Why AI Is Reviving Nuclear Power (and Uranium)

Published: July 18, 2026

The AI Data Centre That Needed Its Own Power Plant

In 1979 the Three Mile Island nuclear plant in Pennsylvania suffered a partial meltdown, the worst commercial nuclear accident in American history. The reactor that melted never ran again, public trust collapsed, and new nuclear construction in the United States effectively stopped for a generation. The plant’s second reactor, Unit 1, kept running cleanly for another forty years, then closed in 2019 because it could not compete with cheap natural gas. That looked like the end of the story.

It was not. In 2024 the plant’s owner, Constellation Energy, announced it would restart that shuttered reactor and rename the site the Crane Clean Energy Center, selling its output under a twenty-year contract to Microsoft to power the company’s data centres. The reason is the same force reshaping the entire electricity industry: artificial intelligence needs more power than the grid can supply, and it needs that power to be constant, carbon-free, and available now. That demand has revived nuclear energy, and behind nuclear energy sits a metal that has to come out of the ground to feed it — uranium. This is the chain that runs from an AI server rack to a mine, start to finish.

Why AI Data Centres Are Turning to Nuclear Power

An AI data centre is a uniquely demanding electricity customer. It runs flat out around the clock, it draws its load in one concentrated place rather than spread across a city, and the companies building it have made public promises to run on carbon-free power. Those three requirements — constant, concentrated, and clean — rule out most of the options a normal factory would consider, and they point straight at nuclear. A reactor produces a steady gigawatt-scale stream of electricity 24 hours a day, emits no carbon dioxide while running, and does it from a footprint small enough to sit beside the servers it feeds.

The comparison is what makes the case. Natural gas is fast to build but emits carbon and now faces multi-year waiting lists for turbines. Solar and wind are cheap and clean but stop when the sun sets or the wind drops, so matching a data centre’s round-the-clock load means overbuilding generation and adding enormous batteries, which inflates both cost and land use. Waiting for the public grid to add capacity means waiting years for transmission lines that may never get approved. Against that field, nuclear is the only source that clears all three bars at once — which is why every alternative a hyperscaler looks at ends up being a downgrade on the requirement it cares about most: firm, clean power it can count on every hour of every day.

That mismatch — between the firm gigawatt AI wants in one place and what a strained grid can actually hand over — is the whole reason nuclear re-entered the conversation. The scale of the demand is what turned it from an idea into signed contracts.

How Much Electricity AI Actually Needs

The scale of the demand is what turned a niche idea into a national power story. US data centres used about 58 terawatt-hours of electricity in 2014. By 2023 that had climbed to roughly 176 terawatt-hours, about 4.4% of all electricity consumed in the country, after more than doubling between 2017 and 2023 largely because of AI servers. A December 2024 analysis from the US Department of Energy’s Lawrence Berkeley National Laboratory projects the figure will reach between 325 and 580 terawatt-hours by 2028 — somewhere between 6.7% and 12% of total US electricity. In under a decade, the industry’s share of the grid roughly triples.

The reason a terawatt-hour figure translates into reactor deals is density. A single large AI campus can want a gigawatt of firm power in one spot, and the utility serving that region simply does not have a spare gigawatt of clean, always-on generation to hand over. The same buildout is straining every input at once, from transmission and transformers to the copper that wires a data centre, and power has become the hardest constraint of all. Faced with grid queues measured in years, the largest technology companies decided to secure generation directly rather than wait in line.

The Nuclear Deals Hyperscalers Have Signed

The clearest evidence that AI is reviving nuclear is the contracts. Microsoft’s agreement to restart Three Mile Island’s surviving reactor is the most striking, but it is one of a wave. Constellation says the Crane Clean Energy Center will add about 835 megawatts to the grid under a twenty-year deal with Microsoft, with the reactor expected back online in 2028. Amazon has taken a different route, buying a data-centre campus next to an existing Pennsylvania nuclear plant and separately backing the small-reactor developer X-energy. Google signed with Kairos Power, Meta contracted for a fleet of advanced reactors, and Oracle disclosed plans to power a gigawatt-scale data centre with three small modular reactors.

Add the announced commitments together and the technology industry has lined up well over ten gigawatts of nuclear capacity, some of it restarts of shut reactors, some new power-purchase deals with operating plants, and much of it advanced reactors that do not yet exist. The split matters: restarting a paused reactor or signing for an operating one adds real power this decade, while the small-reactor deals are bets on plants that will not generate until the 2030s.

Small Modular Reactors and the AI Data Centre

Small modular reactors are the piece of this story most tied to the future rather than the present. An SMR is a reactor built small — typically under 300 megawatts — and assembled from factory-made modules rather than poured on-site like a conventional plant. That design promises faster construction, lower upfront cost, and a footprint compact enough to place directly next to a data centre, which is exactly why hyperscalers find them attractive. Google, Amazon, Oracle, and Meta have all backed SMR developers, betting that a fleet of small reactors can be deployed alongside computing capacity the way a backup generator once was.

The honest caveat sits right beside the promise: almost no commercial SMR is operating yet in the West. Most designs are still working through regulatory approval and first-of-a-kind construction, and the earliest units these deals depend on are targeted for around 2030 or later. There is also a fuel wrinkle. Many advanced reactor designs run on high-assay low-enriched uranium, or HALEU, enriched to between 5% and 20% — higher than the roughly 5% conventional reactors use — and until recently almost the only commercial source of HALEU was Russia. Building Western enrichment capacity for it is a bottleneck the whole SMR wave has to clear, and it links the reactor story directly back to uranium supply.

Is There Enough Uranium for a Nuclear Revival?

