The short answer
Data center site selection is the search for land that can be connected to enough power, soon enough, with the cooling, fiber, zoning and community support to build and run a campus. For AI-scale facilities, power decides first: a parcel near a substation with spare capacity is worth far more than a cheaper one that would wait much longer for a connection.
The practical method works backward from power. Find where grid capacity exists or is coming, find the ground that can reach it, then test that ground for land, water, fiber, hazards and local acceptance before the market reprices it.
Why power is now the binding constraint
Three changes made electricity the first question.
- Loads are far larger. Servers built for AI training and inference pack much more power into each rack than conventional servers, and campuses are now planned in hundreds of megawatts, some at a gigawatt or more.
- The grid was not built for them. Many transmission lines and substations were sized for slow, steady growth in demand. New generation and transmission take a long time to plan, permit and build, while data center demand arrives in large blocks.
- Everyone wants the same headroom. Campuses, factories and growing towns compete for the same spare capacity, and a utility must study each large request before promising anything.
Land with power available is scarce, and land without it, however cheap, is often not a data center site at all.
Interconnection and substation capacity
Being near a power line is not the same as being able to use it. What matters is how much capacity can actually be delivered to the site, how reliably, and when.
- Substation headroom. The capacity left at nearby substations after existing customers and earlier requests. A fully committed large substation is worth less than a smaller one with room.
- Transmission voltage. Large campuses usually connect at transmission voltage, often through a substation built for the site.
- Upgrades and who pays. If serving the load needs new lines, transformers or a substation, the utility's study will say so, and the developer often pays for some or all of it.
- Redundancy. Many operators want two independent supplies.
- Firm or flexible service. Some utilities connect loads faster if they agree to cut consumption when the grid is stressed.
- Commitments. A growing number of utilities ask large new loads to sign long contracts, pay minimum charges or post collateral.
How long it takes to connect
Connection time is often the deciding factor, and it varies more than any other. Where a substation has spare capacity, a load can sometimes be served after a study and modest local work. Where the utility must build a substation, reinforce transmission or add generation, the wait is much longer, and the date can move as requests ahead in line are studied, built or withdrawn. Many campuses are therefore energized in phases, and some bridge the gap with on-site generation or batteries, or locate next to an existing power plant.
Land and zoning
- Size and shape. Buildable acreage for the full campus, a substation, cooling equipment and later phases, net of wetlands and easements.
- Zoning. Whether a data center is allowed by right or needs a special permit or rezoning, with the risk of refusal.
- Height, noise and generators. Local limits and air permits for backup generators shape the design and can rule out a site.
- Price and timing. Land near capacity reprices once the market notices. Options secured early capture the difference.
Water and cooling
Almost every watt a data center draws ends up as heat, and the way that heat is removed decides how much water the site needs. Evaporative cooling uses water to reject heat efficiently and can consume large volumes on hot days. Air-cooled and closed-loop systems use far less water but more electricity. Liquid cooling brings coolant directly to the hottest AI components, but the heat still leaves the site through dry coolers or cooling towers.
For each site, ask whether a reliable water supply will be committed, whether wastewater can be discharged, what happens in a drought, and how many hours a year the climate allows cooling with outside air alone. Water rights should be settled by specialists, not assumed from a map.
Fiber, hazards and the community
Fiber and latency
A campus needs diverse fiber routes, ideally from more than one carrier on physically separate paths. Training large AI models is not sensitive to a few milliseconds of delay, so training campuses can go where the power is. Inference that serves users in real time, and conventional cloud workloads, benefit from sitting closer to population centers and network hubs.
Natural hazards
Check flood risk, including risk beyond the mapped floodplain; exposure to earthquakes, hurricanes, tornadoes and wildfire; and extreme heat, which raises the cooling load just when the grid is most stressed.
Incentives and community acceptance
Many states offer sales tax exemptions on data center equipment, and many localities negotiate property tax agreements tied to investment and jobs. Community acceptance has become a real constraint: residents raise concerns about noise, water use, diesel generators, new transmission lines and local power bills, and some places have paused or tightened approvals. Early engagement and a credible plan for noise and water matter as much as the zoning map.
