US Nuclear Power Fully Embraces AI: How 94 Reactors Use Large Models to Manage Documents

Nuclear power plant control rooms were once considered off-limits for AI. The nuclear industry is known for its caution and risk aversion, and the complexity of reactor operations is enough to give even highly trained managers a headache. It's hard to imagine such an industry proactively embracing a powerful yet error-prone new technology.

But reality has given the opposite answer. According to IEEE Spectrum, nearly all 94 nuclear reactors in the United States have been given the opportunity to integrate AI into their operational processes, and most have already adopted it. This penetration didn't start in the control room—it started with mountains of paperwork.

Three-Way Convergence: Who Is Bringing AI into Nuclear Plants

California-based company Atomic Canyon has launched NIVA (Nuclear Industry Virtual Assistant), a project developed in collaboration with nuclear industry groups. After completing a pilot at nuclear plants operated by Constellation Energy, it is now available to the entire US reactor fleet. Constellation declined to comment.

AI-for-nuclear startup Nuclearn claims that its products are already working with more than 65 partners in the United States, covering numerous traditional nuclear plants, and also including an integration with NuScale Power, which is advancing the development of advanced small modular reactors. Microsoft and Nvidia have also teamed up on projects to apply AI to optimize "the entire nuclear plant lifecycle, from site selection and licensing to design, construction, and ongoing operations."

Jerrold Vincent, CFO and co-founder of Nuclearn, believes there are dual reasons behind this rapid adoption. First, tech companies betting on AI's future are equally enthusiastic about a nuclear renaissance to power data centers. Second, AI happens to be well-suited to tackling some of the nuclear industry's most pressing challenges—convoluted regulatory statutes and complex management and maintenance systems are exactly the types of tasks AI excels at.

The Real Bottleneck in Nuclear Power: Not Enough People, Mountains of Documents

Vincent points out that demand from advanced reactor developers and new nuclear projects is surging, but the existing workforce has been stretched extremely thin. He relays a client's exact words: "I have job openings I can't fill, the regulatory work has to get done—how exactly am I supposed to accomplish this?"

The nuclear power industry's main output these days, besides electricity, is paperwork. Nuclear plants must generate massive volumes of documentation to prove compliance and record responses to any potential issue—from sidewalk cracks to reactor malfunctions. Vincent says every engineering decision must be evaluated, independently verified, and inspected. Nuclearn's AI products help nuclear plants more efficiently search large databases, enabling faster submission of documents to the federal government or finding out how a problem was solved in the past. He compares it to "ChatGPT for the nuclear industry" and considers it a good starting point.

Another hidden concern is the loss of institutional knowledge. Because US nuclear construction has been stalled for so long, fewer and fewer people are entering careers as nuclear scientists or technicians. Trey Lauderdale—Atomic Canyon's CEO, who previously drove AI adoption in the equally conservative healthcare industry—points out that what the industry has left is a highly skilled but rapidly shrinking workforce that must both keep aging reactors running and potentially start up new ones to meet surging electricity demand. His judgment is blunt: without expanding the capacity of this knowledge workforce, a nuclear renaissance is impossible.

How NIVA and Nuclearn Solve the Data Silo Problem

Rob Austin, who leads the EPRI (Electric Power Research Institute) NIVA project, explains that the project began a year ago when Constellation Chairman Joseph Dominguez challenged several nonprofit industry organizations that serve as repositories for compliance, maintenance, and mechanical history information: find a way to make industry knowledge as easily accessible as ChatGPT, Copilot, or Gemini, while keeping it secure.

The task fell to Atomic Canyon. After winning the NIVA contract in 2025, the first step was downloading 53 million pages of public data from the Nuclear Regulatory Commission to train the model on nuclear power-specific operations, terminology, and regulations. The company then partnered with Oak Ridge National Laboratory to build the first batch of nuclear domain-specific AI models. Lauderdale explains that technically these are sentence embedding models—in plain terms, teaching AI to understand nuclear language. "The nuclear industry almost has its own dialect."

"The goal is to make decades of operational experience, hundreds of thousands of operational experience reports, tens of thousands of technical documents, and vast amounts of information searchable by people." — Rob Austin, Electric Power Research Institute

Austin gives an example: a nuclear engineer might encounter a problem with a giant vertical pump for circulating cooling water. The relevant material is simply too vast—EPRI maintenance guides, tens of thousands of incident reports stored by the Nuclear Power Operations Institute. The AI's task is to synthesize these reports and generate a guide to the pump's main failure types, and critically, to show its reasoning process so engineers can trace back to the original documents it cited.

