Data Centers and Nuclear: The Inevitable Marriage of the AI Era

September 20, 2024. Constellation Energy issues a six-page press release. The headline: Three Mile Island Unit 1, shut down in 2019 for economic reasons, will be restarted. The customer on the twenty-year PPA that makes it possible is Microsoft. The same plant that still sits next to Unit 2, where in 1979 the United States had its worst commercial nuclear accident.

The public narrative for forty years has been: nuclear power in the West is dead. Runaway costs, political opposition, Fukushima. Inside eight months of 2024, four hyperscalers wrote checks to change that narrative.

It is worth understanding why. Not because it is a fashion. Because the math of AI workloads does not leave many alternatives.

1. What actually happened in 2024

Four deals, in chronological order:

Amazon / Talen Energy — March 2024. AWS pays $650 million for the "Cumulus" data center campus in Pennsylvania, physically adjacent to the Susquehanna nuclear plant. The PPA allows up to 1.92 GW of dedicated capacity. In November 2024 FERC rejects the request to expand the interconnect — a meaningful regulatory blow — but the underlying deal holds.

Microsoft / Constellation — September 2024. Twenty-year PPA to restart TMI Unit 1, 837 MW, restart target 2028. The site is renamed "Crane Clean Energy Center". It is the first restart of a shuttered commercial reactor in U.S. history.

Google / Kairos Power — October 2024. Agreement for 6-7 Generation IV Small Modular Reactors (molten fluoride salt cooling). First unit online by 2030, ~500 MW cumulative by 2035.

Oracle — late 2024. Announcement of 3 SMRs for a single AI campus. Larry Ellison drops the word "gigawatt" on an earnings call with no further detail.

December 2024 adds Meta, with an open RFP for 1-4 GW of nuclear. OpenAI talks about the "Stargate" project with Microsoft, around 5 GW.

In under a year, the AI industry has committed — on paper — more than 10 GW of nuclear capacity. For scale: the entire Italian nuclear fleet, before the 1987 referendum that ended it, was 1.3 GW.

2. Why NOW: the math of AI load

Power is not a new problem for data centers. The shape of the load is.

A training cluster — say 100,000 H100 GPUs — draws roughly 150 MW continuously. Not peaks, not windows: 24 hours a day for weeks, months. A GPT-5-class training run is measured in GPU-months. Shutting the cluster down means throwing away state and restarting.

This is the consumption profile the grid calls baseload: constant, predictable, insensitive to time of day. It is exactly the profile that solar and wind, on their own, cannot supply.

Capacity factor is the fraction of time the source produces at nameplate. A modern reactor runs at 92% of nameplate for 18 consecutive months between refuels. A solar farm in Spain hits ~22%. To get 1 GW of reliable baseload from solar requires ~5 GW of panels plus tens of GWh of batteries. The economics, today, do not close.

There is also the corporate decarbonization problem. Microsoft, Google, Amazon all carry net-zero pledges by 2030. Natural gas is out. Coal is unspeakable. Nuclear is the only dispatchable source with zero operational emissions.

And finally speed. Building a new nuclear plant in the U.S. takes 12-15 years (see Vogtle 3-4). But restarting a shut reactor (TMI), or co-locating on an existing one (Susquehanna), takes 3-5 years. That is the difference between "in time for the next generation of models" and "too late."

3. The deal table

Note: Meta and Stargate numbers are stated targets, not signed contracts. The distance between announcement and delivered MW in nuclear is measured in years and lawsuits.

4. Europe, meanwhile

The asymmetry with the United States is stark.

France runs 56 operating reactors and plans 6 new EPR2 units by 2035. Yet no European hyperscaler (because no European hyperscaler at that scale exists) has signed a nuclear PPA dedicated to an AI data center. EDF sells nuclear power into the grid, not to a specific campus.

Germany completed its nuclear phase-out in April 2023, in the middle of the AI boom. It is a political choice with a measurable industrial price tag: German electricity costs 2-3x French electricity, and hyperscaler data centers in Germany (Frankfurt) run on gas and intermittent renewables.

The United Kingdom has announced an SMR program with Rolls-Royce, but with timelines past 2032. Poland, Czech Republic, the Netherlands are evaluating new-generation nuclear, but no hyperscaler deals.

Italy is in a peculiar position: the Meloni government publicly reopened the nuclear file in late 2023, but the regulatory framework for new reactors will not be in place before 2027-2028. Meanwhile, AI demand grows month over month.

5. What this means for Southern Italy and the Mediterranean

Italy will not have operating reactors in the next decade. But the question is not just "where does the reactor sit," it is "where does the electron sit."

The eastern Mediterranean has three relevant assets:

1. HVDC interconnections under expansion (Italy-Tunisia ELMED cable, Greece-Egypt, Sicily-mainland upgrade). A data center in Apulia can be physically far from generation and electrically close, provided the trunk holds.
2. Cheap solar — useful for inference workloads, which tolerate diurnal modulation better than training. The architectural separation between training (baseload) and inference (variable) is a conversation that is only now starting.
3. Geography for "nuclear adjacent" — France routinely exports 50-70 TWh per year of nuclear electricity. A dedicated France-Italy HVDC line (Piedmont-Savoie, already operating) could in principle deliver French nuclear megawatts to an Italian campus, with certified origin accounting.

It is not the solution the U.S. is building. It is the one European geography allows.

6. Power is the new chip war

For two years the AI conversation was dominated by Nvidia, TSMC, semiconductor export controls. That was the correct phase: without chips you cannot compute.

The 2025-2030 phase will be dominated by the electron. You can buy the most expensive GPU in the world, but if you cannot inject a megawatt into it for nine months, it is dead iron in a warehouse.

The numbers confirm it. The IEA, in its 2024 report, estimates that global data center electricity consumption will double by 2026, reaching ~1,000 TWh per year — comparable to the entire electricity consumption of Japan. Goldman Sachs projects +160% data center consumption by 2030.

Industrial history teaches that every technology platform inherits the physical constraints of the one before it. Steam needed coal nearby. Electricity let factories leave the rivers. Digital promised to decouple geography from production. AI is coupling it back.

Constellation, Talen, Kairos are not names that were taught in business schools in 2020. They will be in those of 2030. And the world map of data centers, five years from now, will look less and less like the map of network nodes and more and more like the map of reactors — active or restartable.

Whoever controls the electron controls the model. The rest is implementation detail.