The Department of Energy is usually discussed as a single AI customer. It is not one customer. It is two, operating under one Secretary and one budget, with missions that have little in common and governance regimes that look nothing alike. On one side is civilian science and energy: the Office of Science, the national laboratories, the federal power marketing administrations, and the grid. On the other is national security: the National Nuclear Security Administration (i.e., NNSA), which maintains the nuclear stockpile. Both are DOE. Both are investing heavily in AI in 2026. Almost nothing about how they should govern that AI is the same. This post sets out the two regimes, explains why they diverge, and describes what the divergence means for anyone building AI for, or selling AI into, the Department.
Date
10.06.2026
Author
Shailesh Patel
Two missions under one roof
DOE's internal split is old and deliberate. The Office of Science funds and operates open research across the national laboratory system, and DOE oversees grid modernization and the federal power marketing administrations that move wholesale electricity. NNSA runs a separate program: maintaining the nuclear weapons stockpile without nuclear testing. That second mission is the reason DOE operates the largest scientific computing program in the world, because certifying the stockpile by simulation requires it.
The two missions share a department, a Secretary, and in several cases a laboratory campus. They do not share a threat model, a data-sensitivity posture, or a view on how open an AI model should be. Treating them as one customer, with one compliance story, misreads the Department at the first step.
The civilian side is larger and more varied than its public image. DOE funds the Office of Science user facilities, operates the four federal power marketing administrations (i.e., Bonneville, Western, Southwestern, and Southeastern) that sell wholesale power across much of the country, and leads federal grid-modernization work. The national-security side is narrower and deeper, a smaller set of laboratories and plants whose output is measured in assurance rather than throughput.
The science and energy regime
In late 2025 DOE launched the Genesis Mission, a national initiative across all 17 laboratories to build open scientific foundation models trained on the Department's own experimental and simulation data. The program is funded at roughly $320 million, with Argonne National Laboratory leading a consortium to develop the models and industry partners including Google DeepMind, OpenAI, and AMD, whose Lux cluster at Oak Ridge expands near-term capacity. The compute behind the open-science mission includes Frontier at Oak Ridge and Aurora at Argonne, both among the most powerful systems in the world.
The choice to build open models is deliberate. Open weights support reproducibility, external scientific scrutiny, and broad access across the laboratory system, which are values the research mission is organized around. The governance frame on this side is critical infrastructure and research computing. Classification is not the organizing concern. For the grid specifically, operational technology security dominates. NIST Special Publication 800-82 Revision 3 (i.e., the Guide to OT Security) is the current reference, and its central point is that the boundary between information technology and operational technology is now the primary attack surface. IEC 62443 applies as well.
There is a gap. NIST SP 800-82 and IEC 62443 secure the operational technology. Neither says how to authorize an AI model that runs on, or makes decisions against, that operational technology. The posture for science and energy values openness, reproducibility, and reliability, and its security standards were written for control systems rather than for models whose behavior changes after deployment.
The grid makes the tension real. A model that forecasts load, flags anomalies, or helps optimize dispatch sits close to operational technology that controls physical equipment. Open and inspectable is the right default for the science, and it is also a larger attack surface for the operations. Scientific openness and operational security pull in opposite directions on the same model, and nothing in the current standard set resolves that pull. SP 800-82 tells an operator how to secure the control system. It does not say whether a forecasting model is allowed to act on that system, or what to do when the model's behavior drifts.
The national-security regime
NNSA's AI program inverts almost every one of those values. Its principal effort, Artificial Intelligence for Nuclear Deterrence (i.e., AI4ND), is a fiscal year 2026 line within Advanced Simulation and Computing that develops AI for stockpile stewardship under an explicit requirement that the systems be trusted and assured, meaning safe, reliable, and secure enough for high-consequence environments. The compute for national security regime includes El Capitan at Lawrence Livermore, which is an NNSA laboratory rather than an Office of Science one.
