Where Should a Data Center Live? Vertical, Underground, Underwater, or Cold-Climate?

Six Physical Models

The relevant comparison is not novelty versus convention. It is the advantage each physical model creates, the constraint it introduces, and the market condition in which that exchange makes sense.

When Does Vertical Actually Make Sense?

A high-rise data center is not simply a warehouse turned on its side. Power must move through high-capacity vertical risers, cooling must negotiate elevation and longer distribution paths, and structural systems must support concentrated loads floor after floor. Large-equipment replacement requires planned hoist routes, removable façade panels, freight elevators, roof lifting zones, or permanent rigging. Compartmentation, firefighter access, smoke management, emergency power, egress, hazardous systems, and equipment separation must also be solved vertically. The design question is not how tall the facility can become, but what land value, latency requirement, zoning condition, or infrastructure constraint makes stacking capacity financially preferable. In many constrained markets, two to six stories may capture much of the land benefit without all of the penalties of a true tower.

Choose height only after the land, latency, zoning, and infrastructure case pays for vertical complexity.

Could a Data Center Go Underground?

Yes. Lefdal Mine Data Centers occupies a former Norwegian mine with six underground levels and 75 chambers; the company reports up to 120,000 square meters of white space and approximately 200 megawatts of potential capacity, supported by cold fjord-water cooling. Green Mountain's SVG-Rennesøy facility similarly reuses a former North Atlantic Treaty Organization ammunition-storage complex inside a mountain. These projects expose a crucial distinction: adaptive reuse of favorable underground infrastructure is fundamentally different from excavating a new hyperscale campus. Purpose-built underground development must resolve geology, groundwater, waterproofing, drainage, firefighting access, smoke exhaust, emergency egress, equipment delivery, replacement logistics, shafts, cable routing, and construction sequencing. The equation may work for an existing mine, cavern, hardened facility, or unusually favorable geology; for an ordinary greenfield site, complexity will usually outweigh the land-saving benefit.

Treat the ground as part of the structural, waterproofing, mechanical, fire-safety, and operational system.

What About Underwater?

Microsoft's Project Natick showed that subsea computing is technically credible. Its Northern Isles prototype operated on the seabed near Scotland for approximately two years, and Microsoft reported roughly one-eighth the failure rate of its land-based control group. The sealed prototype was intended for long, unattended operation. Its strongest lesson is not that conventional campuses should be submerged; it is that removing people, stabilizing the internal environment, industrializing modules, and designing for lights-out operating cycles may improve reliability. Commercial barriers remain substantial: subsea power and fiber, vessel deployment, permitting, marine review, corrosion protection, retrieval, failure response, refresh cycles, and the inability to repair equipment conventionally. Subsea computing therefore remains more compelling as a specialized modular strategy than as a default replacement for terrestrial Artificial Intelligence (AI) campuses.

Is the Arctic the Best Place for a Data Center?

Cold weather is valuable, but the coldest location is not automatically the best location. Climate-appropriate airside, waterside, and refrigerant-based economization can reduce cooling energy, while modern liquid cooling changes the relationship among outdoor conditions, cooling energy, and water use. Yet a remote Arctic site with inexpensive land and excellent natural cooling may still lose to a warmer location with deliverable electrical capacity, diverse fiber, faster permitting, a strong construction workforce, lower logistics risk, and better customer access. Climate is one variable inside the site-selection system—not a substitute for it.

Cold air cannot compensate for weak power, fiber, permitting, or logistics.

The Real Optimization Problem

Evaluate the data center environment as one system: power availability, time to power, power economics, fiber, climate, cooling, water, natural hazards, land, permitting, logistics, scalability, and community compatibility. Cheap land cannot compensate for unavailable power. Abundant power cannot compensate for an unpermittable parcel. Excellent infrastructure cannot compensate for geometry that cannot accommodate buildings, electrical yards, cooling systems, stormwater, service circulation, buffers, and future phases. Early feasibility should compare several candidate sites rather than attempt to justify the first parcel presented.

Do not allow one favorable variable to create a false positive for the entire site.

DCFR Design Principle

The future of data center architecture will not be one universal building type. Land-rich regions will continue to favor low-rise modular campuses. Dense metropolitan markets will push infrastructure vertically. Existing mines and hardened environments may support effective underground facilities. Subsea systems may remain valuable for specialized applications. Cool climates will continue to offer thermal advantages, and high-density AI will change every typology through liquid cooling and new electrical architectures. The decisive question is which combination of physical form, location, power, cooling, connectivity, risk, cost, and scalability produces the most resilient capacity with the least development friction.

Select the form that strengthens the complete delivery system—not the form that optimizes one isolated metric.