Data center site selection criteria field guide

Site selection, and why 2026 flipped the question

Data center site selection is the structured evaluation of candidate sites against the things that decide whether you can actually build and operate: power, water, connectivity, land, climate, natural-hazard exposure, and the policy and permitting picture. A decade ago the contest was location and the price per acre. The question now is shorter and harder. Where can you get the power, and when.

That inversion is the whole story of 2026. Power availability and the interconnection timeline are the binding constraint on a large build, because the grid cannot deliver tens or hundreds of megawatts of new load on the schedule the project wants. A site with cheap land, good fiber, and a friendly assessor is worth nothing if the utility cannot energize it for seven years. A worse-located site with a firm power commitment on a real date beats it.

So the right site is not the cheapest one. It is the one with firm power on a schedule you can underwrite, a water supply or an air-cooling plan, fiber on diverse routes, buildable land with room to grow, manageable hazard exposure, and a jurisdiction that will permit it. This guide screens a site against each of those. How the power actually reaches the property, the substation and the interconnection study, is the subject of the grid and substation guide. How many kilowatts the building can then deliver per rack is the capacity-planning guide. This one sits upstream of both. It picks the dirt.

Why is power the binding constraint in 2026?

Power is the binding constraint because the load a modern campus wants, tens to hundreds of megawatts and in some cases a gigawatt, is more than most local grids can absorb without years of upstream work. The screening question that now comes before everything else is time-to-power: can the utility deliver the megawatts, and on what date. If that answer does not work, no other strength on the site matters.

The numbers behind the constraint are not subtle. Interconnection timelines across major US markets commonly run several years, and grid-related delays have been adding roughly two to six years to projects in constrained territories. The large-load queues themselves have exploded. ERCOT's large-load queue roughly quadrupled in a single year, and one Texas utility reported its large-load interconnection requests jumping from about 1 GW to 8 GW inside about a year. Nationally, well over 2,000 GW of generation and storage sits in interconnection queues waiting to connect.

The practical consequence is that developers screen for power first and build the rest of the decision around it. The most sought-after sites are the ones that can deliver power in the next 18 to 24 months. Treat every one of those figures as a snapshot that varies by ISO, utility, and the size of the load you are asking for. The direction does not vary: power leads the decision and the rest follows. See the grid and substation guide for how the interconnection study and the utility deal actually work.

What is an interconnection queue?

An interconnection queue is the ordered line of projects waiting for the utility or grid operator to study and approve their connection to the transmission system. Your project takes a number, the operator runs the studies that decide what upgrades the grid needs to carry your load, and you wait your turn and pay your share. The queue, not the construction schedule, is now the long pole on most large builds.

The studies are the slow part. A system impact study that determines what transmission and substation work your load triggers frequently takes 12 to 24 months or longer, and that is before any steel goes up. Projects that energized recently had often sat in the queue for years. Transmission reinforcement, when the study finds you need it, is its own multi-year effort, since major transmission lines run something like seven to ten years from planning to in-service.

This is the schedule risk that kills deals quietly. A site can sit next to a substation, show plenty of headroom on the utility's service map, and still fail, because the surrounding grid cannot absorb a new 100 MW or 300 MW load without network upgrades that take years. Confirm the queue position, the study status, and the realistic energization date in writing with the utility before you treat a site as powered. The queue rules and timelines vary by ISO and jurisdiction, so the utility's own answer controls, not a regional average.

Power cost and source

Once a site can get power on a workable schedule, the next questions are what that power costs and where it comes from. Energy is the largest operating expense over a data center's life, so a difference of a cent or two per kilowatt-hour compounds into real money across a 15- or 20-year hold. The energy rate, the demand charges, and the structure of the utility tariff for a large load all belong in the underwriting, not just the headline price per acre.

The source mix matters for cost and for the carbon commitments many operators carry. Access to low-carbon generation, the ability to sign a power purchase agreement, and proximity to existing or restartable generation have become real siting factors. Large operators have signed long-term PPAs tied to specific plants, including nuclear, to lock both a price and a clean attribute for a campus.

On-site generation has moved from backup to bridge. Where the grid cannot deliver on time, developers are building behind-the-meter generation, gas turbines or fuel cells, to carry the load until the utility interconnection catches up, then run as backup or supplement after. That is a way to beat the queue, not a way to ignore it. The grid and substation guide covers the behind-the-meter option in more detail. Hedge the rate and the source to the specific utility tariff and the project's energy strategy.

