The Feasibility of Nuclear Power for AI Infrastructure and Data Centers

 A technology and economics assessment — September 2026

Copyright: Sanjay Basu

Sitting through the panels at Datacloud USA and Metro Connect Fall in Austin this week, from the Nuclear for AI Summit to the site-selection and power-procurement sessions, I came away convinced that the industry is still conflating two very different timelines, and the conference stage did little to separate them. The energy on the floor around nuclear was real and, I think, largely justified. Firm, carbon-free, co-locatable power is exactly the solution that hyperscale AI wants, and behind-the-meter generation is the one credible path around interconnection queues that now stretch past five years. But too many of the discussions treated announced SMR gigawatts as if they were bankable capacity rather than options on a 2030s bet whose economics remain unproven. Every unit built anywhere in the world so far has cost more and arrived later than promised, and NuScale should still be haunting these rooms. The honest read, which a few of the utility and finance voices got right, is that the AI buildout landing in 2027 and 2028 will be powered by the grid, renewables-plus-storage, gas, and a finite pool of nuclear restarts. Not by new reactors. Nuclear is a genuine part of the answer, but it is a second-wave answer, and I’d have liked to see more of the panels plan for the power mix we actually have rather than the one we’ve contracted to have a decade from now. (These are my own views, not those of my employer.)

Executive summary

Nuclear power has moved from a speculative footnote in data center strategy to a signed, capitalized reality. As of mid-2026, every major hyperscaler, Microsoft, Google, Amazon, Oracle, and Meta, has committed to nuclear power for AI infrastructure, with more than a dozen agreements totaling nearly 10 GW of announced capacity. Yet the feasibility question is not “will nuclear power AI?” It plainly will, at the margin. The sharper question is when, at what cost, and for which portion of the load, and here the answer is far more constrained than the deal-flow headlines suggest.

The core tension is a timing mismatch. AI-driven electricity demand is arriving in 2025–2028. The nuclear capacity being contracted to meet it, particularly small modular reactors (SMRs), largely arrives in the 2030s. That gap is the central feasibility problem, and it shapes everything downstream, which reactor technologies matter, how the economics pencil out, and what a solutions architect should actually recommend to a product or infrastructure organization today.

The demand picture

Why nuclear is even on the table

The scale of AI power demand is what forced nuclear back into serious consideration. The IEA projects global data center electricity consumption roughly doubling from about 485 TWh in 2025 to roughly 950 TWh by 2030, approaching the total power consumption of Japan, and about 3% of global electricity demand. AI-focused data centers grow much faster than the aggregate, roughly tripling over the same period.

The concentration matters more than the aggregate. The US and China account for roughly 80% of the growth. In the US, data centers are projected to drive nearly half of all electricity demand growth through 2030, and the country is on course to consume more electricity for data processing than for producing aluminum, steel, cement, and chemicals combined. Former Google CEO Eric Schmidt testified that data centers would need roughly 29 GW of additional power by 2027 and another 67 GW by 2030.

Three structural features make this demand qualitatively different from historical load growth, and they are what make firm, carbon-free, co-locatable generation attractive:

1. It is geographically concentrated. Demand clusters around a small number of hyperscale campuses rather than spreading across the grid. Next-generation campuses under construction are projected to draw up to 20 times the power of a conventional large data center (which already consumes as much as ~100,000 households).
2. It is firm, not flexible. Training and inference clusters want 24/7 high-capacity-factor power. Intermittent renewables alone, without large storage, don’t match the load shape.
3. It has collided with the grid. US interconnection queues have grown to roughly 2,600 GW with a median wait of about five years, and only around one in five queued projects ever reaches operation. Wholesale prices near hyperscale facilities have spiked sharply. PJM, the largest US grid operator, flagged in January 2026 that data center demand was jeopardizing grid reliability.

This is the opening nuclear exploits: firm, carbon-free, high-density power that can, in principle, be co-located and bypass the interconnection bottleneck.

What has actually been signed

The 2024–2026 deal wave falls into three technically distinct categories, and conflating them is the most common analytical error.

1. Restarts of existing reactors (near-term, real). Microsoft’s 20-year power purchase agreement with Constellation to restart Three Mile Island Unit 1 (rebranded the Crane Clean Energy Center) is the emblematic case: an 835 MW plant, roughly $1.6B refurbishment, targeting ~2027, with Microsoft taking 100% of output. This is the most feasible category because the asset already exists and the design is proven. Amazon’s arrangement around the Susquehanna plant is similar in spirit — using existing nuclear capacity to feed a co-located campus.

