Pentagon prepares to train AI on classified data as new reactors rethink nuclear waste

The Pentagon is weighing a new phase in the military use of generative AI: companies may be allowed to train special versions of their models on classified data inside secure environments. At the same time, another technology track, new nuclear reactors, is moving from theory into practice. Along with the promise of cheap carbon-free energy, it brings new questions about how to handle waste.

Secret models for the military

According to MIT Technology Review, the US Department of Defense wants to build secure environments where AI developers can fine-tune military versions of their models on classified data. This is not simply about accessing a model inside a classified network. It means deeper customization for real military tasks, from intelligence analysis to operational planning support. The logic is clear: a model that learns from material unavailable on the open internet may better understand the context, terminology, and scenarios that military personnel deal with.

This setup differs markedly from what happens today. Large models are already used inside closed networks to answer questions and analyze documents, but the model weights themselves are usually not trained on classified datasets. Under the new configuration, the Pentagon wants to host copies of the models in data centers certified for classified information while retaining ownership of the data. In rare cases, AI company employees may receive security clearances if a system cannot be tuned or tested without them.

Where the risks lie

The main risk is that classified information could become part of the model itself rather than simply sitting alongside it in a database or search index. That makes it harder to audit the system, separate access levels, and verify what conclusions the model can draw for a user with a given clearance.

Control over usage rules is a separate issue. The Pentagon's deals with OpenAI and xAI made clear that the military wants not an experimental pilot but full-fledged infrastructure for combat and analytical tasks.

"We will not allow any company to dictate the terms of how we make operational decisions," the

Pentagon said.

The dispute around Anthropic showed that the conflict is not only about technology but also about the limits of what is acceptable. The company publicly objected to scenarios involving mass surveillance and fully autonomous weapons, while the Pentagon insists on the broadest possible access to models for lawful military purposes. That turns defense contracts with AI labs into a new kind of bargaining: the government wants powerful models without unnecessary restrictions, and developers are trying not to lose control over exactly how their models are used.

• More accurate answers on closed military data
• Stricter requirements for model isolation, logging, and auditing
• Separate military versions of neural networks tailored to the armed forces
• Growing political and ethical pressure on AI companies

Reactors change the rules

MIT Technology Review also highlights another infrastructure story: the new generation of nuclear reactors. Most operating nuclear plants still follow a familiar template: a large facility, water cooling, low-enriched uranium, and centralized storage of spent fuel.

New projects promise to break from that standard. Small modular reactors, microreactors, and designs with alternative coolants and fuel configurations are in development. They are meant to be built faster, run more flexibly, and fit better into a power grid shaped by the data-center era.

The problem is that a new reactor architecture also changes the waste profile. The core methods, such as cooling pools, dry cask storage, and long-term isolation, are not going away, but there is no universal approach for all the new designs. Different fuel and coolant types produce different materials, with different packaging and logistics requirements.

Small reactors are especially tricky: if they are deployed at many sites, storing waste at each one will be inconvenient and expensive. That is why some companies are already considering a centralized model: shipping microreactors and their spent materials back to a single hub, such as the manufacturing site or a dedicated storage facility.

Experts believe the industry will not have to rewrite the entire nuclear waste management system from scratch, but it will have to refine it. For now, many conclusions rest on models and calculations, and the real picture will only become clearer once commercial units come online.

What it means

Both stories point the same way: the key technologies of 2026 look less and less like universal consumer tools and more and more like strategic infrastructure. AI is moving into classified military networks, while the energy sector is searching for new reactors to meet rising demand from industry and AI data centers.

For the market, this means one thing: the winners will include not only the builders of models and reactors, but also those who know how to integrate them safely into real-world systems.