Do data centers need a lot of water?

Yes, if the site uses evaporative cooling, a large data center can consume a substantial volume of water, and water availability has become a real screening criterion. Evaporative and water-cooled systems cut the electrical cost of cooling, but they do it by evaporating water, and a large share of what they draw, often most of it, leaves as vapor and does not return to the local supply. In a water-stressed or drought-prone region that is both a permitting problem and a community one.

There is a direct tradeoff between water and power. Evaporative cooling lowers the energy penalty of cooling and improves PUE, but it spends water. Air cooling and closed-loop or refrigerant systems spend little or no water, but they raise the power draw and the PUE, which means more of the constrained resource you are already fighting for. You are choosing which scarce input to lean on, and the right answer depends on the site's climate and its water rights.

So the water question has two acceptable answers, and a site needs one of them. Either secure a water supply with the rights and discharge permits to use it, or commit to an air-cooled or low-water design and budget the extra power. Water usage effectiveness, WUE, is the metric here, stated in liters per kilowatt-hour, and zero-water and low-water designs are now in the field specifically to take this constraint off the table. Confirm the water rights, the cost, and the discharge permit with the local authority, because all three vary sharply by jurisdiction.

Connectivity and fiber

A data center has to move its traffic, so the site needs fiber, and it needs more than one path of it. The screening criteria are how many carriers reach the site, whether the routes are physically diverse, and how far the nearest long-haul or metro fiber actually is. A site with a single fiber path has a single point of failure that no amount of internal redundancy fixes.

Diverse routing is the part people underestimate. A well-built site brings fiber in through two separate entrances, on physically separated paths, so a backhoe through one conduit does not take the building dark. Carrier neutrality matters too: a site that can reach many carriers and the cloud on-ramps gives tenants choice, lower latency, and cost control, while a site captive to one provider does not.

Proximity to existing fiber drives cost and schedule. Building new long-haul fiber to a remote site is expensive and slow, so sites near existing routes and exchange points have a real edge. The relationship to power is what makes this hard, and it gets its own section below: the places with abundant power and the places with dense fiber are often not the same places.

Does latency matter for site selection?

Latency matters for some workloads and barely at all for others, so it shapes site selection only when the workload is latency-sensitive. Latency is the round-trip delay between the user, or another data center, and the site, and it scales with distance. For interactive services, financial trading, real-time gaming, and synchronous data replication between sites, a few milliseconds is the difference between working and not.

For those workloads the site has to sit close to the users or the exchange it serves, which is the whole logic of edge data centers: smaller facilities placed near population centers to hold the delay down. Synchronous replication between two sites is the strict case, since it commonly tolerates only single-digit milliseconds of round-trip latency, which caps how far apart a paired campus can be.

For large-scale AI training, latency to users is close to irrelevant. A training cluster grinds for days or weeks and ships a result, so it can sit wherever the power is cheapest and most available, far from any city. That is exactly why training campuses are landing in remote, power-rich locations while latency-sensitive workloads stay near the metros. Match the site's distance to the workload, and do not pay a metro premium for a workload that does not need it.

Land and buildable acreage

Land still matters, but it is now a later question, screened after power and fiber. What you are checking is whether the parcel has enough buildable area for the build plus expansion, whether the ground can carry the load, and whether anything on the site will slow you down. The usable acreage after setbacks, wetlands, easements, and drainage is the number that counts, not the gross acreage on the listing.

Size scales with the build. A mid-size campus in the tens of megawatts commonly wants something on the order of tens of acres, and a hyperscale campus often wants 50 acres or more, with extra land banked for the phases that come later. Think in acres per megawatt and per phase, not in a single lump, because the substation, the generator yard, the cooling plant, and the setbacks all eat land that never shows up as white space.

Then the dirt itself has to cooperate. Geotechnical conditions decide the foundation cost, so the soil bearing capacity, the water table, and the presence of rock or fill all feed the budget. Flood zones, wetlands, and contamination can shrink the buildable area or trigger mitigation that adds months. A cheap parcel with bad soils, a wetland through the middle, and a contamination history is not cheap once the geotech and environmental reports come back.

Will the jurisdiction permit it?

Whether the jurisdiction will permit the project, and how long that takes, is now one of the largest schedule risks after power, so it gets screened early. The questions are whether the zoning allows a data center by right or only by special exception, what the entitlement path looks like, and how the local permitting timeline and any community opposition are likely to play out.