2. Uprates and PPAs from operating fleets (near-term, real). Meta’s deals include purchasing output from existing Vistra and Constellation plants. This is effectively contracting for already-generated firm power and is feasible today, though it reallocates existing clean generation rather than adding new supply.

3. New-build SMRs and advanced reactors (medium-to-long term, uncertain). This is where most of the headline gigawatts live and where feasibility is genuinely contested. Meta’s agreements with TerraPower (Natrium sodium fast reactor) and Oklo (Aurora) target the ~2032–2035 window. Google’s up-to-1,800 MW Elementl Power agreement and its 500 MW Kairos Power deal are advanced-reactor bets. Amazon’s ~$700M investment in X-energy covers up to twelve Xe-100 units. These commit real capital, but almost none of it delivers electrons before the 2030s.

The single most important distinction for feasibility. Categories 1 and 2 are engineering-and-contracts problems that are largely solved. Category 3 is a first-of-a-kind (FOAK) deployment problem that is not.

The SMR economics reality check

SMRs are the technology most often presented as purpose-built for data centers, factory-fabricated, standardized, and faster to build. The promise is credible in theory and unproven in practice.

The cost story. SMR overnight capital cost is currently estimated around $10,000/kW for early units, versus roughly $6,600/kW for conventional large nuclear (IEA figures). SMR proponents argue the shorter build (a targeted 3–5 years versus 6–10+ for large plants) reduces financing cost enough to compensate, because interest accrues over fewer years. On paper, with a 5% financing rate, that math can bring an SMR’s all-in cost close to or slightly below a conventional plant’s despite the higher per-kW capital cost. The NOAK (“Nth-of-a-kind”) target CAPEX is $4,000–7,000/kW — but that depends on a manufacturing learning curve that doesn’t exist yet.

The LCOE gap. Current SMR levelized cost of electricity estimates cluster around $80–150/MWh for FOAK units, with NOAK targets of $50–80/MWh. Set against alternatives available today — utility-scale solar at $30–50/MWh, onshore wind at $25–45/MWh, and combined-cycle gas at roughly $40–75/MWh — SMRs are currently two to three times the cost of the cheapest firm and non-firm options. The NOAK targets would close much of that gap, but they are targets, not observed costs.

The cautionary tale. NuScale’s Carbon Free Power Project in Idaho is the reference-class failure. Its LCOE target rose from $58/MWh to $89/MWh even with subsidies (about $119/MWh without), project cost escalated from roughly $5.3B to $9.2B, it couldn’t attract enough utility offtakers, and it was cancelled. This is the base rate the industry is fighting against. Vogtle Units 3 and 4 — proven AP1000 designs, not novel SMRs — came in around $35B and roughly seven years late. Bent Flyvbjerg’s megaproject research puts average nuclear cost overruns near 120%.

The honest reading: SMR economics can work at NOAK scale with a real order book driving learning rates, but every unit built so far, globally, has cost more and taken longer than promised. The IEA’s base case sees only 10–25 GWe of SMR capacity installed globally by 2035–1–3% of global nuclear capacity. That is evolution, not the overnight revolution the deal announcements imply.

The timing mismatch

The crux of feasibility

Layer the two timelines and the problem is stark:

• Demand: meaningfully outpaces committed supply by roughly 2027–2028.
• New nuclear supply: restarts arrive ~2027; SMRs and advanced reactors mostly arrive 2030–2035.

For any AI capacity being planned for 2025–2028, new-build nuclear is simply not a supply option. Renewables plus storage and natural gas have 1–2 year build timelines and are available now; new nuclear does not compete on that horizon. Nuclear’s role in the near term is therefore almost entirely restarts and existing-fleet PPAs. A finite pool. The genuinely additive nuclear capacity is a 2030s story that helps with the second wave of AI buildout, not the current one.

This is why the sober framing among grid analysts is that renewables and natural gas will “take the lead” in meeting near-term data center demand, with nuclear as a firm-power complement that scales later.