The posture here is assurance over openness. Provenance, classification controls, and demonstrated reliability take priority over reproducibility and broad access. An open-weight foundation model trained for science is not, by default, acceptable in this environment, and the community's caution is visible in the April 2026 Asilomar Principles for Governing AI Applications in Nuclear and Biological Security, which address autonomy and assurance directly. NNSA has issued a request for information on AI capabilities and signaled intent to tune commercial tools for national-security use rather than adopt them as delivered. The same underlying technology that the science mission wants open, the security mission requires bounded and assured.
DOE's October 2025 AI strategy sketches how the security mission intends to use AI, including AI-enhanced physical security at the laboratories, faster regulatory processing for nuclear materials, and generative assistance for nuclear engineering design. In each case the requirement is the same. The tool must be bounded and assured before it touches the mission. NNSA has signaled that it will tune commercial models for national-security use rather than adopt them as delivered, which is a different acquisition and authorization path than the open-model strategy on the science side.
Where the two regimes collide
The difficulty is not that the two regimes exist. It is that they exist in close physical and institutional proximity. Argonne and Livermore are both DOE laboratories. A foundation model built for open science and a model assured for nuclear security can be developed by overlapping technical communities, on hardware from the same suppliers, under the same departmental leadership. The line between them is a governance line rather than a geographic one, and governance lines are easier to cross by accident than fences are.
The shared supply base sharpens the point. El Capitan, Frontier, and Aurora are all DOE exascale systems near the top of the global rankings, and they are built from the same small set of processor and accelerator suppliers. Technical staff move between laboratories over a career. The distance between an open-science model and an assured-mission model is therefore organizational and procedural rather than technical, which is exactly the kind of distance that governance must hold on purpose.
The gap that matters most is internal, and it is not unique to either mission. It is the distance between research computing, which is permissive by design, and production AI, which has to be authorized and monitored. DOE runs the leading research-computing environments in the country. Turning a research model into a production capability crosses a boundary that the research-computing posture does not police. On the mission side that boundary is a national-security boundary. On the grid side it is a critical-infrastructure boundary. They are different boundaries, and a control set written for one does not transfer to the other.
What this means for building or selling AI to DOE
The practical consequence is a short list of disciplines for anyone working with the Department. The first is to identify which regime a workload belongs to before designing anything, because the requirements diverge immediately. The science and grid side expects open or inspectable models, reproducible results, and conformance with operational-technology security guidance. The mission side expects assurance, provenance, and classification controls, and treats model openness as a risk rather than a feature.
The second discipline follows from the first. A single offering presented as covering AI for the Department, or a single compliance posture described as meeting DOE requirements, is a category error. The two missions do not share acceptance criteria, and a proposal that assumes they do will be mis-scoped for both. A vendor or architect who scopes to one regime at a time, and says plainly which one, will be read as understanding the Department. One that treats DOE as a monolith will not.
A third discipline is to settle data sensitivity and classification before model selection rather than after. On the science side the data is often meant to be shared, and openness is an asset. On the security side the data defines the controls, and exposure is the risk. A model chosen before that question is answered will usually be wrong for one side or the other, and reworking the choice late is expensive on both.
The model authorization boundary argued earlier in this series applies on both sides, but the authorized envelope is different. On the science side the concern is the reliability and reproducibility of a model that may eventually touch the grid. On the mission side it is assurance in a high-consequence, classified setting where the cost of a wrong output is not measured in downtime. The same discipline, drawing a boundary around what a model is authorized to do, produces two different documents depending on which mission is asking.
An earlier post in this series argued that FedRAMP authorizes infrastructure while a model's behavior goes unauthorized. DOE shows the next layer of the same problem. The authorization that matters is not only model-level, it is mission-level, and the Department runs two missions that authorize differently. A later piece will go deeper on grid operational-technology AI specifically, where the critical-infrastructure regime meets equipment that was never designed to host a model whose behavior changes after deployment.
—
DOE is not one AI customer. It is two, and the first step in working with either is knowing which one you are standing in front of.
—
Shailesh Patel is CTO of Keystone International Ventures and a SAFe Program Consultant (SPC). He writes about the intersection of AI architecture, federal technology, and the delivery frameworks that connect them.