Zoning is the first gate. Data centers usually land in industrial districts, light or heavy, and sometimes in technology or business parks with a special approval. A site zoned correctly for the use, by right, is far faster than one that needs a rezoning or a special exception, where the timeline can stretch from a couple of months to the better part of a year as the approval goes through public review. Some jurisdictions have added discretionary review specifically for data centers, which lengthens the path further.

The blunt reality of 2026 is that permitting and entitlement now drive a large share of project delays and cancellations, more than the construction itself. Moratoria have been proposed or adopted in a growing number of places, and projects have been blocked or reversed after community pushback even where the zoning initially allowed them. Confirm the zoning, the entitlement path, the noise and setback rules, and the political temperature with the local authority before you commit, because all of it varies by jurisdiction and it is moving fast.

Community acceptance and social license

A data center needs the community to accept it, and that acceptance is no longer automatic. The friction points are real and local: the noise from generators and cooling equipment, the water the site draws, the strain on the grid that residents see in their own electricity bills, and the sense that a large facility brings little lasting local employment after construction. A project can clear every technical hurdle and still die in a town meeting.

The trend has turned. Public opposition has blocked or delayed tens of billions of dollars of US data center development in recent years, and the policy mood at the state and local level has shifted from competing for these projects to questioning them. Tax-incentive rollbacks and moratoria are downstream of that change in sentiment.

The work that earns a social license is unglamorous and early. Engage the community before the application, be straight about the noise, water, and traffic, show the tax-base and infrastructure benefits in terms residents care about, and design to mitigate what they will object to. A site in a place that wants the project, with a track record of approving them, is worth more than a marginally better parcel in a place that has started saying no.

Climate and free cooling

Climate is a site criterion because a cooler, drier location lets the cooling system run on outside air for more of the year, which lowers both the power and the water the site spends on cooling. The mechanism is free cooling, or economization: when the outdoor air or the wet-bulb temperature is low enough, the plant can reject heat without running the chillers, and the energy penalty of cooling drops.

The number that governs this is the wet-bulb temperature, not the dry-bulb you read off a thermometer. Water-side economizers do their best work where the wet-bulb sits below roughly 78 degrees F for a good part of the year, and the more hours a site spends below that line, the more cooling it gets close to free. Cooler and drier climates therefore deliver a lower PUE for the same design, which is why so many large campuses have gone to the northern tier and the high desert.

This ties straight to the water section and to the cooling design, so it is not a standalone preference. A cold, dry site can run dry coolers and economizers and spend little water. A hot, humid site leans harder on either evaporative cooling, which spends water, or mechanical cooling, which spends power. Match the cooling design to the climate, and credit the climate in the energy model rather than treating it as a footnote.

Natural-hazard risk

Natural-hazard exposure is a screen you run to avoid building a critical facility where the ground, the weather, or the fire season will threaten it. The hazards that matter are seismic activity, flooding, wildfire, tornado, and hurricane, and each one either rules a site out or adds hardening cost that belongs in the comparison.

Some of this is avoidance, plain and simple. You do not put a data center in a mapped floodplain if you can help it, because the flood risk, the insurance, and the downtime exposure are not worth it, and where a site touches a flood zone you raise the critical equipment above the flood elevation. High seismic zones drive structural cost and equipment anchoring. Wildfire-prone areas bring smoke, which fouls air-side economizers and forces them shut during the exact hot, dry conditions when you most wanted free cooling, on top of the fire and power-shutoff risk itself.

The right move is to map the hazards early, against FEMA flood data, the published seismic hazard maps, and the regional wind and wildfire exposure, and to price the hardening for the hazards you cannot avoid. A site in a low-hazard region starts ahead. A site in a high-hazard one is not disqualified, but it carries a cost and a risk that has to show up in the scorecard rather than getting discovered after the deal closes.

Grid reliability

Beyond whether the utility can deliver the power, there is the question of how reliably it delivers it once you are connected. Grid reliability is the outage history and the quality of the local utility, and it feeds directly into how much on-site backup you have to build. A site on a weak feeder with a history of outages forces more generator and UPS investment than a site on a strong, redundant part of the grid.

What you look at is the utility's outage record for the area, whether the site can be served by two independent feeds from different substations, and how the utility has performed for other large customers nearby. Redundant feeds matter because a single feed is a single point of failure, and the cost and feasibility of a second feed is part of the site comparison, not an afterthought.

