The physics and fuel costs of nuclear are almost irrelevant to whether a nuclear plant is a good investment. What actually determines nuclear economics is overnight capital cost, how long construction takes — because interest accrues the entire time with zero revenue — and who eats the risk if either blows out. Nuclear baseload economics is a financing problem wearing an engineering costume.
Plant Vogtle is the clean case study. Georgia Power targeted $14 billion and 2017 completion for two AP1000 reactors. It landed at $36.8 billion and seven years late, making it the most expensive power project in U.S. history, with much of that cost passed to ratepayers through construction surcharges. The all-in actual levelized cost came out around $169/MWh, compared with a $79–200/MWh range for other dispatchable and renewable alternatives — nuclear was on the wrong end of that comparison.
But that is the first-of-a-kind (FOAK) penalty, not nuclear’s ceiling. Industry modeling for the next AP1000 builds — once a workforce and supply chain exist — projects overnight costs falling from $10,000/kWe to $7,800/kWe to $6,200/kWe across the second, third, and fourth two-unit plants, with construction time compressing from 7 to 5.5 years, and levelized cost potentially reaching $50/MWh with tax credits applied. This is the crux of baseload nuclear economics: it is terrible on plant #1 and potentially excellent on plant #5, but someone has to survive plant #1 to get there. In the U.S., that has traditionally been ratepayers — which is why Vogtle triggered a political backlash in Georgia and why most utilities have been unwilling to try again.
The Overnight Capital Cost Problem
The scale of the capital cost gap across nuclear technologies is the defining constraint:
The gap between the red bar and the blue bar is the entire SMR value proposition: factory fabrication and standardization are supposed to eliminate the site-specific megaproject risk that destroyed Vogtle. Whether they actually do is the open question.
The SMR Timeline: A Genuinely Mixed Picture
No Western SMR has entered commercial operation yet. NuScale — the only U.S. design with NRC certification — had its flagship U.S. project, the Carbon Free Power Project in Idaho, cancelled in 2023 after its target cost climbed 53% to $89/MWh and utilities lost confidence. That continued into 2026: NuScale’s Q1 revenue collapsed 96% year-over-year, the company posted a $664 million net loss for 2025, and its TVA/ENTRA1 6-gigawatt program — the largest planned SMR deployment in U.S. history — remains in early planning with no binding customer contracts.
But real construction is happening elsewhere. GE Hitachi’s BWRX-300 became the first SMR under construction in North America at Darlington, Ontario, after Canadian regulatory approval. X-energy completed a $1.02 billion IPO in April 2026 and got NRC approval for a HALEU fuel facility. Kairos Power’s Hermes demonstration reactor is under construction in Oak Ridge. Rolls-Royce SMR selected its UK site at Wylfa, with first concrete possible as early as 2027. And outside the West, China’s HTR-PM has been operating commercially since 2023, with the ACP100 targeting first grid operation in 2026.
The realistic read: first Western units operating — late 2020s to early 2030s, plausible. Cost-competitive fleet-scale economics (the NOAK cost curve that is the actual point of building “small” and “modular”) — that requires building several units in series, and realistically that is mid-2030s in the West. NuScale’s specific stumble does not kill the concept, but it is a live demonstration that the FOAK cost-overrun pattern which plagued Vogtle has not been solved by making reactors smaller — it has been relocated, not eliminated.
What This Means for AI Data Centers
Hyperscalers are running two tracks at once, and they are solving different problems.
Near-term bridge — existing, proven nuclear. Microsoft’s $16 billion, 20-year deal to restart Three Mile Island Unit 1 (now the Crane Clean Energy Center) targets power by 2027. Amazon expanded its offtake with Talen Energy to 1,920 MW from the already-operating Susquehanna plant. Google struck a deal to restart Iowa’s nuclear plant by 2029. None of this is new technology — it is using AI’s capital and appetite for firm power to reopen or extend plants that already have a construction track record, which sidesteps the FOAK risk entirely.
Long-horizon bet — SMR developers. Google’s “order book” arrangement with Kairos, Amazon’s $700 million equity stake plus offtake from X-energy, and Meta’s sprawling 6.6 GW commitment across TerraPower, Oklo, and Vistra are structurally different. These are prepayments for capacity that mostly does not exist yet. What they actually buy is offtake certainty — exactly the thing nuclear project finance has always lacked, and exactly what utilities asking ratepayers to underwrite a NuScale-style FOAK project can no longer easily get away with post-Vogtle. Hyperscaler balance sheets are absorbing that risk instead.
Two implications worth sitting with. First, this is why nuclear is showing up in AI conversations right now even though the physics have not changed: it is the interconnection queue, not the reactors, that is binding. The U.S. queue has swollen past 2,600 GW with roughly five-year average waits, pushing companies toward direct power purchase agreements and even off-grid colocation. Second, do not mistake deal announcements for delivered megawatts — the SMR side of this is a 2030s bet dressed up in 2026 press releases, and it carries NuScale-shaped execution risk. If a Kairos or Oklo project has its own cost-overrun moment, the loss lands on a tech company’s balance sheet rather than a ratepayer base, which is a healthier place for that risk to sit — but it does not make the risk go away.
The Project Finance Constraint
The nuclear story is about risk transfer. For decades, the FOAK penalty sat with ratepayers — until Vogtle proved that ratepayers would revolt. The hyperscaler nuclear deals are not an energy story. They are a project finance story: tech companies are using their balance sheets to absorb the first-of-a-kind risk that utilities can no longer pass on to consumers. The near-term bridge (existing nuclear restarts) works because it sidesteps the FOAK problem entirely. The long-horizon bet (SMRs) works only if factory fabrication actually eliminates site-specific megaproject risk — which NuScale has not yet proven.
The alpha is not in nuclear power itself. It is in the layers that enable nuclear power to scale: HALEU fuel supply, NRC-certified designs, and the offtake certainty that makes project finance viable. The constraint is not whether nuclear can work. It is whether it can work at a cost that does not require a balance sheet to absorb the first-of-a-kind penalty.