AI centers are increasing demand of electricity. Nuclear is low carbon and more reliable than wind and solar.
" The deal would help enable a revival of Unit 1 of the five-decades-old facility in Pennsylvania that was retired in 2019 due to economic reasons. Unit 2, which had the meltdown, will not be restarted.
Constellation plans to spend about $1.6 billion to revive the plant, which it expects to come online by 2028."
And Oracle plans to use 3 SMR’s for their new data center. Wowzers.
At the time it shut down, TMI Unit 1 was licensed to operate until 2034. As reported in the following link, Constellation Energy will apply to extend the license until 2054.
~ ~ ~ ~ ~
Meanwhile, Holtec is progressing with plans to restart the currently shuttered Palisades plant in Michigan. Palisades was shut down in 2022.
~ ~ ~ ~ ~ ~
Not to be outdone, NextEra Energy is evaluating restart of the Duane Arnold nuclear plant in Iowa, which was shut down in 2020 after a wind storm damaged a cooling tower.
~ ~ ~ ~ ~ ~
Finally, earlier this year, Amazon Web Services (AWS) purchased a data center located adjacent to the currently operating Susquehanna nuclear plant in Pennsylvania.
~ ~ ~ ~ ~ ~
It is interesting how things have changed in just a few years. If Constellation, Entergy (previous owner of Palisades), and NextEra had foreseen the increasing demand for computing power, they never would have shut down their plants in the first place.
_ Pete
So SMRs are not hot today. Restart of old 1970s reactors is the new hot nuclear topic. Which is the cheaper option and which can be accomplished the fastest?
Jaak
Because SMRs are new, getting them permitted is likely to take longer. More questions to answer, etc. Previously permitted old reactors should be less risky for regulators and faster. But of course once SMRs are running we hope that changes and they become faster and easier.
The Federal Energy Regulatory Commission (FERC) recently issued a ruling that could impact these sort of power purchase agreements between tech companies that operate large data centers and the power companies that supply the electricity.
FERC rejected a proposed agreement between Talen Energy, which owns the Susquehanna nuclear power plant in Pennsylvania, and Amazon Web Services, to increase the amount of power supplied to a data center. This has caused a sell-off in the stocks associated with some of these companies.
_ Pete
FERC rejected the Amazon/Talen proposal because it was “behind the meter”, i.e., they would avoid grid taxes. IIRC, the Microsoft deal is not behind the meter.
DB2
Another recently retired US nuclear power plant may restart in the next few years. The Duane Arnold Energy Center was Iowa’s only nuclear plant when it shut down in 2020. The owner now sees growing demand for power, and has filed a request with the NRC to change the licensing status. NextEra is also looking buying new natural gas generators from GE Vernova.
~ ~ ~ ~ ~ ~
Separately, Bill Gates’ TerraPower company has signed an agreement with a large data center developer to supply power from new nuclear power plants that TerraPower is developing.
From the link:
Tech companies are scrambling to determine where to get all the electricity they’ll need for energy-hungry AI data centers that are putting growing pressure on power grids. They’re increasingly turning to nuclear energy, including next-generation reactors that startups like TerraPower are developing.
“The energy sector is transforming at an unprecedented pace after decades of business as usual, and meaningful progress will require strategic collaboration across industries,” TerraPower President and CEO Chris Levesque said in a press release.
~ ~ ~ ~ ~ ~
A question:
With the recent news of China’s DeepSeek AI, I wonder if the projected increase in electricity demand will really be as big as some anticipate? Supposedly, DeepSeek operates on a smaller number of computer processors, so the power demand is less. On the other hand, beware of Jevons Paradox. Efficiency improvements in a technology often result in even more demand, and energy requirements go up even more.
_ Pete
The small reactors for the likes of Google are being rethought in light of Deepseek’s breakthrough.
But the IRRs in the planning departments would have sunk many of those deals anyway. Deals so loudly shouted would have been canceled in a hush tone.
WSJ article (don’t recall which one) says much less power to train but MORE power to run the inference. This sounds like more power overall, since one AI is in widespread deployment you train less (compared to when creating and testing) and run inference more via many users/subscribers.
Mike
I understand the tech companies are already constrained by lack of power for data centers in general, in addition from new demand from AI.
I should note that most other articles say (or imply) the opposite so we have to get someone hosting a server to actually verify this.
Mike
However, as has been observed elsewhere, lower power, cheaper, etc. makes use of the tool more attractive, leading to more use overall, resulting in more total power use.
Based on My engineering experience, I think the need for AI electricity demand is highly overblown. Remember the early 1970s most utilities invested in nuclear power because it was thought to be cheap, easy to build, and safe. Most of these utilities spent lots of money on engineering and construction on these nuclear projects, but dozens were abandoned. Nuclear power was too expensive, to difficult to build, and nuclear accidents happened.
Keep in mind in the '70s there was a long term history of increasing demand for electricity. And nuclear cost was competitive. But then came the oil embargo and conservation efforts. More energy efficiency. Demand flattened. So much planned growth in generating demand turned out not to be required. Not unlike the dot com internet fiber boom.
And in Europe…
Meanwhile, utility EDF has identified four sites on its own land that it will offer for data centers…
With strong interest already secured from its financial partners, Fluidstack’s Phase 1 of the project will be supported by an initial investment of EUR10 billion (USD10 billion) and is set to become operational in 2026. The facility’s Phase 1 will ultimately host about 500,000 next-generation AI chips.
“The new facility will leverage France’s abundant, carbon-free, and predominantly nuclear energy to provide up to 1 gigawatt of dedicated AI compute power, reinforcing the country’s leadership in AI infrastructure, energy security, and digital sovereignty,” Fluidstack said. “Designed for scalable expansion beyond 1 GW by 2028, this project positions France as a premier global AI hub, offering unparalleled compute capacity for next-generation AI models.”
DB2