Microsoft, Amazon, Google and Meta have all signed nuclear deals to feed their AI data centers, and the pitch is seductive. Firm, carbon-free power running more than 90% of the time, on twenty-year contracts, matched to a load that never sleeps. So yes, nuclear is real, and everyone serious is in. But calling it the key is generous.
In the United States, nuclear supplied only about 20% of data-center electricity in 2024, behind natural gas at more than 40%, according to the IEA. The reactors doing the promised work are mostly a 2030s proposition. The honest read is nuclear as the headline and gas as the workhorse, and the whole question is whether the prestige bet arrives in time to matter.
What Has Actually Been Signed
As of September 2026, the SMR Intel tracker counted 13 US nuclear-to-data-centre deals totaling roughly 9.8 GW. The four hyperscalers account for most of the named capacity.
Meta leads the pack at up to 6.6 GW, spread across TerraPower, Oklo, Vistra, and Constellation. Amazon invested in X-energy, which has since closed $700 million-plus in funding rounds, for a planned SMR fleet. Google signed the first corporate SMR agreement, 500 MW with Kairos Power. Google has also contracted output from NextEra’s planned Duane Arnold restart in Iowa, about 615 MW, with a DOE loan closed in September 2026. Microsoft locked in a twenty-year PPA for the full output of a restarted reactor, about 835 MW.
For a decade, the tech energy story was wind and solar bought by the gigawatt. Nuclear on a hyperscaler balance sheet would have sounded implausible three years ago. What changed is the shape of AI demand. Training clusters want power that is on all the time, and nuclear’s capacity factor, above 90% against roughly 34% for wind and 24% for solar per US Department of Energy data, is the cleanest way to buy it without burning gas.
The Timeline Problem
Here is where the headline and the reality separate. Almost none of that 9.8 GW is delivering today.
The one deal actually moving electrons is Amazon’s. Talen’s Susquehanna plant in Pennsylvania already powers an adjacent Amazon campus, and that arrangement shifted to a full grid-connected PPA in spring 2026. But even here, the fine print bites. The contract ramps up to its full 1.92GW no later than 2032. Power is flowing, but not at headline volume.
Everything else is ahead of us. Microsoft’s restart of Three Mile Island Unit 1, rebranded the Crane Clean Energy Center, targets first power in 2027, and it leads the pipeline because it restarts an existing reactor rather than building a new one. The small modular reactors that make up most of the committed capacity, the Oklo, TerraPower, and X-energy machines, are 2030s hardware. Meta has the largest book and the longest wait, with deliveries stretching towards 2035.
So the pipeline is real and growing fast. The IEA notes that conditional SMR offtake agreements ballooned from 25 GW at the end of 2024 to 45 GW by early 2026. But a signed offtake is not a running reactor, and the gap between the two is close to a decade.
Who Fills the Gap
If nuclear is a 2030s answer and AI demand is a right-now problem, something has to bridge the years in between. That something is gas, and increasingly, the grid connection itself.
Natural gas is the largest single source of new supply to data centers through 2030, adding over 130TWh of annual generation on IEA numbers, much of it built behind the meter in the US to dodge grid queues. The near-term buildout is being powered by turbines, not reactors, which is why gas turbine supply from the likes of GE Vernova has become one of the tightest bottlenecks in the entire chain.
The other bottleneck is the wire. Interconnection queues, not generation, are increasingly what decides whether a campus comes online, and the Susquehanna case triggered a broader FERC rulemaking on how large co-located loads connect at all. For anyone pricing this, the margin over the next five years sits less with the reactor developers and more with the deregulated operators who own running plants today, Constellation, Vistra, and Talen, plus whoever can supply a turbine or clear an interconnection.
Is it Even Reachable?
There is a harder question underneath all of this. Even if nuclear delivers on schedule, is it available to anyone other than a top-four hyperscaler?
The evidence says not really, at least not yet. Direct nuclear PPAs need a deregulated power market and a counterparty large enough to underwrite twenty years of offtake. Most US reactors are owned by regulated utilities that cannot sign bilateral deals with a single buyer, and only a handful of operators can transact this way at all. That structurally narrows the field to companies with hyperscaler balance sheets. The SMR developers promising to widen access are real, but they will not deliver commercial megawatts until 2027 at the earliest and mostly later. For a mid-sized AI infrastructure player siting capacity in 2027, nuclear is not a procurement option. Gas, renewables plus storage, and a good grid connection are.
The Verdict
Nuclear is not the key to the AI buildout. It is the prestige bet on the far end of it, and a rational one, because past 2030, the reactors and SMRs coming online genuinely reshape the mix towards firm clean power. But the buildout happening now, the one deciding which models get trained this year and next, runs on gas and grid, not fission.
The nuclear story is a bet on the next decade dressed up as a solution to this one. For investors, the near-term margin is in the boring middle of the stack. For anyone actually building, the reactor is a 2030s luxury, and the workhorse is already in the barn.
Author: Ayanfe Fakunle
See Also:
What Country in Europe Has the Most Data Centres?
Data Centres Are Draining Fresh Water. Why Not Just Use the Sea?