Every reactor in every one of these deals runs on the same fuel, and that raises the question the market is now wrestling with: is there enough uranium coming out of the ground to feed a nuclear expansion this large? The answer is not that the world is short of uranium in the crust — there is plenty in the ground — but that mines are not producing it fast enough, and the buffer that used to cover the gap is thinning. The sections that follow put numbers on how much uranium a nuclear revival needs, why mine supply cannot simply ramp to meet it, and how investors are positioning around the squeeze. What follows is market analysis for information only, not investment advice.

How Much Uranium the Nuclear Revival Will Need

Start with the fuel appetite of a single reactor. Keeping one large 1,000-megawatt reactor running for a year takes about 200 tonnes of natural uranium, which is milled, converted, and enriched down into roughly 20 to 27 tonnes of finished fuel. Multiply that across the world’s operating fleet and the numbers get large fast. The World Nuclear Association’s 2025 fuel report puts global reactor requirements at roughly 68,920 tonnes of uranium in 2025, up about 3% on the year, and projects that figure rising to just over 150,000 tonnes by 2040 in its reference scenario — a near doubling.

That projection was drawn up as reactor restarts, life extensions, new builds in Asia, and the first SMRs all stacked onto demand at once, before the full weight of the AI-driven deals was even visible. Each restarted reactor and each new SMR adds another 100 to 200 tonnes a year of uranium demand on top of the baseline, and the United States has floated ambitions to quadruple its nuclear capacity by 2050 — an increase that would, on its own, require roughly doubling today’s global uranium production just for one country. Demand is not the uncertain side of this equation. Supply is.

The Uranium Supply Squeeze

The gap is already here. In 2024 the world’s uranium mines produced 60,213 tonnes of uranium, about 90% of reactor requirements that year. The other tenth came from what the industry calls secondary supply: stockpiles built up over decades, uranium recovered from old weapons, and material squeezed out of the enrichment process. For years that secondary cushion papered over the shortfall between what mines dig up and what reactors consume. The problem is that the cushion is finite and shrinking, while demand is climbing.

Underinvestment made it worse. According to Sprott, 2025 marked the thirteenth consecutive year in which uranium contracting fell short of replacement needs, meaning utilities kept drawing down inventories instead of signing enough new supply. That is the setup behind the price move: after years near $50 a pound, uranium surged past $101 a pound in January 2026 for the first time in about two years, before consolidating into the mid-$80s through the second quarter as the spot market caught its breath. Long-term contract prices, which matter more for financing new mines, broke above $90 for the first time since 2008.

Why Uranium Mining Can’t Simply Ramp Up

High prices are supposed to summon new supply, and eventually they will — but uranium responds slowly, for reasons built into the industry. A new mine takes ten to fifteen years to move from discovery through permitting, financing, and construction to first production, so a price signal in 2026 does not become metal until the late 2030s. Miners also carry scars: the 2011 Fukushima accident crushed the uranium price for a decade and bankrupted or idled much of the sector, which made producers cautious about restarting even as prices recovered. Supply that was switched off does not come back the moment the price rises.

Concentration compounds the fragility. A single country, Kazakhstan, mines about 39% of the world’s uranium, with Canada at 24% and Namibia at 12% — so the market leans heavily on a handful of producers and one dominant one. Most of that Kazakh output uses in-situ leaching, and the country has faced its own constraints on sulphuric acid and expansion. On top of mining, enrichment — the step that turns mined uranium into reactor fuel — is dominated by Russia, adding a geopolitical chokepoint downstream of the mine. Even US domestic production, which policymakers want to grow, fell about 44% in the third quarter of 2025 as older operations wound down. The supply side is inelastic in exactly the moment demand is inflecting.

None of these constraints is permanent, but all of them take years to ease — which is why the gap between what reactors need and what mines deliver is set to persist even as prices climb.

What the Nuclear Revival Means for Uranium Investors

A rising demand curve meeting an inelastic supply curve is the classic setup that draws investors to a commodity, and uranium has attracted a wave of them. The exposure ranges from owning the physical metal to owning the companies that mine it, enrich it, or burn it — each with a different relationship to the underlying squeeze and a different risk. The purest way to hold the metal is a physical uranium trust, which buys and stores actual pounds and, in doing so, pulls supply off the spot market and tightens it further. Miners such as Cameco and Kazakhstan’s Kazatomprom offer leveraged exposure to the price. Diversified uranium and nuclear ETFs spread the bet across the sector, and a speculative tier of small-reactor and enrichment developers is tied to the buildout rather than the metal.

The chokepoints are the risks. Supply concentration in Kazakhstan and enrichment concentration in Russia mean geopolitics can move the price in either direction. The small-reactor developers powering the AI narrative are mostly pre-revenue, betting on plants that will not operate until the 2030s, and the uranium price itself is famously volatile — it has round-tripped from booms to decade-long busts before. The demand story is real and the supply story is tight, but neither guarantees the timing or the direction of any particular security.

The Outlook for Nuclear Power and Uranium

The next decade points toward tight uranium and expanding nuclear, with the pace set by how fast mines and enrichment can catch up to demand that is now inflecting upward. The World Nuclear Association’s reference case has reactor requirements nearly doubling by 2040, and the AI-driven deals of the mid-2020s were mostly signed after that projection was drawn, which suggests the demand side has more upside than downside. On supply, the response is coming — idled mines are restarting, Kazatomprom is guiding production higher, and the West is finally investing in its own enrichment — but the ten-to-fifteen-year lead times mean the gap is more likely to widen before it closes. Further out, fusion sits on the horizon as the technology that could eventually change the entire equation, though the magnets that make fusion possible are themselves still being proven, and commercial fusion power remains decades away. For the foreseeable future, the metal that powers the AI era is uranium, and the reactors that burn it are being switched back on faster than the mines can refill.

A Short History of Nuclear Power

Nuclear Power and Uranium FAQ

Related Articles