Nuclear Bets Grow for AI Data Centers, But Power Lags Years

What happened

A wave of nuclear-industry activity has attached itself to the AI boom, but the underlying story splits into two very different timelines. On one side, tech companies keep signing headline-grabbing power deals with nuclear developers. On the other, fuel and reactor technology that could make small modular reactors safer and more sitable is still years from commercial scale. Standard Nuclear, an Oak Ridge, Tennessee, company led by nuclear engineer Kurt Terrani, has emerged as a bet on the second half of that story: it went public with a reported $576.9 million contract backlog and deals to supply TRISO fuel to Antares Nuclear, Radiant Industries and the U.S. Army, arguing that fuel incapable of melting down could make small reactors easier to site near data centers, military bases and industrial plants 1.

Meanwhile, hyperscalers have been signing actual power agreements. Microsoft committed to a 20-year, roughly $16 billion purchase of the entire output of the restarted Three Mile Island Unit 1 — now rebranded the Crane Clean Energy Center — with the reactor targeted for a 2027 or 2028 return to service 910. Google struck a deal for up to 500 megawatts from Kairos Power's molten-salt-cooled reactors, with Tennessee Valley Authority set to buy the electricity for delivery around 2030 910. Amazon has backed X-energy's Xe-100 high-temperature gas reactor design and is spending more than $20 billion converting Talen Energy's Susquehanna site into a nuclear-linked AI campus 910. Meta has assembled the largest and most diversified nuclear portfolio, with potential commitments spanning TerraPower's Natrium reactor, Oklo's Aurora design, and existing-fleet purchase agreements with Vistra and Constellation, adding up to several gigawatts on paper 9.

Why AI data centers are driving this

The reason all of this is happening now is electricity, not ideology. The International Energy Agency estimates global data centers consumed about 415 terawatt-hours in 2024 — roughly 1.5% of global electricity use — and projects that figure could reach 945 TWh by 2030 in its base case, with AI-optimized "accelerated servers" growing consumption around 30% annually 68. That demand is landing in concentrated geographic clusters rather than spreading evenly across grids, which makes it harder to absorb than typical household electrification, and it's arriving on a two-to-three-year data-center construction timeline that outruns the years-to-decades timeline of new generation and transmission projects 6. Individual AI campuses can require gigawatt-scale power — Stargate's Abilene, Texas project and a planned Meta campus in Louisiana are both cited as potentially needing up to 5 gigawatts each 7.

Nuclear's pitch to this market rests on round-the-clock output, low operating emissions and a small physical footprint, qualities that make it a theoretically strong match for AI workloads that can't tolerate interruptions. TRISO fuel, at the center of Standard Nuclear's business, is designed to reinforce that pitch: uranium kernels wrapped in ceramic layers that the International Atomic Energy Agency says can withstand temperatures near 1,700°C, reducing reliance on the massive containment structures conventional reactors require. But TRISO and related advanced fuels still depend on HALEU — uranium enriched between 5% and 20% — which the U.S. Nuclear Regulatory Commission notes brings its own transport, security and licensing hurdles that haven't yet been resolved at commercial scale.

Where the reporting agrees

Across the trade coverage, the IEA's modeling and even the harshest critique, several facts are not in dispute. Data-center electricity demand is rising fast and is heavily driven by AI-specific hardware, not general IT growth 68. Nuclear currently supplies a modest but real share of data-center electricity — the IEA puts it at about 15% globally and around 20% in the United States today — while gas, coal and renewables carry the larger present-day load 68. Every outlet describing hyperscaler activity agrees on the basic shape of the four biggest deals: Microsoft's Three Mile Island restart PPA, Google's Kairos Power SMR agreement, Amazon's X-energy investment and Susquehanna campus, and Meta's multi-partner nuclear strategy 910. There is also consensus that most of this capacity is not yet generating electricity — the IEA, the Bulletin of the Atomic Scientists and industry trackers alike describe the bulk of announced gigawattage as planned, contracted or under construction rather than operating 67910. And nobody disputes that new nuclear construction, especially of first-of-a-kind SMR designs, has historically run over budget and behind schedule, with Vogtle in Georgia — final cost above $36 billion, more than double its original estimate — cited repeatedly as the cautionary example 7.

Where it doesn't

The real divergence is over what these facts mean. Forbes-style coverage of Standard Nuclear frames safer fuel as a genuine unlock: if TRISO reduces meltdown risk, it could widen where and how small reactors get built, expanding a market well beyond data centers into military and industrial uses 1. The IEA's modeling treats nuclear as a credible, if slow-arriving, contributor — its base case has SMRs entering the data-center supply mix meaningfully only after 2030, with the U.S. reaching over 20 GW of hyperscaler-financed SMR plans, while gas adds the largest near-term supply 8.

The Bulletin of the Atomic Scientists rejects that framing almost entirely, arguing the nuclear-for-AI narrative is closer to greenwashing than to an energy strategy. It points out that Amazon's original announcement only funded a feasibility phase and secured a right, not an obligation, to buy power, and that Google's Kairos deal disclosed no pricing terms at all 7. It also sets corporate nuclear spending against corporate data-center budgets — Amazon and Google each expected to spend up to $200 billion and $185 billion respectively on data centers, dwarfing nuclear commitments in the hundreds of millions 7. On the raw numbers, the Bulletin calculates that a 300-MW SMR run continuously would produce about 2.6 TWh a year, against U.S. data-center demand the Bulletin puts at 312.6 TWh in 2025 — and notes U.S. nuclear generation actually declined slightly from 2020 to 2025 even as data-center demand surged, meaning nuclear has contributed nothing to the increase so far 7.

There's also a numbers mismatch worth flagging directly: different trackers cite different totals for aggregate hyperscaler nuclear commitments — one industry tracker puts it near 9.8 GW across roughly 13 deals, another cites more than 10 GW signed in a single year, and the IEA separately notes tech companies have announced plans to finance more than 20 GW of SMRs 8910. These aren't necessarily contradictory — they measure different things (signed deals versus financing plans versus a specific 12-month window) — but they get cited almost interchangeably in coverage, which risks overstating how much capacity is actually locked in.

The verdict

The evidence best supports a narrower reading than either the optimistic fuel-innovation framing or the fully dismissive one. Safer fuel like TRISO is a real technical development that could matter for siting and licensing small reactors, and Standard Nuclear's backlog shows genuine commercial interest exists. But the Bulletin's math is hard to argue with on timing and scale: no U.S. SMR is under construction today, nuclear's actual contribution to the current AI demand surge is effectively zero, and the dollar figures hyperscalers have committed to nuclear remain a rounding error against their data-center capital budgets. The IEA's own modeling — produced independent of that critique — arrives at a similar structural conclusion from the supply side: gas and renewables will carry the 2020s, and nuclear becomes meaningful only toward and after 2030. What's happening now looks less like a nuclear boom already underway and more like options being purchased against one — bankable, plausible, and still fundamentally unbuilt.