- Demand for electricity to fuel data centers is driving a boom in nuclear power projects
- But there's concern about the influence of AI in these facilities and their safety more generally
- Gartner and the U.S. Nuclear Regulatory Commission said there's little reason to worry
Chernobyl. Fukushima. Three Mile Island. Should we be worried that an AI-induced boom in nuclear energy will lead to the next big meltdown? Or that AI itself – running within a nuclear facility – might do the same?
“No, that’s the simple answer,” Gartner VP Analyst Bob Johnson told Fierce. Here’s why.
AI apocalypse?
Artificial intelligence (AI) is infiltrating just about everything these days, so it seems logical to worry about the impact it could have on some of the world’s 440 nuclear power reactors.
But not only are regulators all over the issue, but Johnson said there’s little reason for nuclear facilities to leverage AI in areas that would impact the safety of operations in the first place.
“I would say that there’s no reason for a control system in a nuclear plant to use AI,” Johnson said. He explained that while AI is great for sifting through mountains of data to find patterns – something very useful for, say, semiconductor design – the operation of a reactor is relatively simple.
“The control problem of a nuclear power plant has been solved long before AI even was thought of, and you don’t need AI to run a power plant,” he said. “Nuclear power is needed to run the data centers that run AI but it doesn’t need AI to run the power plant.”
And, looking specifically at the U.S. (which is home to 94 of the aforementioned 440 commercial reactors), even if a nuclear plant operator did want to use AI in its facility, it would need to secure approval from regulators.
“If a U.S. nuclear power plant wants to deploy AI in safety-related applications, it must seek NRC approval or document how it determines the use would have no more than a minimal effect on events covered by the plant’s safety analysis,” U.S. Nuclear Regulatory Commission (NRC) Public Affairs Officer Scott Burnell told Fierce.
Asked if AI is in use at nuclear facilities today, Burnell told Fierce that while AI could theoretically be applied to boost safety, the NRC has yet to receive any applications for safety-related use cases.
So, what kinds of AI applications are nuclear operators looking at?
“We are aware of efforts to develop AI-assisted drafting of applications, AI-assisted searches of the agency’s document database, and AI-enhanced analysis of plant risks, but none of those directly affect facility safety,” Burnell said.
Turns out the NRC has been working hard since 2023 to prepare for the AI wave. To date, it has established an Artificial Intelligence Steering Committee, created a database to track new and completed nuclear-related AI projects, and has been training its workforce to ensure it has the knowledge base to review AI-related applications.
Burnell acknowledged that the NRC’s existing regulations were developed well before AI became available, but insisted the agency’s regulatory framework is flexible enough to account for the changes AI might bring.
“This framework ensures any modifications or changes, including those involving AI, do not adversely affect the safety and security of the facilities,” he said.
And the NRC continues to work to identify gaps in its framework and flesh out new AI-specific regulations.
Last month, it also published a paper alongside nuclear regulators from Canada and the U.K. which detailed principles for safe use of AI in nuclear settings. And the agency’s AI Project Plan offers an in-depth view of the steps it plans to take through 2027 on AI. Among other things, it expects to award a contract in the current quarter to conduct a survey to identify both the tools to evaluate AI safety and any regulatory gaps that need to be addressed.
Is it safe?
But even without AI in the mix, is nuclear energy safe? Johnson said the answer is a resounding yes.
In 2015, researchers from Switzerland, Denmark and the United Kingdom set out to determine “how often and with what severity” nuclear disasters occur. The long and the short of what the study found? “Nuclear power accidents are decreasing in frequency, but increasing in severity.”
Put another way, there has been a “significant reduction” in the number of nuclear incidents since the 1970s. But those which have happened – Chernobyl and Fukushima – account for the vast majority (84%) of the damage seen in the researchers’ dataset.
But Johnson said one has to think critically about whether those disasters were anomalies. And, he argued, they both were. In the case of Chernobyl, he said the plant was built using a design that was known to be unsafe and rejected by every other country in the world except Russia.
Meanwhile, Fukushima’s facility actually had two reactors – one that had been upgraded to withstand a tsunami and one that hadn’t. The fact that the upgraded reactor didn’t melt down shows that we already know how to prevent a similar incident from happening in the future, Johnson noted.
Indeed, in the wake of Fukushima, the NRC required commercial nuclear power plants to implement a range of upgrades focused on mitigating the impacts of natural disasters to maintain plant safety.
And looking even further back to the Three Mile Island incident, Burnell told Fierce the NRC boosted its safety standards following that event as well. How? It ramped the use of “resident inspectors,” aka staffers who live near the operating plants. It also began tracking plant responses to malfunctions, revised its risk analysis methods, and required plant operators to enhance safety-related systems and staff training.
New kid on the block
But what about the small modular reactors (SMRs) that cloud hyperscalers like Amazon and Google and others are eyeing? Well, Burnell said the NRC is still reviewing applications for those, but so far has approved permits for a number of SMRs that use molten salt, liquid sodium or gas-based cooling designs.
Among those approvals is a permit for a test version of Kairos’ system (which Google plans to use). The NRC is also “ready to discuss” nuclear plans from X-energy and Energy Northwest for an Amazon project and Dominion Energy for an SMR to be used by Amazon in Virginia.
“Any U.S. nuclear power plant, regardless of size, must meet the NRC’s requirements to operate safely, to safely shut down under adverse conditions, and to protect the public if an accident occurs,” Burnell said. “Small modular designs, given the reduced size of the reactor core and other factors, could more easily meet those agency requirements.”
Johnson noted there are a few things that make SMRs potentially even safer than existing facilities. The first is the use of designs which automatically default to a safe mode when there is no power. For example, old designs may have required power to lift a fuel rod out of the reactor core to shut down. Now, though, the reverse is true: power is required to lift the rod up to the core. Thus, when there is no power, the rod drops out of the core, halting the nuclear reaction by default.
The second is standardization. Because SMRs are smaller than legacy plants, they can all be manufactured in a facility using a standard set of processes before being shipped to their ultimate destination. Contrast that with larger power plants, each of which has to be built to a custom specification.
Dark matter
But what about the money factor? With hyperscalers now front and center on the nuclear issue, could they use their deep pockets to change nuclear safety standards in their favor?
Likely not. Burnell noted the NRC is an independent agency.
“And to be clear, the Google and Amazon announcements involve those companies purchasing the electricity generated by the plants. The utilities or reactor vendors would own and operate the facilities and be responsible for meeting NRC regulations without input from Amazon or Google,” he concluded.
Update 10/28/2024 11 am ET: This story was update to amend the third bullet point to specify sources rather than using the general term 'experts.'