The U.S. Nuclear Plant Fleet Is Leaning Into AI

AI tools have moved into the U.S. nuclear power industry, with nearly the entire fleet of 94 American reactors offered the chance to integrate AI into operations this year and most taking it. California-based Atomic Canyon launched NIVA, the Nuclear Industry Virtual Assistant, in August, developed with nuclear-industry groups, pilot-tested at Constellation Energy plants and now available to the whole U.S. reactor fleet. Constellation declined to be interviewed. Separately, startup Nuclearn says its products work with more than 65 U.S. partners, including an integration with NuScale Power Corp., which is working toward advanced small nuclear reactors. Microsoft and Nvidia have collaborated on a project applying AI to what they describe as the entire life cycle of a nuclear plant, spanning site permitting, design, construction and continuous operations.

Nuclearn CFO and cofounder Jerrold Vincent attributes the fast adoption to two factors: technology companies backing AI also want a nuclear revival to power data centers, and AI suits the industry's hardest problems, such as intricate regulations and complex management and maintenance systems. The industry faces aging hardware and an aging workforce. "Every engineering decision has to be evaluated, independently verified, and checked," Vincent said, describing Nuclearn's products as helping plants search large data repositories to file federal paperwork faster or find how a past problem was solved.

The NIVA project began a year ago when Constellation chairman Joseph Dominguez challenged nonprofit industry groups to make regulatory, maintenance and mechanical knowledge as convenient as a large language model while keeping it secure, according to Rob Austin, the Electric Power Research Institute's leader for the project. When Atomic Canyon won the NIVA contract in 2025, it first downloaded 53 million pages of public Nuclear Regulatory Commission data to train the model, then partnered with Oak Ridge National Laboratory to build nuclear-specific models. Lauderdale described these as sentence-embedding models, a way of teaching AI to seek nuclear language, which he called almost its own vernacular.

In 2025, the Institute of Nuclear Power Operations, EPRI and the Nuclear Energy Institute moved to create unified datasets, an idea that led to NIVA. Secrecy remains a hurdle: Nuclearn offers only single-location solutions so plant data never leaves a site, and Austin said NIVA can only search databases that already required plant-level security clearances. Both Nuclearn and NIVA models report to humans and do not touch plant operations. An EPRI pilot is testing whether AI can flag concerns in weld-inspection data, with a human reviewing all recommendations. Lauderdale said he cannot see AI making plant-operation decisions in the short or long term, citing regulatory and risk concerns, and expects humans in the loop for a very long time.

Why it matters: Nuclear operators and the vendors serving them now have fleet-wide access to AI assistants aimed at documentation and data search rather than control-room decisions, with data confined to secure, site-level or clearance-restricted systems. Adoption is being driven by a shrinking specialized workforce and pressure to keep aging reactors online.