The race to deploy data centers for Artificial Intelligence is hitting a more physical than algorithmic limit: available energy. In this context, a proposal is gaining momentum in Washington and the tech industry due to its unusual approach: repurposing decommissioned nuclear reactors from U.S. Navy ships and submarines to directly power large-scale computing campuses.
The initiative is driven by HGP Intelligent Energy, a U.S. company that has presented the idea to the Department of Energy (DOE) with a clear goal: to set up a data center in Oak Ridge, Tennessee, alongside the scientific and industrial ecosystem historically centered around its national laboratory. The plan involves using two naval reactors capable of providing continuous power of approximately 450 MW and 520 MW, respectively.
Why is this idea on the table now?
For years, the energy debate surrounding data centers revolved around expanding the power grid, renewable PPA contracts, and, at best, using gas as a “cushion” to ensure supply. The explosion of AI (training, persistent inference, and 24/7 availability needs) has brought back a word many thought was outdated:
Meanwhile, major tech companies have been exploring nuclear alternatives, especially around the concept of SMRs (Small Modular Reactors). The key difference here is that HGP is not proposing to build a new reactor but to reutilize existing nuclear hardware that was designed to operate for decades under extreme conditions.
What exactly is HGP proposing?
According to information circulating in international media, HGP has presented the DOE with a plan that includes:
- Reusing pressurized water naval reactors (PWR), the type that has powered U.S. military vessels for decades.
- Connecting these reactors to a data center campus with stable demand, rather than selling electricity on the wholesale market.
- Seeking financial support: the company aims for a loan guarantee from the DOE to kick off the project, with a total estimated investment of $1.8 to $2.1 billion.
- Working on a timeline that, if realized, would place the project toward the late 2020s (with 2029 cited as a target).
If successful, this initiative would set a significant precedent: it would be the first conversion of a military reactor for civilian use with the explicit goal of powering digital infrastructure.
What does the proposal promise (and what can’t it promise)?
This idea is attractive because it addresses a real sector challenge. An AI data center not only needs megawatts; it requires constant megawatts, predictably, and without depending on the local grid’s margin, substation availability, or reasonable interconnection timelines.
However, it’s important to differentiate between the narrative and reality:
- What it could provide: base load power, partial independence from the grid, and potentially faster deployment than building a new nuclear plant from scratch—if regulations allowed.
- What it cannot automatically do: bypass regulations, nor eliminate the main project bottlenecks: licensing, fuel, safety, and political approval.
The real challenge isn’t technical: it’s regulatory (and fuel-related)
This is where the plan becomes complicated. Naval reactors were not designed for civilian regulatory frameworks. Sources analyzing the initiative point out clear tensions:
- Licensing and regulation: the NRC’s (nuclear regulatory authority) framework is intended for commercial plants, not for “sealed” naval reactors with military logic.
- Type of fuel: many naval reactors use highly enriched uranium, a sensitive point due to security and proliferation concerns, and poorly aligned with civilian standards.
In other words, although “recycling” might seem faster than building from scratch, the actual speed depends on whether the U.S. is willing to create a specific regulatory pathway or adapt the existing one without delaying the project through bureaucratic processes and litigation.
Quick comparison: how does a “recycled naval reactor” stack up against other options?
| Option for powering a large data center | What it offers | What holds it back |
|---|---|---|
| Reused naval reactor (HGP proposal) | High, stable base power; leverages existing assets | New regulatory framework, fuel and safety concerns; strong political debate |
| SMR (new small modular reactor) | Civil design from the ground up; better suited for long-term licensing | Timelines, cost, industrial uncertainty, supply chain issues |
| Renewables + batteries | Good ESG fit and cost competitiveness | Challenging to guarantee 24/7 without oversizing; grid dependency |
| Gas (turbines) | Relatively quick deployment; firm power | Emissions, fuel prices, regulatory pressure |
What could go very right (if it works)
If the project manages to navigate legal hurdles, it could open a path many have pursued for years: “dedicated energy” for computing, independent of the public grid.
In a positive scenario, the message to the market would be compelling: if you can secure hundreds of megawatts of constant power, you can plan GPU deployments, cooling infrastructure, phased growth, and long-term contracts without relying solely on local substations and permits.
And what could go very wrong (if it fails)
If the movement is perceived as a regulatory shortcut or a security risk, the reputational and political costs could be immense. Additionally, there’s the classic business risk: spending years and capital on engineering and permits only to face an administrative “no,” or to see the project delayed so much that the temporal advantage is lost.
Therefore, the key isn’t whether the idea is “ingenious,” but whether the U.S. wants (and is capable) of transforming it into an industrial and regulatory policy without unleashing chaos.
Frequently Asked Questions
Can a private owner use a military nuclear reactor in a civilian project?
In practice, not directly and automatically. That’s precisely why the proposal focuses on permissions, licenses, and a regulatory fit that today isn’t designed for this kind of conversion.
How much energy are we talking about for an AI data center?
In HGP’s plan, it involves two reactors with an approximate capacity of 450 MW and 520 MW, comparable to regional electrical infrastructure.
Would this be cheaper than building a new nuclear plant or SMR?
The estimated cost ranges from $1.8 to $2.1 billion. It’s argued that this could be lower than constructing new nuclear options in some scenarios, but the regulatory framework and timelines are usually decisive factors.
Why not just rely on renewables and batteries?
Because many AI data centers require constant 24/7 power with high predictability. While renewables plus storage can work, they often involve oversizing and depend heavily on grid support for full reliability.
Sources: Bloomberg and Artificial Intelligence News