A data center site is only as good as its worst constraint

Consider a hypothetical site.

It is a former industrial parcel on the edge of a European town. The land is affordable. A transmission substation stands nearby. A highway runs within a few kilometers. The local authority wants investment, and the owner has enough space for the first phase and several more after it.

On Monday, the site looks nearly perfect.

By Friday, it is dead.

The utility confirms that the nearby substation has no capacity for another large connection without reinforcement. The work might take eight years.

A second parcel has power sooner, but the two telecommunications providers serving it use the same bridge and last-mile duct. One construction accident could cut both routes.

A third has power and fiber, but its flood assessment relies on historical conditions that no longer describe the expected risk over the project’s life.

A fourth clears the technical tests and fails in planning. The buildings do not fit local height limits, cooling equipment would sit too close to housing, and the expansion plan asks the municipality to accept far more than the first phase suggests.

None of these failures is unusual.

Data-center site selection is often presented as a search for the best location. In practice, it is a process of elimination. A site can perform well across nine criteria and still be unusable because the tenth is fatal.

Cheap land cannot compensate for power that arrives too late. Low-cost electricity cannot compensate for inadequate fiber. A cool climate cannot compensate for flood exposure. Strong infrastructure cannot compensate for a project that cannot obtain permission to build.

The best site is rarely the one with the highest average score.

It is the one without an unresolved constraint capable of stopping the project or undermining its operation later.

The substation can lie

Power is usually the first filter because almost every other part of a data-center project can be designed only after the electrical envelope is understood.

A substation on a map does not establish that power is available.

The relevant question is whether the network operator can deliver the required capacity, under acceptable connection terms, on a timeline the project can finance.

The International Energy Agency reports that European data-center projects can face connection waits ranging from two to ten years, with some established hubs facing queues near the upper end.

A parcel may sit beside infrastructure whose spare capacity has already been allocated. The network may require a new transformer, substation, cable route, or upstream reinforcement before another large load can connect.

Even a connection offer needs careful reading. Is the capacity firm? Does it cover the first phase or eventual buildout? Which party funds reinforcement? What milestones preserve the place in the queue? What happens if the project fills more slowly than forecast?

The IEA estimates that unresolved power-system constraints could delay around 20% of planned data-center projects. It recommends locating more facilities where generation and grid capacity are already available rather than continuing to concentrate demand in the same constrained hubs.

That does not make every uncongested region a good location. It makes deliverable power a stronger criterion than proximity to power infrastructure.

The distinction between connection capacity, IT load, peak demand, and annual consumption also matters. A proposed 100 MW facility can mean several different things, while available electricity and available grid capacity are not the same.

A site waiting eight years for power is not necessarily a bad site. It may suit a long-term project with patient capital.

It is a bad site for a project whose commercial case assumes operation in three.

Fiber follows roads, bridges, and ownership

A data center can have enough electricity to run and still be unable to perform its useful function. Information has to enter and leave.

A map showing fiber nearby is only the beginning. The project needs to know which carriers are available, how much capacity they can deliver, how long construction will take, what latency the workload can tolerate, and whether the physical routes are genuinely independent.

Two provider logos do not necessarily mean two resilient connections. Both carriers may lease capacity on the same cable or follow the same duct, railway crossing, tunnel, bridge, or roadside corridor.

The UK government’s telecoms resilience guidance warns that services bought from separate suppliers may still share infrastructure. It recommends checking that geographic routes and points of presence are physically separate, rather than assuming contractual diversity provides physical resilience.

The workload determines how much this matters.

A backup archive may tolerate higher latency and a temporary loss of one route. A real-time inference service, financial platform, or distributed database may have stricter requirements around response time, throughput, synchronization, and failover.

Large AI and scientific workloads introduce another complication. The results may be compact while the datasets and model files needed to produce them are enormous. A site with attractive power can lose much of that advantage if moving data into and out of it is slow, expensive, or restricted.

A site-selection study should trace the physical network, not merely count carriers. Where do routes enter the property? Where do they converge outside it? Which bridges, tunnels, roads, and exchanges do they share? What capacity exists now? What upgrades depend on third parties? Can the workload move elsewhere if a route or facility fails?

The data center may be modular. Its network dependencies are spread across a landscape it does not control.

Cooling turns weather into infrastructure

The third promising parcel has power and fiber.

It sits in a hot, dry region where land is plentiful and renewable generation is growing. Climate appears in the commercial model as an annual average.

That average hides the week that matters.

A data center has to reject heat during the hottest operating conditions, not only typical ones. Cooling equipment, electrical capacity, water requirements, and hardware performance all depend on that design point.

ASHRAE’s current guidance for AI data centers treats power, cooling, and architecture as one interdependent system. A decision about processor density changes electrical distribution, cooling technology, pipework, structural requirements, leak controls, and the outdoor equipment needed to release heat.

Climate affects whether outside air or dry coolers can reject heat efficiently. High temperatures can increase fan, pump, compressor, or chiller demand. Humidity changes the usefulness of some evaporative methods. Dense GPU equipment may require liquid cooling closer to processors, while final heat still needs a route into the surrounding environment. The complete heat path belongs inside the site decision.

Water belongs there too. A site may have a secure electrical connection and an unsuitable long-term water position. Regional supply may look adequate while summer restrictions, agricultural demand, treatment capacity, or the proposed cooling system create a local problem. Water volume, source, consumption, and drought behavior need to be understood in the place where the facility will operate.

Heat recovery creates another apparent advantage that depends entirely on location. Captured heat has value only when a compatible user is close enough, requires the available temperature, and needs heat when the data center produces it. A line on a map labeled “potential heat user” does not prove a viable project.

