How SMRs Could Solve the Data Center Energy Crisis
AI is eating the grid alive — and small modular reactors may be the only power source that actually fits the bill.
Something extraordinary happened in April 2026: the IEA published a report finding that the pipeline of conditional offtake agreements between data center operators and SMR projects had grown from 25 gigawatts at the end of 2024 to 45 gigawatts today. In 16 months, the industry doubled its nuclear commitments. That’s not a trend. That’s a panic buy.
The reason is straightforward: AI is consuming electricity at a pace the grid was never designed to handle. Data center electricity demand jumped 17% in 2026, and AI-focused facilities grew even faster, according to the same IEA report. The five biggest tech companies, Amazon Web Services, Google, Meta, Microsoft, and Equinix, poured more than $400 billion into data center capital expenditure in 2025 alone — a figure the IEA notes now exceeds global investment in oil and gas production combined. And demand is set to increase by another 75% in 2026.
The grid is not keeping up. Grid interconnection queues in many U.S. regions now stretch five to ten years, according to Shumaker, Loop & Kendrick’s analysis of the nuclear-AI intersection. Northern Virginia, home to the largest concentration of data centers on earth, has utilities openly warning that substations and high-voltage lines are approaching their limits. PJM Interconnection — the grid operator covering much of the U.S. Mid-Atlantic — has issued warnings about summer peak capacity shortfalls.
So hyperscalers are looking off-grid. And nuclear, specifically SMRs, is starting to look less like an idealistic long shot and more like the only option with the right profile. Whether the industry can actually deliver in time is a very different question.
Why renewables alone can’t do this
Hear me out before dismissing this section as pro-nuclear boosterism. The problem with solar and wind for data centers is not that they’re bad. It’s that AI workloads don’t care what the weather is doing. 🌤️
A large-scale AI training run consumes electricity continuously. When you’re spending millions of dollars in compute time to train a frontier model, you cannot pause for cloud cover over a solar farm in Texas. Nuclear SMRs operate at capacity factors exceeding 92%, compared to roughly 25–30% for solar and 30–40% for wind, per market research cited by Data Center Dynamics. That gap is the whole ballgame. You can’t battery-storage your way out of it when you’re drawing hundreds of megawatts around the clock.
There are other mismatches worth naming:
Solar and wind require vast land areas; a single SMR plant occupying a few acres can match what a solar farm covering many square miles would produce
AI data centers have “rapid and large swings in demand,” per the IEA, which strains the technical capabilities of on-site gas plants trying to track load
Renewable energy certificates (RECs) are increasingly failing to satisfy corporate sustainability commitments, because they allow companies to claim clean power while actually drawing from fossil sources 24/7
Grid interconnection backlogs of 5–10 years mean that even if a utility agrees to add renewable capacity, hyperscalers can’t count on timely access to transmission
None of this means renewables are useless for data centers. They’re already covering enormous swaths of data center load, and the Brookings Institution’s April 2026 analysis notes that Big Tech accounted for 43% of all clean energy power purchase agreements globally in 2024. But renewable PPAs work beautifully for general corporate sustainability targets and don’t work well for the specific, non-negotiable uptime requirements of a 100,000-GPU AI training cluster. ⚡ That’s where nuclear steps in.
The deals that changed the conversation
The shift in nuclear’s reputation happened fast, and it happened because a few big bets got placed very publicly.
Microsoft committed to a 20-year, 835-megawatt power purchase agreement to restart Three Mile Island, targeting 2028 — a move that felt almost deliberately provocative given Three Mile Island’s history. The deal with Constellation Energy is worth an estimated $16 billion over its life. Microsoft has also assembled an internal nuclear team, hiring directors of atomic technology from Ultra Safe Nuclear and the Tennessee Valley Authority.
Amazon went further. AWS backed 5 gigawatts of X-energy SMR capacity through a $500 million direct investment, signed an agreement with Energy Northwest in Washington state to fund four Xe-100 reactors producing 320 MW with expansion potential to 960 MW, and committed $20 billion to convert the Susquehanna nuclear site into an AI data center campus. Matt Garman, CEO of AWS, said plainly that “nuclear is a safe source of carbon-free energy that can help power our operations.” 🔬
Google made history in October 2024 with what the industry called the first corporate SMR purchase agreement — a 500 MW deal with Kairos Power for a fleet of molten-salt reactors, with the first unit targeted for 2030. Kairos has already received construction permits from the U.S. Nuclear Regulatory Commission for two demonstration facilities in Oak Ridge, Tennessee.
Switch, the data center colocation developer, signed a non-binding master agreement with Oklo for 12 gigawatts of advanced nuclear power through 2044 — described by Power Magazine as one of the largest corporate power agreements in history. And X-energy, Amazon’s SMR partner, recently completed a record-breaking $1.02 billion nuclear IPO on Nasdaq, with the offering 15 times oversubscribed. The company failed to close a $1 billion SPAC merger in 2023. In two years, the entire investment thesis shifted.