Confidentiality is the main obstacle to modernization. AI systems that help with document and data retrieval are not subject to the strictest rules, but most nuclear data itself is proprietary information or protected by federal law. Nuclearn only offers on-premise deployment solutions, ensuring that nuclear plant data never leaves the site. Austin says NIVA can only search databases that already require plant-level security clearance, and the same rules apply when using AI assistants to sift through massive data. Technology cannot eliminate human barriers, but it does provide new incentives for more open data sharing.

Austin emphasizes that the goal is to make information more accessible and usable while ensuring data security and controlled access to critical information—and that's exactly what NIVA does.

Sentence Embedding Model is the core of NIVA's technical approach. Simply put, it maps a piece of text (a sentence or a paragraph) into a high-dimensional numerical vector, where semantically similar texts are also close together in vector space. This way, the system doesn't need to match keywords word-for-word but retrieves by "meaning": an engineer inputs a description of a circulating cooling water pump failure, and the model can find historical reports that discuss the same type of failure even with completely different wording. This precisely solves the retrieval challenge posed by the nuclear industry's complex terminology and varied document styles. Compared to general-purpose large models, sentence embedding models have fewer parameters, can be deployed locally, have lower inference costs, and are better suited to handling domain language like "nuclear industry dialect," with fine-tuning enabling the model to master industry-specific expressions.

On-premise deployment means all software is installed in the nuclear plant's own server room or intranet environment, data never leaves the facility, and model inference is completed locally rather than uploading data to the cloud. This is one of the prerequisites for AI adoption in the nuclear industry: nuclear plant operational data, maintenance records, and compliance documents are mostly proprietary information, some of which is protected by US federal regulations—external transmission could cross legal and security red lines. Through on-premise deployment, AI assistants search this data within the plant, satisfying security review requirements while giving employees a ChatGPT-like convenient experience. The alternative is the public cloud API calling model, which is fast to deploy and low-maintenance but requires data to leave the site—essentially unfeasible for nuclear scenarios.

From Document Retrieval to On-Site Inspection? Not Yet

Currently, both Nuclearn's and NIVA's large language models only report to humans and don't touch plant operations. But an inevitable question arises: what if they could?

Both Vincent and Austin say this is still far off, but they can see a path. Austin says nuclear plants have countless welds, all currently inspected manually for quality and durability. An EPRI pilot project is testing whether AI can review the massive data generated by these inspections to see if it can flag areas worth attention. The project still keeps humans in the loop reviewing all AI recommendations, and no nuclear plant allows AI to take over inspections. But it marks a step toward AI directly participating in plant operations.

Lauderdale is more conservative: whether short-term or long-term, he doesn't see an environment where AI makes decisions around plant operations, both due to the nature of regulation and the nature of risk. You don't want machines making decisions that could affect operations, so the human-in-the-loop state will persist for a very, very long time.

Human-in-the-Loop (HITL) is an AI system design principle: the model can generate recommendations, flag anomalies, or make preliminary judgments, but the final decision must be reviewed and confirmed by a qualified human operator or engineer. In nuclear plant scenarios, this means AI output can only be "supplementary information" and cannot directly trigger equipment actions or change operating parameters. The value of this model is that it leverages AI's ability to quickly discover patterns in massive data while preserving human accountability for risk consequences. For regulators, HITL is also an acceptable transitional approach because it doesn't require re-writing safety standards for autonomous decision-making systems, thereby lowering the compliance barrier for AI adoption.

Industry Observation: AI and Nuclear Power's Mutual Reinforcement

The essence of this combination is bidirectional dependence. Tech companies need nuclear power to supply data centers, and nuclear power needs AI to fill workforce gaps, digest regulatory documents, and preserve institutional knowledge. The downstream community of Diablo Canyon, where Lauderdale is based, gives him a direct view of this symbiotic relationship: a nuclear renaissance can power energy-hungry data centers, but only if AI can help the industry thrive again.

It's worth noting that the entire industry's AI applications are strictly limited to the auxiliary level—document retrieval, compliance support, knowledge management, inspection recommendations—all preserving final human decision-making authority. This restraint is not due to insufficient technical capability but is an inevitable choice determined by the nuclear industry's risk characteristics. For other high-risk industries, this may offer a reference implementation paradigm: first apply AI to information-intensive but risk-controllable areas, then gradually move toward core operations.