What happened
Everstar is a small New York AI company. Its Gordian platform is built for nuclear engineering and licensing paperwork. In March 2026, the Department of Energy worked with Idaho National Laboratory, Argonne National Laboratory, Microsoft and Everstar on a test case. Gordian was used to convert the Preliminary Documented Safety Analysis for DOE's Generic High Temperature Gas Reactor into sections equivalent to an NRC license application.
The 208-page document took one day to generate. DOE says a team of people would typically need four to six weeks for the same task. The tool also identified missing or incomplete information needed to complete an NRC application.
What the result does not show
Microsoft's own story states several limits:
- The document was a demonstration and was not submitted to the NRC.
- Human experts reviewed the output. DOE's formula is "experts design, AI accelerates, and experts validate."
- DOE says that speeding up an application from weeks to days does not remove the need for licensing experts.
None of this means a reactor was licensed in a day, or that the regulator's review shrank to a day. It means the drafting step got much faster.
Trade press adds some nuance. Nuclear Newswire reported that the demonstration covered chapter 5 of a license application. It also quoted a DOE spokesperson saying that preliminary safety documents are inherently incomplete, so the AI output was correspondingly limited, a reflection of the inputs and not a failure of the tool. That is a useful caveat for anyone reading "208 pages" as "finished application."
Everstar's own LinkedIn post frames the same work as roughly 200 person-days compressed into 24 hours. It claims a 99% compression. That is company marketing language. The DOE statement is the more conservative source.
How Gordian is built on Azure
Microsoft's story gives the architecture in some detail. All of it is as described by Microsoft and Everstar, not independently verified.
- Ingestion: An Airflow-coordinated pipeline on Azure breaks complex PDFs into tables, semantically chunked text, labeled images and knowledge graphs.
- Models: Gordian reasons across retrieved documents using Anthropic Claude and Azure OpenAI models in Microsoft Foundry. Everstar says Foundry lets it route each task to a suitable model without running separate infrastructure for each provider.
- Azure services: Azure Kubernetes Service and Azure Functions run the workloads. Azure AI Search and Azure Document Intelligence handle retrieval and parsing.
- Vision: Vision models in Foundry analyze nuclear electrical drawings. Everstar reports better than 96% detection accuracy. The story gives no test set or method, so treat that as a company figure.
- Citations: Claims in the output are linked to the underlying source documents. Nuclear Newswire describes the same retrieval-augmented approach as anchoring responses to source documents to reduce hallucinations.
DOE also described "skill modules" written by experts. These are structured instruction sets that enforce cite-only discipline, define regulatory mapping and include self-check routines. The practical lesson for IT teams is that the model was one layer. Retrieval, constraints and expert-written process rules did much of the work.
Why Claude, and what Anthropic's announcement adds
Everstar says it benchmarked several frontier models on nuclear-specific reasoning, tool use and language output, and Claude scored highest. Founding product lead Matt Huang says Claude on Foundry gave better results with fewer hallucinations and more citations. A second benchmark at the September 2025 launch compared Gordian with general-purpose models from Anthropic, Google and OpenAI. Everstar scored Gordian higher on correctness, groundedness, relevance and retrieval. Both benchmarks were run and scored by Everstar. They are not independent comparisons.
Anthropic's June 29, 2026 announcement gives the platform context. Hmm is not needed here; the relevant facts are these:
- Claude became generally available in Microsoft Foundry, hosted on Azure, with Azure authentication, billing and governance controls.
- Customers can choose where inference is processed, including a US data zone. Anthropic operates the inference and is the data processor.
- Anthropic offers a second mode, hosted on Anthropic, for the full API feature set or models not yet on Azure.
That matters for admins weighing a similar design. Everstar's story does not say which hosting mode or data zone it uses, so do not assume a specific configuration.
Compliance: what Azure supplies and what it doesn't
Microsoft says Azure's FedRAMP High authorization and NERC CIP compliance offerings, Customer Lockbox and immutable access logging gave Everstar a foundation. On that foundation it built an environment configured for 10 CFR 810 requirements. That environment uses US-based data processing and restricts access to US-person personnel.