The land carries its own history

Power and fiber can make an industrial parcel look ready for reuse. The ground may tell another story.

A former industrial site can contain contaminated soil, buried structures, undocumented utilities, unstable fill, or drainage designed for another use. A greenfield site may require roads, utilities, habitat mitigation, or flood defenses that erase the apparent advantage of inexpensive land.

The project needs room for buildings or modules, cooling and electrical equipment, access roads, security setbacks, drainage, construction staging, maintenance, replacement equipment, and safe movement around the operating facility.

Heavy equipment must also reach the site. Transformers, batteries, cooling systems, prefabricated modules, and servers may arrive on vehicles that require suitable bridges, turning radii, road strength, gates, and lifting areas. A parcel that looks accessible in a passenger car can become difficult once abnormal loads and cranes enter the plan.

Ground conditions affect foundations and cost. Flood risk affects electrical placement, drainage, access, and the ability to continue operating during an emergency. Wildfire, extreme heat, wind, seismic activity, storms, drought, and neighboring industrial hazards may change the design or insurance case.

Swiss Re’s 2026 assessment identifies power supply as a major source of business-interruption exposure and describes water stress, liquid-cooling leaks, fire, extreme weather, and connected operational systems as related risks.

Historical records remain useful, but the project is being built for future conditions. A flooded access road can isolate an otherwise dry facility. A water restriction can change cooling operation. Extreme heat can reduce thermal headroom while the grid is under pressure.

The site is not only the land inside the fence. It includes the routes and services the facility needs when conditions are at their worst.

Permission and operation are part of feasibility

A technically promising parcel can still fail because the project does not fit the rules or the place around it.

Zoning, building height, noise limits, setbacks, visual impact, water permissions, environmental review, grid agreements, construction traffic, and land-use policy vary by jurisdiction and site. They should not appear as one row labeled “permitting.” Each can change the project design.

The order matters. A developer that secures land before understanding the planning position may acquire a parcel for a project it cannot build. A community told about the proposal only after major decisions have been made may reasonably conclude that consultation is ceremonial.

Modularity can make the initial commitment smaller and more legible. It does not remove the need to describe eventual scale and cumulative impact.

Noise is particularly easy to underestimate. A value measured one meter from a machine does not reveal what reaches a neighboring property after source count, tone, barriers, weather, distance, load, and nighttime background are considered. Noise has to be assessed at the listener.

Construction has its own sequence. Foundations, utilities, access, logistics, inspections, commissioning, and testing remain real even when equipment arrives prefabricated. The visible installation is only one part of the build.

A parcel can pass engineering and planning review and still be difficult to operate. Data centers need maintenance, spare parts, monitoring, security, incident response, specialist contractors, and access to people who understand the electrical, mechanical, network, and IT systems.

A remote location may offer power and inexpensive land while increasing response times and limiting the technical workforce. An urban location may have strong network access and labor availability while creating tighter constraints around land, noise, traffic, and expansion.

A strong design does not operate itself. The site has to support the people, supply chains, procedures, and responsibilities required to keep it real.

Expansion can invalidate the first answer

Site-selection reports often reward room to grow. Empty space does not guarantee expandable infrastructure.

The grid connection may support the first phase and require years of reinforcement before the second. Fiber routes may lack capacity or diversity at greater scale. Cooling may fit the current workload and fail to support a denser hardware generation.

The first phase may comply with planning limits while the eventual buildout changes height, noise, traffic, water demand, or the relationship with nearby homes.

A site should therefore be assessed twice: against what will be built now, and against the maximum credible development the first investment is intended to enable.

Modular infrastructure can make phasing more deliberate. Capacity can be added in physical increments, and each increment assessed against the conditions that exist at the time.

A modular unit does not create another grid connection, fiber route, access road, permit, or maintenance team. It divides capacity into clearer steps. The site still determines how many steps are possible.

Do not average away a veto

Site-selection teams often use weighted scorecards. Power receives one weight. Fiber another. Land, tax, climate, labor, water, permitting, and risk receive their own scores.

The method helps compare viable options. It becomes dangerous when a fatal constraint disappears inside the average.

A site with excellent land, climate, and community support does not become 85% viable when the utility cannot deliver power. A parcel with available electricity and no lawful route to build is not an attractive site with a permitting weakness. A location with two carrier contracts sharing one physical route does not have resilience because the spreadsheet counts two providers.

Some criteria are trade-offs. Others are gates.

The first stage should identify the gates: legal use, deliverable power, adequate connectivity, buildable land, acceptable physical risk, an operable cooling path, and a credible route through approval.

Only after those conditions survive does it make sense to compare cost, speed, efficiency, incentives, and expansion options.

Infrastructure is rarely defeated by the average of its conditions. It is defeated by the one condition nobody resolved.

The site is part of the product

Policloud develops and deploys physical, modular data-center infrastructure for defined sites.

Modularity can widen the range of locations worth considering. It cannot rescue a poor site. The unit still needs deliverable power, viable fiber, a cooling path, suitable ground, physical access, permission, security, and an operating structure.

The question is not whether a standardized unit can be placed on a parcel. It is whether the unit, workload, and place form an operable system.

A viable site emerges when several promises describe the same project. The utility’s connection date matches the development schedule. Fiber routes are available and physically diverse. Cooling works in the local climate without depending on a resource the site cannot reliably supply. The ground supports the infrastructure. The planning authority understands the full proposal. The operator can maintain it. The workload fits the power, network, location, and operating model.

There is no perfect data-center site. Every location carries constraints.

Good site selection identifies which ones can be engineered, which can be contracted, which can be mitigated, and which should end the project before more capital is committed.

The best parcel is not the one with the most attractive features.

It is the one whose worst constraint has been found, understood, and made acceptable.