Two years ago, talking about small nuclear reactors powering a data center was still the stuff of conference panels. Today it’s a line item in the investment plans of Google, Amazon and Microsoft. In a handful of quarters, the topic went from PowerPoint to the trading floor, with direct commitments now exceeding ten billion dollars. That shift is worth pausing on, because it says something broader about the state of power supply for compute infrastructure.
Why nuclear, and why now
The trigger is simple: AI consumes enormous amounts of power, and the grid can no longer keep up. Training clusters for large language models demand dense, stable, around-the-clock power. Yet in the major hubs, grid interconnection queues are now measured in years. In the US, the PJM grid operator alone reports a queue of more than 286 GW of pending interconnection requests — more than its current installed capacity.
Faced with that wall, hyperscalers are chasing every form of firm power available. Solar and wind remain essential, but their intermittency doesn’t match the profile of an IT load that runs 24/7 and demands “five nines” availability (99.999%). Nuclear, by contrast, delivers capacity factors of 90 to 95% and carbon-free electricity. On paper, it’s the ideal answer to a need for clean, permanent base load.
SMRs: what the acronym actually covers
SMR stands for Small Modular Reactor. Three features set it apart from conventional plants.
First, size: these reactors typically top out around 300 MW electrical, versus 900 to 1,600 MW for a traditional unit. That’s precisely the scale of a large data center campus or cluster.
Second, modularity: these reactors are designed to be factory-built and assembled on-site, rather than constructed piece by piece in the field. The goal is to cut timelines, capital costs and the overrun risk that plagues large nuclear projects.
Third, proximity: with reduced emergency planning zones, an SMR can be sited close to the load it powers. That’s the whole appeal of the “behind the meter” model: the reactor feeds the data center directly via a microgrid, without clogging already-saturated national transmission lines. The reactor’s waste heat can even be recovered for other uses, including cooling.
The technologies vary: small pressurized water reactors (Holtec SMR-300, NuScale), high-temperature gas-cooled reactors (X-energy Xe-100), or molten salt reactors (Kairos Power). All aim at the same spec sheet: reliability, compactness, series production.
The deals that tipped the market
The best way to gauge how serious the trend is: follow the money.
Google and Kairos Power. In October 2024, Google signed the first corporate agreement to deploy multiple SMRs in the US: up to 500 MW of molten salt reactors by 2035, across six to seven units, with the first reactor expected by 2030. The first concrete milestone, the Hermes 2 plant in Tennessee, will deliver up to 50 MW to the Tennessee Valley Authority grid to power Google data centers in Tennessee and Alabama.
Amazon and X-energy. In April 2026, startup X-energy pulled off a closely watched IPO on the Nasdaq (ticker XE), raising about $1.02 billion at a valuation of roughly $9.1 billion at the offering price. The stock jumped 27% on its first day. Behind the success: Amazon, whose commitment to support deployment of more than 5 GW of Xe-100 reactors by 2039 effectively underwrites much of the order book. Each Xe-100 unit produces around 80 MW electrical, from a high-temperature helium-cooled reactor using TRISO fuel.
Microsoft and Constellation. In September 2024, Microsoft secured the restart of Three Mile Island Unit 1, via Constellation Energy, for an estimated $1.6 billion — that’s 835 MW of carbon-free capacity committed through 2054. It isn’t an SMR, but it’s the same reflex: lock in firm nuclear power to underpin AI growth.
EDF, Holtec and the Cottam project in the UK. On the site of the former Cottam coal-fired power station in Nottinghamshire, EDF and Holtec submitted a joint proposal to develop SMR-300 reactors to power a 1 GW data center campus. One nuance is worth flagging: the data center will initially be supplied by the grid and renewables, with the SMR element only becoming operational in the 2030s. It would nonetheless be one of the first SMR-coupled data centers in Europe.
Timeline, or the great misunderstanding
This is where a cool head matters. The announcements are massive, but they mostly describe deployments on a 2030-2039 horizon. None of these SMRs is powering a data center today. The first credible commissioning dates land around 2030 — and even then, subject to regulatory approval.
Because the real bottleneck is no longer financial — the money is there — but regulatory. The speed at which safety authorities (the NRC in the US, France’s ASN, the UK regulators) certify these new designs will determine the actual timeline. SMRs remain, to date, a promising but not yet commercially proven technology. The more optimistic analysts estimate that by 2035, SMRs could cover roughly 10% of the expected rise in US data center electricity demand. That’s significant, but a long way from a near-term silver bullet.
What it means if you’re sourcing power today
If you’re planning AI, HPC or cloud workloads, two readings coexist.
In the long run, next-generation nuclear will expand the pool of carbon-free, permanent capacity. That’s good structural news: more firm power will eventually reach the market, and some campuses will be directly backed by a dedicated reactor.
In the short and medium term, none of this solves your need for 2026, 2027 or 2028. The power you book over the next two years will come from the grid, from renewable power purchase agreements, and from operators’ existing capacity. The nuclear announcements shouldn’t be an excuse to defer a sourcing decision: on the contrary, they confirm that firm power is scarce and strategic, which only strengthens the case for securing what’s available now, early.
For a data center operator or developer, the signal is just as clear. Tomorrow’s capacity will increasingly be structured around the energy question: backing from a firm source, microgrid, decarbonization profile. Being able to present that dimension will become a sales argument in its own right with buyers.
In short
The marriage of nuclear and data centers is no longer a hypothesis: it’s a funded trajectory, driven by grid saturation in the face of AI. But between the 2026 commitments and the first megawatts actually delivered, there’s a gap of several years, marked out by regulatory uncertainty. SMRs will reshape supply in the long run. Until then, securing available capacity remains the most concrete decision an infrastructure team can make.
If your roadmap touches the coming years, the time to map your power options is now — not 2030.
Looking to secure data center power for your AI, HPC or cloud workloads? Describe your needs and receive a tailored proposal within 48 hours.