territory, and every hyperscaler deal since has had to thread the same needle between private benefit and public grid cost.

The fuel supply chain is a quieter concern. The US relies on foreign enrichment for a large share of its low-enriched uranium, and a fleet of restarts, uprates and new SMRs would multiply demand on a supply base that was sized for a shrinking market. Fuel contracts are signed years ahead, and new enrichment capacity has its own multi-year build time.

For the SMR path specifically, caution is warranted. NuScale, the company that carried the flag for small reactors in the US, saw its flagship carbon project canceled in late 2023 after subscription costs rose, and other advanced-reactor developers have slipped schedules. Kairos Power's fluoride-salt-cooled design is promising on paper but remains pre-commercial: no US SMR has yet delivered a single grid megawatt-hour. Google's order is a bet on engineering maturing on schedule, which is a bet the nuclear industry has historically lost more often than won.

06 What Could Make the Curve Bend

Every projection in this story rests on AI demand continuing to grow, and that assumption deserves scrutiny. On the demand side, model usage genuinely keeps rising, and agentic applications that run continuously would push consumption up, not down. But on the efficiency side, the compute required per unit of AI capability has fallen dramatically across model generations, and better inference scheduling, quantization and purpose-built silicon could all flatten the electricity curve. The IEA's doubling scenario assumes significant efficiency gains already; a faster efficiency trajectory would shrink it.

There is also a substitution question. If grid buildout, renewable additions and storage costs keep falling, hyperscalers may find that a mixed portfolio of wind, solar and batteries delivers firm-enough power for most workloads, reserving nuclear premiums for the most latency- and availability-critical sites. In that world, today's nuclear deals still make sense as hedge capacity, but they do not become the backbone of AI power.

The measured facts, as distinct from projections, are these: data centers consumed roughly 415 TWh in 2024, a handful of large nuclear contracts totaling a few gigawatts have been signed, and no SMR yet operates commercially in the US. Everything else, including the doubling forecast and the strategic framing, is interpretation. The bet the hyperscalers are making is not that nuclear is necessary for AI in any physical sense; it is that paying a premium for certainty is cheaper than the alternative of being wrong about availability.