(Credit: Westend61 / Westend61 via Getty Images)
President Trump today announced $92 billion in AI-related investments, one of which is 10 new nuclear reactors to be built by Westinghouse in partnership with Google.
Construction will begin by 2030, interim CEO Dan Sumner told Trump today at a conference at Carnegie Mellon University, CNBC reports. But given the company's history, that's easier said than done.
The US has built just three reactors in the past 30 years, according to the US Energy Information Administration (EIA). The first is a 2016 project in Tennessee. Two followed, both at the same site in Georgia—one in 2023 and one in 2024. (The latter is not shown on the graph below, which contains data up to 2023.)
Both Georgia reactors are the same Westinghouse model the company plans to build again for its next set of 10, called the AP100. But that project came in $18 billion over budget and seven years behind schedule. The federal government under President Obama issued an $8.33 billion loan to help guarantee its completion, but Westinghouse still filed for bankruptcy in 2017 in large part due to its troubles in Georgia, Reuters reports.
What's different this time? Westinghouse is throwing AI at the problem. It's partnering with Google Cloud AI streamline production, making it "an efficient, repeatable process."
"By partnering with Google Cloud...we can accelerate the deployment of new AP1000 units while implementing powerful AI technologies that will optimize the construction and operations of nuclear power plants," says Sumner. The CEO says the nuclear industry is "reinvigorated" after Trump issued four executive orders to promote the industry in May.
It remains to be seen if Google's AI can save Westinghouse's nuclear reactor business, but the company says it's already "successfully achieved a first-of-its-kind proof of concept" using Google Cloud tech. It's using three Google services: Vertex AI (a machine learning platform), Gemini (Google's flagship large language model), and BigQuery (a cloud-based data warehouse). Together, they are able to "autonomously generate and optimize AP100" construction plans, Westinghouse says.
But using AI to construct nuclear reactors might be a recipe for disaster. Hallucinations remain a stubborn issue with large language models, so Westingouse will need to double (triple?) check every output. Any issues with these projects could tank the industry, which has fallen out of favor after the disasters at Chernobyl (1986), Three Mile Island (1979), and Fukushima (2011). Could AI make things worse, or create a black box in which even Westinghouse isn't quite sure how the reactor was built?
The tech industry is willing to take the risk in its desperate quest for more energy to power its AI data centers. Microsoft is reviving the Three Mile Island plant to power its AI ambitions. Meta CEO Mark Zuckerberg this week announced plans to create several data centers to achieve "superintelligence." Nuclear power could be an attractive, environmentally friendly option—if it works.