Copyright: Sanjay Basu

Additional feasibility constraints worth naming

• Fuel supply. Several advanced designs (including X-energy’s Xe-100 and TerraPower’s Natrium) depend on HALEU. High-assay low-enriched uranium, whose Western supply chain is a genuine bottleneck. The ~$80B Brookfield/Cameco acquisition of Westinghouse is partly a bet on controlling fuel and technology supply ahead of this constraint.
• Regulatory throughput. Each novel design needs its own NRC licensing review. A proliferation of distinct SMR designs fragments regulatory attention and slows the whole category; standardization is the friend of feasibility here.
• Behind-the-meter vs. grid-connected. Co-locating a reactor directly with a campus (behind the meter) is what lets nuclear bypass the interconnection queue, its single biggest structural advantage. But it also concentrates single-point-of-failure risk and raises novel regulatory questions about islanded operation and grid support obligations.
• Fusion is not a plan. Microsoft’s PPA with Helion and similar fusion agreements are real contracts but should be treated as options with a low probability of delivering material power this decade. They do not belong in a feasibility base case.

Assessment

Where nuclear is and isn’t feasible for AI

The BWRX-300 (a boiling-water design leveraging proven BWR technology, targeting first Western commercial operation around 2029 at Darlington) is worth watching as the most credible near-term new-build, precisely because it is the least novel.

Recommendations for an AI infrastructure/solutions architecture team

Framed for planning decisions rather than energy-market speculation:

1. Treat nuclear as a 2030s firm-power hedge, not a near-term supply solution. For anything siting before 2029, plan around grid PPAs, renewables-plus-storage, and gas. Use nuclear commitments to secure future firm capacity and green credentials, not current electrons.
2. Distinguish sharply between deal categories in any capacity model. Restart/PPA megawatts are bankable. Announced SMR gigawatts should be risk-discounted heavily for schedule slip and cost escalation. Model them as options, not committed supply.
3. Prioritize sites with nuclear adjacency or interconnection headroom. Proximity to existing plants (restart or uprate candidates) and to sites viable for behind-the-meter reactors is becoming a first-order site-selection criterion, and land values within ~25–50 miles of planned reactor sites are already repricing.
4. Build for power-source heterogeneity. The realistic 2030 data center energy mix is diverse. Renewables (IEA projects renewables reaching ~50% of data center supply by 2030), gas, grid, and a growing nuclear slice. Infrastructure and workload-scheduling designs that assume a single firm source are fragile. Designs that can shift flexible workloads (much batch training) toward cheap-and-clean windows extract more value from any mix.
5. Watch three leading indicators, not the announcements: (a) whether the first Western SMR (BWRX-300 class) hits its ~2029 cost and schedule; (b) HALEU supply-chain maturation; © NRC licensing throughput for repeat designs. These, not new PPA press releases, will tell you whether the 2030s nuclear thesis is actually landing.

Bottom line

Nuclear power for AI infrastructure is feasible, but feasibility is heavily front-loaded onto existing reactors and back-loaded onto a 2030s SMR bet whose economics remain unproven. The near-term contribution is real but capped by the finite supply of restartable and upratable plants. The transformative contribution , factory-built SMR fleets delivering firm, carbon-free power at $50–80/MWh, is technically plausible and commercially bet-upon, but it depends on the nuclear industry doing something it has not yet done at any scale. Build repeatedly, on time, and on budget. Until a Western FOAK SMR proves that learning curve is real, the responsible planning posture is to bank the restarts, option the SMRs, and design AI infrastructure to run on a heterogeneous power mix.

Key sources

• IEA, Energy and AI / Key Questions on Energy and AI (2025–2026) — demand projections, mix outlook
• Brookings, Global energy demands within the AI regulatory landscape (2026) — US demand concentration, Schmidt testimony
• CBS News, Bloomberg, Forbes, Energy Connects (Jan–Jul 2026) — hyperscaler nuclear deal reporting
• SMRIntel; energy-solutions.co (2026) — deal inventory, SMR LCOE/CAPEX benchmarking
• softwareseni.com; IEEFA (2025–2026) — SMR cost/timeline critique, NuScale case
• NIRS / Kim & Macfarlane (2026); ScienceDirect techno-economic analysis — SMR overnight cost and construction-time modeling
• GlobSec (2024) — SMR vs. conventional financing math
• Lawrence Berkeley Lab / Bloom Energy figures via underhyped.ai — interconnection queue, US capacity gap