Deep Fission reactors 1.6 km down for AI data centres

Clean megawatts are in short supply, permitting moves slowly, and patience is wearing thin across technology corridors.

A striking proposal from the United States enters this high-pressure environment: place compact nuclear reactors more than a mile beneath the ground and connect them directly to new campuses. The proposition relies on geology, drilling technology and the demand for dependable electricity at a predictable price.

Why bury reactors 1.6 km down

Deep Fission, a US start-up, says small nuclear units could be installed in 76 cm boreholes drilled to approximately 1.6 km deep. Endeavour Energy, the company behind Edged data centres, has agreed a target of up to 2 GW for its AI-ready facilities. The partners present the approach as clean, dispatchable generation that avoids the land requirements, lengthy schedules and grid-integration challenges that can deter major above-ground projects.

Two promised advantages stand out: a smaller surface footprint and a stronger safety envelope delivered by the rock itself.

The two big advantages

The first benefit concerns land use and cost. Because a deep-well reactor is almost entirely underground, the above-ground infrastructure is limited to a relatively small pad, substation and supporting equipment. The companies say this could shorten construction programmes and cut costly civil-engineering requirements, including enormous containment structures. They are also aiming for a delivered price of €0.05 to €0.07 per kWh, an attractive proposition for operators facing rising electricity tariffs.

The second advantage is safety. At a depth of 1.6 km, the surrounding geology provides a passive barrier. It contains radiation, offers protection from external incidents and gives operators more time to respond should a problem occur. The design is intended to lower the potential for airborne releases while making physical interference more difficult.

Rock becomes a permanent shield. No giant dome. No skyline-changing tower.

How the deep-well reactor would work

The proposed system is similar to a downhole heat source with a sealed primary loop. A narrow shaft would be drilled, the reactor module lowered into place, and heat exchangers linked to surface equipment that runs turbines or high-efficiency generators. The borehole supplies shielding, while purpose-built casings control pressure, temperature and fluids. Remote supervision and replaceable modules are intended to make maintenance periods easier to manage.

The case becomes clearer when the electricity demand is considered. The International Energy Agency estimates that data centres consumed about 1.3% of global electricity in 2023, equivalent to roughly 260 to 360 TWh. AI training workloads operate for extended periods, inference demand is expanding, and nearby grids frequently lack sufficient capacity. Locating power generation alongside computing facilities has an obvious logic, while nuclear power offers the availability profile sought by hyperscalers.

Attribute	Surface SMR	Deep-well SMR
Surface land use	Dozens of acres with visible structures	Small pad and substation
Shielding	Engineered containment buildings	Geologic barrier plus casing
Siting politics	Intense community scrutiny	Lower visual impact, fewer neighbours
Cooling approach	Often needs large water systems	Closed-loop systems, careful groundwater isolation
Security posture	Perimeter-heavy, above ground	Hard to access, below grade
Maintenance	On-site crews, larger components	Modular service, constrained access

What it could mean for AI-scale data centres

Provided the technology passes licensing and financing milestones, Endeavour intends to supply Edged sites with up to 2 GW of nuclear capacity. That amount could support several campuses and offer a fixed price over decades. Colocation providers could design their services around assured electricity, rather than seeking substation upgrades or waiting for connection slots in overloaded regions.

Stable power at the fence line changes site selection and speed-to-market for new compute.

The market signal grows louder

Large technology companies have begun exploring contracts supported by nuclear generation. Google has entered a framework agreement to purchase electricity from a small modular reactor developer. Other cloud and chip businesses are investing in advanced nuclear start-ups or agreeing early offtake arrangements. The underlying argument is consistent: clean, local and reliable power is preferable to volatile wholesale pricing when GPU clusters cost billions and cannot operate without electricity.

Questions that regulators will ask

The concept is ambitious, but it must still address established nuclear issues alongside new questions involving drilling and geology.

• Licensing pathway: How will agencies classify deep-well units within existing reactor regulations?
• Seismic and subsurface risk: What would occur during major ground movement or fault displacement at depth?
• Groundwater protection: How will casings, liners and seals stop any contact with aquifers?
• Emergency planning: What form would an off-site plan take when the core is beneath rock?
• Decommissioning: Once its operating life ends, how would the module be recovered or entombed?
• Fuel and waste: Which fuel type would be used, and how would spent assemblies be managed?

Deep Fission says geology narrows the routes through which an accident could develop. That assertion will be tested against modelling, test results and independent assessment. The nuclear sector has experienced gaps in public trust, so rigorous measurement, transparent reporting and straightforward explanations will be as important as the engineering itself.

Costs, timelines, and real-world hurdles

A target cost of €0.05 to €0.07 per kWh is compelling, but it depends on repeatable drilling, standardised modules and reliable financing. Grid connections remain relevant for backfeed and excess generation, although campus microgrids could manage most day-to-day operations. Compared with a conventional plant, construction may proceed more quickly if permissions, supply chains and drilling teams are available at the right time.

However, substantial risks persist. Underground work can produce unexpected conditions. Maintaining casing integrity for decades requires cautious design, and deep maintenance depends on dependable remote equipment. Any interaction with groundwater could damage public acceptance. During planning hearings, clear communication about sampling, monitoring and protective barriers will matter greatly.

What this means for cities and states

Areas competing to attract AI factories are facing an electricity constraint. Solar and wind can provide low-cost power, but they do not deliver continuously. Batteries can assist for several hours rather than several days. Gas can meet peak demand but increases emissions. A compact nuclear module close to the load addresses the duty-cycle issue and avoids protracted disputes over long-distance transmission, which can delay schemes for years.

Put power under the parking lot, not 200 km away behind a contested transmission line.

Extra context that helps frame the bet

Small modular reactors span numerous designs and capacity ranges. Deep-well concepts sit at the micro end of that spectrum, where individual units can supply tens to hundreds of megawatts. This capacity is better suited to a data-centre cluster than an entire city. The configuration also supports phased development: increase computing capacity, install another module, then repeat.

Cooling arrangements deserve close consideration. A sealed primary loop can transfer heat to a secondary loop, which releases it through dry coolers, hybrid towers or water-based systems. Locations facing water stress are likely to favour air-cooled or hybrid systems. Developers may also recover low-grade heat for nearby buildings, greenhouses or absorption chillers, improving overall site efficiency.

One practical indicator of progress is the appearance of test wells, regulator pre-application submissions, and fuel and drilling supply agreements. Their emergence would shift the proposal from a pitch deck to a project plan. The data-centre industry relies on roadmaps; electricity now needs one as well.