The American energy story took an unexpected turn somewhere between the cooling towers of central Pennsylvania and the pumpkin patches of Wisconsin. Microsoft continued to make purchases after paying $76 million for a 407-acre pumpkin farm in Mount Pleasant, which was valued at about $600,000. Now the pumpkins are gone. A two-square-mile data center campus, part of a five-site network called Stargate, is rising in their place.
If it is ever constructed, it could consume as much electricity as five nuclear reactors operating at full capacity. The quiet side of the AI narrative is that scale. The headlines go to the chatbots. The electricity bill doesn’t.
| Topic Snapshot | Details |
|---|---|
| Subject | Big Tech’s nuclear power acquisitions for AI data centers |
| Key Players | Microsoft, Amazon, Google, Meta, Apple |
| Trigger Event | Microsoft’s deal to restart Three Mile Island’s Unit 1 reactor (2024) |
| Estimated Power Demand | Data centers may use 8–12% of total US electricity by 2028–2030 |
| Flagship Project | Stargate — a five-site supercomputer network projected at over $100 billion |
| Power Required for Stargate | Roughly 5 gigawatts (equal to five average nuclear plants) |
| Typical Data Center Downtime Cost | Over $8 million per day |
| Nuclear Plant Lifespan | 80+ years with refueling cycles of 18–24 months |
| Public Perception Baggage | Three Mile Island (1979), Chernobyl, Fukushima |
| Main Challenges | Construction timelines, regulatory hurdles, public fear, proliferation risks |
For this reason, in 2024, Microsoft agreed to purchase all of the output from a restarted reactor at Three Mile Island—yes, that Three Mile Island—for the following twenty years. The name still makes most Americans over forty shudder. Even though the actual damage was less than that of Chernobyl or Fukushima, the 1979 partial meltdown is still the worst nuclear accident in US history. No one died. However, the release of The China Syndrome a few days prior to the accident heightened the actual fear. Nuclear power in America never truly recovered after Hollywood and reality collided.
Apparently, up until now. Amazon, Google, and Meta have all inked agreements of their own, some for current facilities and others for small modular reactors that aren’t yet commercially viable. The industry seems to be placing a wager on a technology that has yet to demonstrate that it can be developed on schedule and within budget. In that regard, history is not encouraging. Georgia’s most recent nuclear project, Vogtle, ran billions over budget and years behind schedule. However, the technical reasoning is fairly straightforward. Data centers are constantly awake. Reactors don’t either. Despite all of their benefits, solar and wind do sleep.
Data centers will use roughly 8% of US electricity by 2030, up from about 3% now, according to Goldman Sachs. The Department of Energy projects that by 2028, it will be closer to 12%. These are the kinds of figures that change utility plans, reorganize grids, and make a restarted reactor in Pennsylvania seem more necessary than nostalgic.
It’s difficult to ignore the irony. Environmental organizations battled for decades to shut down facilities like Three Mile Island. The very symbol of the future, artificial intelligence, is currently being trained by companies that are writing checks to keep them open. Next-generation reactor startups have received investments from some of those same tech billionaires. TerraPower is owned by Bill Gates. Oklo was supported by Sam Altman. It’s really unclear if any of these projects will produce a functional commercial reactor by the end of the decade.
There are dangers that no one enjoys discussing. Long-term waste storage, nuclear proliferation, and the straightforward query of what would happen if a small reactor were constructed next to a data center and something went wrong. These issues have been brought to light by the Bulletin of the Atomic Scientists, and they have good reason to do so.
However, it seems as though the decision has already been made in the boardrooms as you watch this play out. Wind farms won’t stop the AI race. Investors appear to think that the first person to secure dependable gigawatts will win the next ten years. Perhaps that’s correct. Perhaps it isn’t. In any case, the pumpkins will not return.