None of this removes the need for on-site backup, since every data center carries generators and UPS regardless. But a reliable grid lets that backup be backup, rarely called on, while an unreliable one makes it part of the daily operation, which costs fuel, maintenance, and runtime hours. Weigh the grid's reliability alongside its capacity, because the two are different questions and a site can pass one and fail the other.

Tax and incentives

Tax treatment can move the economics of a site, but it is a tiebreaker, not a reason to build somewhere you cannot get power. The common incentives are sales- and use-tax exemptions on the servers, equipment, and sometimes the electricity, plus property-tax abatements negotiated with the local jurisdiction. Across the US a large number of states offer some version, and they compete on the strength of the package.

The caution for 2026 is that this picture is tightening. A growing number of states are moving to cap, condition, or repeal data center incentives as the power demand pushes up residents' electricity bills and the politics turn. An incentive that exists today may be narrower or gone by the time a multi-year project energizes, so underwrite the deal on economics that survive without the abatement, and treat the incentive as upside.

A clean incentive package still helps, and the dollars are the smaller part of why. A deal that the local jurisdiction understands and supports, and that can be explained to both an investment committee and the community, is a signal that the entitlement path and the social license are real. Confirm the specific incentives, the claw-back conditions, and the duration with the jurisdiction, since the terms vary widely and they are changing.

Why sites with both power and fiber are scarce

The sites that have both abundant power and rich fiber at the same time are scarce, and that scarcity explains a lot about where data centers cluster. Power tends to be abundant where land is cheap and remote, near generation and open transmission. Fiber tends to be dense where people and existing data centers already are. The two conditions pull in opposite geographic directions, and the rare place that has both becomes very valuable, very fast.

That is the cluster effect. Once a region proves out both inputs, Northern Virginia being the obvious example, more operators pile in to be near the fiber, the talent, the supply chain, and the proven utility relationships. The cluster then strains its own power and water, which is what pushes the next wave of development out to new regions that have to build the fiber to match their power.

For site selection the lesson is concrete. Do not assume a power-rich remote site has the fiber, and do not assume a fiber-rich metro has the power headroom left. Confirm both independently, early, because a site that is strong on one and weak on the other is common, and the missing input is usually the expensive and slow one to add.

Expansion and the campus

A data center is rarely a single building, so the site has to have room to grow and the power has to be able to grow with it. The screening question is whether the parcel supports the full campus plan in phases, with land banked for later buildings, and whether the utility can deliver power in stages that match the build-out rather than all at once on day one.

Phased power is the part that ties back to the interconnection. Securing a large block of capacity for the full campus up front is ideal but often impossible given the queue, so developers frequently stage the power, taking what the grid can give for phase one while the upgrades for later phases work through the studies. The land plan and the power plan have to be designed together, because expansion land with no path to expansion power is just a field.

Buy or option the land for the whole vision, not the first building. The cost of a larger parcel up front is small against the cost of being landlocked when the campus succeeds and there is nowhere to put the next phase. The expansion plan is also where the capacity-planning discipline starts, since the watts-per-rack the later phases will carry are higher than the first phase, and the site has to be sized for that trajectory.

Workforce

A site needs two different workforces, and they are easy to forget in the rush to lock up power and land. Construction needs a regional pool of trades deep enough to build a large, technical project on schedule, and operations needs technicians who can run the facility for its life. A remote, power-rich site can be hard to staff on both counts.

The construction labor question is whether the local market can supply the electricians, mechanical trades, and controls specialists a data center build demands, or whether you are importing and housing crews, which adds cost and risk. The operations question is longer-lived: a site far from any population center may struggle to hire and keep the technicians who run it. Weigh the labor market alongside the physical inputs, because a site that is perfect on power and impossible to staff is still a problem.

Supply chain and long-lead gear

The site decision now collides with the equipment supply chain, because the gear that connects a campus to power has lead times measured in years. The long-lead items are the large transformers, the switchgear, and the generators, and their availability has become a development constraint in its own right, not a procurement detail handled later.

The numbers are stark. Large power transformer lead times have stretched from roughly two years before the boom to about four years or more for high-capacity units, driven by shortages of grain-oriented electrical steel and copper and by demand outrunning factory capacity. Switchgear and large generators commonly run a year or more. A site can have land, zoning, and a tenant and still miss its energization date because the transformer does not arrive in time.