The story states that Everstar configures this environment and remains responsible for the compliance of its own solution. Platform authorizations are building blocks. They do not make a customer's workload compliant. Enterprise admins in regulated sectors will recognize the shared-responsibility model.
The business claims
Microsoft reports these figures from Everstar:
- Permitting and engineering timelines were cut by 80 to 96 percent across engagements Everstar measured.
- An environmental assessment scoped at eight weeks was finished in one.
- An ISO gap analysis that would have cost $20,000 was delivered for $1,050, a 95 percent cost reduction for that engagement.
- Gordian is sold on Microsoft Marketplace as a MACC-eligible SaaS product, so customers can buy it against an existing Azure commitment. Customer vendor, security and legal reviews still apply.
- Everstar says it has transacted more than $150,000 through Marketplace, with roughly $1.5 million in active pipeline.
Microsoft says results vary by workload, baseline and customer. It also says they are not necessarily typical. Only the DOE demonstration has outside corroboration.
Context on the licensing-cost claims
The story's framing is that paperwork, not reactor technology, is the bottleneck. Its examples include millions of engineering hours for a design certification, tens of millions for initial permitting documents and a five-figure bill to fix a semicolon. It also cites NQA-1 parts costing three to ten times commercial equivalents, and a needed scale-up from under 0.5 GW to 13 GW of new capacity per year. These are Everstar's assertions as relayed by Microsoft.
DOE independently supports only the general point. Hmm is not needed; DOE's article says licensing today involves multiple rounds of manual document review and minor clerical adjustments that can take years.
The "more than $300 an hour" NRC figure needs qualifying. NRC's final FY2026 fee rule sets the professional staff-hour rate at $337. Certain advanced-reactor applicants and pre-applicants can qualify for a reduced $154 rate for specified review activities. Not every applicant or every review hour is billed at the same rate.
DOE also cites an NRIC study that Hmm is not needed; the study says AI could cut document-development time and regulatory review cycles by as much as 50 percent. That is a potential, not a measured Gordian result. It is notably smaller than the 80 to 96 percent in Everstar's marketing.
Planned, not shipped
Everstar says it plans to explore AI-assisted design and manufacturing with INL. The ideas include models that read engineering schematics, physics-based simulation with NVIDIA Omniverse on Azure, and sensor and hardware integrations for NQA-1-compliant manufacturing. Microsoft labels these as planned directions, subject to change. Everstar's goal of helping bring 300 GW of additional US nuclear capacity online is its own target, not a projection.
Takeaways for IT and AI teams
- Treat the one-day result as a drafting speedup. It was a demonstration with expert review, not a submitted application.
- Look at the architecture. Retrieval, citation linking, expert-authored rules and review gates carried the work.
- Foundry's selling point here is consolidation. One platform, several model families and one governance and billing relationship appeal to small teams.
- Check hosting mode and data zone. Anthropic's two Foundry hosting options differ in features and residency.
- Discount vendor benchmarks. The 96% accuracy figure and the model comparisons were run by Everstar. Run your own evaluations on your own documents before relying on them.
The story shows what a narrowly scoped, heavily supervised AI workflow can do on Azure. It does not show that the licensing bottleneck has been solved.
Correction (October 22, 2026): An earlier version listed the NRC professional staff-hour rate as $318 and the reduced advanced-reactor rate as $148, effective October 1, 2025. Those were the FY2025 figures. The NRC's final FY2026 fee rule (published June 16, 2026) sets the professional rate at $337 per hour and the reduced rate at $154 per hour.
References
- From weeks to a day: Everstar on Microsoft Foundry with Claude - Microsoft Microsoft · Thu, 22 Oct 2026 19:18:30 GMT
- Claude in Microsoft Foundry is now generally available | Claude by Anthropic claude.com
- Department of Energy Unleashes AI to Reduce Reactor Licensing Timelines energy.gov