AI Just Reopened American Nuclear. Inside the Eighteen-Month Shift.

For thirty years US utility nuclear was in retreat. New plants got cancelled, old plants got retired, and the orthodoxy in the industry was that we were done building reactors in this country. Then Microsoft signed a 20-year power purchase agreement to restart Three Mile Island Unit 1. Amazon bought a 480 MW direct feed from Talen Energy's Susquehanna plant with provisions to scale to 960 MW. Google signed with Kairos Power for small modular reactors. Oracle announced three SMRs of its own. In eighteen months, AI capital reopened American nuclear.

Walk through how it happened, why it works for AI specifically, and what the FERC fight looks like that could unravel the whole structure if it goes the wrong way.

The Eighteen-Month Shift

Chronologically, the major deals:

Nothing like this sequence had happened in three decades. The previous decade was the opposite: plants closing early because natural gas was cheaper and operators could not justify continued operation. Indian Point shut in 2021. Palisades shut in 2022. Diablo Canyon was scheduled to close in 2025 before California intervened. The industry consensus was that the existing fleet would retire on natural schedules and nothing new would replace it.

The Microsoft TMI announcement in September 2024 was the moment that consensus broke. Wall Street took it as a signal, and inside six months every other hyperscaler had announced something.

Why Nuclear Fits AI Workloads

Three structural reasons, in order of importance.

One, the load profile. AI training runs pull close to peak power 24 hours a day for weeks or months at a time. Inference is bursty in a different way than web traffic but still relatively flat compared to consumer demand. Nuclear is baseload-by-design. A typical plant runs at 90% plus capacity factor with refueling outages once every 18 to 24 months. The mismatch between nuclear's output curve and traditional grid demand (peaky in evenings, dipping overnight) is what made nuclear economics hard in deregulated markets. AI training is flat. The mismatch disappears.

Two, the contract length. Nuclear economics work when you can write 15 to 20 year power purchase agreements at predictable prices. They do not work when you have to compete in spot markets against natural gas. Hyperscalers want exactly the kind of long-dated price certainty that nuclear plants want to sell. The PPAs underwriting the new deals are 20-year, take-or-pay structures. That is enough to underwrite a restart. It is also enough to underwrite an SMR build, which is why the SMR vendors are suddenly real businesses again.

Three, the carbon math. Every hyperscaler has a public net-zero commitment by 2030 or 2040. Adding two gigawatts of AI training capacity makes those commitments harder unless the underlying power is firm and carbon-free. Solar plus wind plus battery can be carbon-free but is not yet reliably firm at gigawatt scale. Nuclear is the only proven dispatchable clean source for that duty cycle.

The FERC Fight That Could Unravel It

In November 2024 the Federal Energy Regulatory Commission blocked an amended interconnection service agreement that would have let Amazon scale its Susquehanna draw from 480 MW to 960 MW behind the meter. The technical objection: a campus that big drawing power from a generator inside the substation, but not paying full transmission cost-share, was a structural change to how grid costs get allocated. If hyperscalers can co-locate at nuclear plants and avoid transmission fees, the rest of the ratepayer base ends up subsidizing infrastructure that does not serve their load.

Talen and AWS filed a rehearing request. The matter remains unresolved. Three outcomes matter:

If FERC sides against the bypass structure, every direct-feed deal has to renegotiate. Microsoft TMI may or may not survive in current form because its structure is different (Constellation owns the plant and sells power into the grid, then bills Microsoft through a virtual PPA, not a direct feed). Amazon Susquehanna would have to either pay full transmission cost-share or restructure as a virtual PPA. The latter is uglier because it loses the latency and reliability benefits of co-location.

If FERC sides for the bypass structure, every existing nuclear plant becomes a potential AI campus site. The economics turn sharply favorable for operators who can lease land adjacent to existing reactors. Vistra Energy, Constellation, Dominion, and a handful of other nuclear operators with surplus land near plants effectively get a new revenue stream worth tens of billions over the next decade.

If FERC delays or punts, the deals stall. Some restructure as virtual PPAs. Some get cancelled. The bigger SMR commitments (Kairos, X-energy) are less affected because they are net-new builds that get built directly into the grid, not co-located bypasses. But the restart-existing-plant deals get harder.

The SMR Pipeline Behind the Restarts

The restart deals are short-term wins (existing reactor, NRC license renewal pathway, 2028 or 2029 operational target). The structural shift is the SMR pipeline behind them.

Kairos Power has DOE Hermes demonstration approval, targeting first criticality mid-2027, with Google as anchor customer for up to 500 MW of follow-on commercial deployment. X-energy is closer to commercial with Dow Chemical and Amazon committing. NuScale was the first SMR design certified by the NRC and remains a credible US option despite the UAMPS cancellation. Oklo, TerraPower (Bill Gates), Westinghouse AP300, and a handful of others are in earlier stages.

The pre-pandemic estimate for first commercial SMR criticality in the US was 2032 to 2035. AI capital has compressed that to 2027 to 2030 for the leading designs. Compressed how: by writing the offtake agreements that let SMR vendors get financed, and by lobbying the NRC to streamline license review for the licensing bottleneck.

What To Watch

The FERC ruling on the Amazon-Talen rehearing request. It is the load-bearing decision for whether the direct-feed structure survives. Expected sometime in 2025 to 2026.

NRC license decisions on TMI Unit 1 restart and Palisades restart. Both are mid-2026 to 2028 timelines. If either slips, the announcement-to-electrons window opens further.

Kairos Hermes first criticality in 2027. The first real SMR demonstration in the US in decades. If it goes well, the commercial SMR market is real. If it slips, the SMR pipeline gets longer.

New deals from hyperscalers not yet announced. Meta has not publicly signed a nuclear deal. Apple has not. xAI has not. All three have load profiles that need nuclear-shaped firm clean power, and all three have public clean-energy commitments. The next twelve months of announcements will tell us whether the eighteen-month wave was the front edge of a deeper structural shift or the peak of a short cycle.

Bet right now is that it is the front edge. AI demand is not going down, and the grid is not building transmission fast enough to serve it from anywhere else. The steel is going to follow the megawatts. The megawatts are increasingly going to come from uranium.

We track the active nuclear deals on the AI infrastructure page: Microsoft Three Mile Island, Amazon Susquehanna. Companion analysis of the broader buildout at The AI Buildout, Plain English.