RTS Nuclear + AI Briefing, Issue 04 · September 2026
In September 2026, AI was assigned to critical-path nuclear work: a licensing file, a waste processing weir, and an interconnection queue. That is where AI begins to matter, and where its outputs begin to be audited.
Blykalla pointed Microsoft’s AI at a permitting file. Savannah River ran an AI supported analysis, widened a weir by a tenth of an inch, and pushed the Salt Waste Processing Facility past its design throughput. Google, NVIDIA and Emerald AI pulled twenty organizations into an alliance built to make data center load flexible enough to get connected.
Licensing: Blykalla takes Microsoft’s AI to the licensing file
On September 15, 2026, Blykalla, the Swedish developer of the lead cooled SEALER reactor, announced a collaboration with Microsoft to apply AI to permitting and licensing. The stated target is to cut the documentation process “from years to months.” That remains a company projection, not a demonstrated licensing outcome.
Blykalla is not first through this door. Idaho National Laboratory and Aalo Atomics have both taken Microsoft AI to reactor licensing work. What is new is a developer applying it to its own submittal rather than to a research exercise.
The RTS perspective
The bottleneck is real. For an advanced reactor the documentation package is not paperwork around the schedule, it is the schedule.
But a compression claim has to survive a defensibility test. A regulator does not accept a faster document. It accepts a document it can trace. The questions that matter: does every generated assertion map to a qualified source, is provenance captured at the moment of generation, and will the review record still hold up under audit two years from now?
Human in Command® is the control that makes this defensible, not a tagline. Teams that build traceability into the generation step will move faster than teams that generate first and reconstruct the trail afterward.
Source: NucNet
Operations: a tenth of an inch, and 106 percent of design
Engineers at Savannah River Mission Completion investigated a throughput problem at the Salt Waste Processing Facility with a technical analysis that incorporated AI. The answer was a weir, the circular component inside the centrifugal contactors that controls liquid separation. They widened its inner diameter by one tenth of an inch.
Throughput moved from the 21.6 gallon per minute design rate to 23 gallons per minute, 106 percent of the original design. Before the change, the facility processed 17 million gallons of waste feed in almost six years. After it, the facility moved more than 3 million gallons in five months.
The RTS perspective
This is the most useful nuclear AI story of the month, and it will get the least attention. No new plant. No new model. An existing facility, an analysis, one part, and a number that moved.
The value came from pointing AI at a bounded engineering question where ground truth was measurable and the engineers kept the decision. It did not operate the plant or touch a safety function. The regulatory surface stayed small, so the work shipped. When operators ask RTS where to start, this is the answer: the bounded analysis problem with a measurable outcome, not the enterprise assistant.
Source: American Nuclear Society
Grid and demand: the grid decides who gets built
On September 16, Emerald AI, Google and NVIDIA launched the AI Energy Management Alliance, joined by seventeen more organizations including Anthropic, National Grid, AES, Constellation, NRG and RWE. The goal is to make data center load flexible enough to interconnect faster. Google committed to cutting 1 gigawatt of demand when grid conditions require it.
The RTS perspective
This looks like a data center story. It is a nuclear demand story. How much firm power the AI build out needs, and when, depends on whether flexible load is real. “Verifiable and enforceable” is a measurement and verification problem, a discipline nuclear already runs on. Developers should stop selling megawatts and start selling an availability profile they can defend with data.
Source: Axios
From the NRC desk: Part 53 takeaways from Workshop #7
RTS Founder and Chief Nuclear Officer Sam Rieck presented “Nuclear AI as the Key to Part 53” at the NRC Artificial Intelligence Workshop #7 on September 15, 2026, in Rockville, Maryland. The research was developed with N. Prasad Kadambi, author of NUREG/BR-0303. Sam’s takeaways:
- Part 53 and agentic AI have arrived at a mutually enabling moment. 10 CFR Part 53 supplies a technology inclusive, risk informed, performance based structure. Agentic systems may make the decomposition and traceability work practical on a real project schedule.
- The anchor is §53.230. Limiting the release of radioactive material is the primary safety function, supported by control of reactivity, heat generation, heat removal and chemical interactions, carried down through a traceable objectives hierarchy.
- The agent builds the structure. The human owns the judgment. Every generated element still has to be attributable, reviewable and owned.
- There is a readiness gap. No complete objectives hierarchy based Part 53 application exists yet as a reference example.
- RTS recommended a topical report pathway, defining methodology, controls, traceability expectations and the human command model now.
- Trust was treated as a system property, not a model claim. Duke reported cutting an aging management review from roughly five months with ten engineers to about one month.
What this signals
- Licensing is the schedule. Three developers are now pointing AI at submittal documentation. The differentiator will be provenance, not speed.
- Bounded beats broad. A weir and a tenth of an inch beat an enterprise rollout, because the delta was measurable and the regulatory surface stayed small.
- Verification is becoming the product. “Verifiable and enforceable” on the grid side is the same requirement as an auditable trail on the licensing side.
- The regulator is not waiting. Seven AI workshops in, the NRC is building its own readiness.
None of this rewards moving fast for its own sake. It rewards moving fast in a way you can show your work on: nuclear AI that is audit ready from day one, with Human in Command®.
Talk to RTS
One conversation. No pitch deck. Schedule a roadmap session or browse RTS research.
Comments are closed