Site compute underground in disused mines, caverns and bunkers

The proposal

Build inside existing subsurface voids: worked-out mines, rock caverns, and decommissioned bunkers. The overburden that made the space expensive to excavate is free once excavation is done, and it is an excellent acoustic and visual barrier.

Why it works

Rock holds a stable temperature year-round, removing the seasonal swing that sizes conventional cooling plant. Several sites pair this with naturally cold water: Lefdal Mine in Norway uses fjord water and hydroelectric power in a former olivine mine, and Iron Mountain's Boyers facility sits more than 200 feet down in limestone with a 35-acre underground lake. Pionen in Stockholm and SubTropolis in Kansas City are long-running examples of the bunker and cavern variants (Data Center Knowledge).

For the siting problem specifically: nobody objects to a data centre in a mine. Noise does not propagate to any residence, there is no visual impact, and the land above remains as it was.

Implementation path

Suited to regions with a legacy of hard-rock mining and stable geology, and to operators willing to accept an irregular floor plan. Site selection turns on three questions: whether the rock is competent enough to avoid extensive ground support, whether there is a cold water source or sufficient ventilation, and whether power and fibre can reach the portal economically.

Trade-offs and limitations

• Rock insulates, it does not sink heat. The surrounding mass buffers temperature but has poor thermal conductivity, so heat still has to be moved out actively. The advantage is a stable inlet condition, not free cooling.
• Geometry constrains density. Existing voids were shaped by ore bodies, not rack rows. Vertical expansion is blocked, and high-density, high-power layouts used in AI training halls are hard to fit.
• Operating complexity. Humidity control, radon in some geologies, ventilation for personnel, and harder logistics for equipment delivery and removal.
• Cost. Specialised construction and access raise capital cost relative to a greenfield shed, and the supply of suitable sites is fixed and geographically uneven.

Where it fits

Best suited to secure hosting, storage, and moderate-density compute in regions with the right geology, rather than gigawatt-class training campuses. It is the one relocation approach with a multi-year commercial operating record and no unresolved environmental question.