Site selection criteria at a glance
| Criterion | What to check | Red flag |
|---|---|---|
| Power capacity | Substation headroom, planned upgrades, transmission voltage | A line next to the site with no capacity behind it |
| Time to connect | Utility study results, upgrades required, phasing | A date that depends on projects not yet approved |
| Land | Buildable acreage, easements, room to expand | Wetlands or easements that cut the usable area |
| Zoning and permits | By-right use, height and noise limits, generator permits | A rezoning facing organized opposition |
| Water and cooling | Committed supply, discharge, drought exposure, climate | A design relying on water no one has committed |
| Fiber | Diverse routes, carriers, distance to network hubs | One carrier on one path |
| Hazards | Flood, seismic, storm, wildfire and heat exposure | Critical equipment inside a flood zone |
| Incentives and community | Tax treatment, local politics, past decisions on similar projects | A moratorium or a pending ordinance |
Screening from region to parcel
- Define the load. Megawatts at first phase and full build, timeline, redundancy, cooling approach and latency needs.
- Screen regions. Rank them by where grid capacity exists today and where more is being added, the price of power, water stress and policy.
- Find the substations. Those with headroom, or with upgrades already planned.
- Find the ground that can reach them. Parcels within practical reach that pass the size, zoning and hazard tests.
- Read the land market. Clusters of transactions near the same capacity show where competitors are moving.
- Diligence the shortlist. Utility talks and load studies, then water, fiber, title, environmental conditions and local opinion.
- Secure before announcing. A site known to have power is no longer priced as ordinary land.
Common mistakes
- Buying land before confirming power. Proximity to a line is not capacity.
- Reading a rating as output. Next to a power plant, check what it produces now; a rating says only what it was built to produce.
- Treating the connection date as fixed. It moves with other requests in line, equipment supply and permits for new lines.
- Leaving water and the neighbors to the end. Either can sink a site that passed every other test.
- Forgetting what the campus does to the area. New power reorders nearby land, and the parcels around a site reprice too.
How FABRIC does it
FABRIC: find the power before the land is listed
FABRIC is built around the constraint that now decides data center site selection: capacity. It shows where grid capacity exists today, where more is being added, and which ground can reach it, before a site is ever listed.
A campus shows up in the land records long before it shows up in a press release, so FABRIC reads land transactions across a whole region, maps them and ranks them. It reads what a site is producing now rather than what it was rated to produce, and what is being built around it. It ranks who stands to gain as new power reorders the property nearby, and it ranks every parcel in a region by how close it is to becoming a site, so the shortlist exists before the first visit. How the ranking is built stays proprietary.
For a specific project, the Energy practice assesses where a large load can actually connect, what power is worth hour by hour at that point, the cooling constraint, and which parcels sit within reach of the right substation while still priced as ordinary land. Its cooling screen runs before engineering and legal diligence, and is clear about what it cannot establish, including interconnection queue position and water rights.
Questions
What is the most important factor in data center site selection?
For large and AI-scale facilities, power: how much grid capacity can be delivered to the site, and how soon. Land, water, fiber and zoning still have to work, but without a credible power connection none of them matter.
How do you check power availability for a data center site?
Identify the nearest substations and transmission lines, ask the utility what capacity is available and what upgrades a new load would need, and request a formal load study. A power line next to a parcel does not mean capacity is available.
How long does it take to connect a data center to the grid?
It varies widely. Where a substation has spare capacity, connection can follow a study and modest local work; where new substations, transmission or generation are needed, it takes far longer.
How much water does a data center use?
It depends mostly on the cooling design and the climate. Evaporative systems use the most water, while air-cooled and closed-loop systems use far less water but more electricity.
Where should AI data centers be located?
Training campuses can go wherever large amounts of power, land and cooling come together, because training is not sensitive to small network delays. Inference and user-facing workloads benefit from sitting closer to population centers.