What’s worth noting is that these aren’t just press releases. They represent genuine financial commitments with 20-to-30-year terms. That kind of contract forces both parties to treat nuclear not as a backup plan, but as core infrastructure. 📈
What makes SMRs actually fit for this job
Beyond the philosophical appeal of always-on, carbon-free power, there’s a specific technical compatibility between SMR architecture and data center needs that’s worth unpacking.
Traditional large nuclear plants are engineered for bulk baseload power delivery into the grid. They’re designed to run at full capacity continuously, with output distributed across regions. Data centers need something different: dedicated, on-site generation that bypasses grid congestion entirely, scales incrementally as the campus grows, and doesn’t compete with surrounding communities for electricity.
SMRs are sized for exactly this. The BWRX-300 produces 300 MW. NuScale’s VOYGR-6 configuration delivers 462 MW. X-energy’s Xe-100 produces 80 MW per module. These outputs map directly onto the power requirements of large hyperscale campuses — and crucially, a single operator can deploy multiple modules over time, expanding capacity as racks fill up rather than overbuilding on day one. 💡
The colocation benefit is real too. Last Energy, a startup building commercial SMRs in Europe, argues that on-site generation means:
No dependence on transmission and distribution systems that were built decades ago for very different loads
No competing with local residential and commercial ratepayers for grid capacity
Potential for waste heat recovery to power on-site cooling systems (a major data center operating cost)
Fixed-price long-term contracts, typically 20–30 years, that protect against electricity price volatility
The refueling interval is also quietly impressive: SMRs typically refuel every three to seven years, compared to the annual refueling shutdowns conventional reactors require. For a data center operator modeling 10 years of operating costs, that matters.
The honest case against betting on SMRs now
I’d be doing readers a disservice if I only presented the bull case. The honest version is messier.
The only commercial SMRs operating today are in Russia and China. Canada’s first commercial SMR project, the BWRX-300 at Ontario Power Generation’s Darlington site, is under construction and not expected to be operational until around 2030. In the United States, no commercial SMR has ever generated a single watt of electricity for a paying customer. ⚠️
Licensing is genuinely slow. The Nuclear Regulatory Commission’s Atomic Safety and Licensing Board recently allowed formal opposition to an X-energy SMR project in Texas on financial-qualification grounds, per Data Center Knowledge’s March 2026 coverage. Lux Research estimated that first-of-a-kind SMRs could cost nearly three times more than natural gas — $331/MWh versus $124/MWh — when factoring in cost overruns and delays. Even optimistic projections don’t see broad SMR deployment before the early 2030s.
Meanwhile, the data centers that Google and Amazon are building right now need power right now. The solution for the immediate gap, not the 2035 solution, is largely natural gas. The IEA itself notes that data center developers are “advancing a large number of projects with onsite natural gas-based power generation” as a bridge while longer-term solutions mature. That’s a carbon problem that 45 GW of SMR offtake agreements doesn’t solve today.
There’s also the question of whether many of those agreements survive contact with reality:
Some are conditional and non-binding frameworks, not signed PPAs
Regulatory approval timelines can and do slip
Community opposition, seismic studies, water rights, and site-specific engineering can each add years
SMRs require roughly 15 million gallons of water daily per site for cooling — not a trivial constraint
The data center industry seems to be pricing in a version of the future where SMR licensing gets significantly faster and costs compress through manufacturing scale. That may happen. It may not. Does your risk model handle both scenarios?
What this means for where nuclear goes next
Here’s what I think the data center-nuclear nexus actually changes, beyond the obvious deal flow. 🚀
The IEA’s Key Questions on Energy and AI report makes a point that doesn’t get nearly enough attention: “the momentum behind AI could accelerate the commercialization of new energy technologies.” That’s a polite way of saying that hyperscaler demand might do for SMRs what smartphone demand did for lithium-ion batteries — fund the manufacturing scale-up that makes the technology cheap enough to deploy broadly.
NuScale’s modules. X-energy’s Xe-100. Kairos Power’s molten-salt design. GE Hitachi’s BWRX-300. These designs are all chasing the same vision: a factory-built reactor, shipped to site, assembled like an industrial product rather than hand-built like a cathedral. If Amazon and Google collectively commit to 5–10 GW of capacity, they generate the order volume that justifies factory investment. That factory investment drives down per-unit costs. Lower costs open the technology to customers who aren’t named Amazon.
The analogy to Wikipedia’s account of learning curve economics in manufacturing is direct: every doubling of cumulative production typically yields a fixed percentage cost reduction. Nuclear has never had the production volume to ride that curve. It might finally be getting it.
What none of this resolves is the gap between the paperwork and the power plant. 45 gigawatts of offtake agreements is not 45 gigawatts of generating capacity. It’s a bet that the industry figures out how to build these things faster than it has historically managed to do. The data center industry is wagering enormous sums on that bet being right.
Whether you think that’s a rational calculation or motivated reasoning probably depends on how long your investment horizon is — and how much you trust an industry where the first commercial deployment is still four years away. What’s your read: are the hyperscalers leading a genuine nuclear renaissance, or buying very expensive optionality on a technology that may arrive too late?