This pulls procurement forward into site selection. The realistic question is no longer just when the utility can interconnect, but whether the equipment timeline matches that date, and ordering long-lead gear early, sometimes before the design is finished, has become normal practice. Treat the equipment lead time as part of the schedule, and confirm the current quotes, because they shift with the market. The grid and substation guide covers the interconnection side of the same schedule. Tie the gear order into the precon schedule so the long-lead items are on order before they sit on the critical path.

How do you score data center sites?

You score data center sites with a weighted scorecard that rates each candidate across the criteria that matter, with power and schedule weighted heaviest, so the comparison is explicit instead of a gut call. The point is to make the tradeoffs visible: a site that wins on land and loses on power should lose overall, and a scorecard forces that result rather than letting a cheap parcel win on price.

Weight the criteria to the project. Power availability and the interconnection timeline carry the most weight in 2026, because they are the binding constraint, followed by water or the cooling plan, fiber, land, hazard, cost, and the policy and incentive picture. A latency-sensitive workload weights connectivity and location higher. A training campus weights power and cost higher and latency near zero. There is no universal weighting, only the right one for the workload in front of you.

Use the score to rank and to shortlist, then put the top candidates through due diligence before any commitment. The scorecard is a decision tool, not the decision. It tells you which two or three sites are worth the cost of a power study, a geotech, and an environmental review, and it gives the investment committee a defensible reason for the choice. The table below is a starting template. Set the weights to the project.

Due diligence before you commit

Due diligence is the confirmation step that turns a high-scoring candidate into a site you can commit to, and it is where the assumptions in the scorecard either hold or fall apart. The studies cost money and time, which is why you run them only on the shortlist, but skipping them is how a deal that looked great on paper becomes a write-off after closing.

The core studies are a short list. A power study, often a formal interconnection study or a load-serving capacity study from the utility, confirms the megawatts and the date in writing rather than off a service map. A geotechnical investigation confirms the soils carry the building and prices the foundation. An environmental assessment, typically a Phase I and a Phase II if it flags anything, finds contamination and wetlands before they become your liability. A title search and a survey confirm you control what you think you control, with the easements and rights mapped.

Confirm, do not assume, is the whole discipline. The utility's verbal headroom becomes a study. The seller's clean site becomes a Phase I. The plan acreage becomes a survey of buildable area after setbacks. Each one can kill or reprice the deal, which is exactly why you do them before you commit, not after. Hedge every timeline and result to what the studies, the utility, and the jurisdiction actually return.

What to document

A site decision that nobody can reconstruct later is a decision nobody can defend when the project is challenged or the deal goes sideways. The record is the scorecard, the power study, the diligence reports, and the assumptions behind each score, kept where the team and the investment committee can find them. Capture it as you go, because reassembling a siting decision from memory months later does not work.

For each candidate, record the power position and the source of the number, the water and cooling plan, the fiber and carrier picture, the land and buildable acreage, the hazard findings, the cost and incentive terms, the score and weights, and the diligence results as they come back. Note who confirmed each item and when, because a utility's answer in January is not the same as its answer in June. A field tool such as FieldOS is a practical place to keep the site evaluation, the scorecard, and the diligence records together with the photos and the dates attached, so the decision trail stays intact from screening through commitment.

Common mistakes

• Assuming the utility can deliver the power because the service map shows headroom, without confirming the queue and the energization date.
• Ignoring the interconnection timeline and underwriting a schedule the grid cannot meet.
• Going to land with no water supply and no air-cooling plan, then discovering the cooling problem after closing.
• Picking cheap land with no fiber, or a single fiber path with no diverse route.
• Overlooking natural-hazard exposure: building in a floodplain or a high-wildfire zone, or skipping the seismic check.
• Treating zoning and permitting as a formality in a jurisdiction that has turned against new data centers.
• Buying land for the first building with no room and no power path for expansion.
• Forgetting the long-lead gear, so the transformer schedule blows past the interconnection date.
• Scoring sites on price and gut instead of a weighted scorecard, then committing before due diligence.

Field checklist

Standards and references

There is no single code that governs data center site selection the way the NEC governs the wiring, so the practice draws on the utility's interconnection process, environmental and geotechnical standards, and industry siting guidance, combined with the local land-use law. Naming the right framework for each piece is what keeps a site decision defensible.

The power side runs through the utility and the grid operator. The interconnection study process is defined by the regional transmission operator or ISO and the utility's own tariff, with the FERC interconnection rules sitting above them, so the timelines and the queue mechanics come from those bodies, not from a building code. The grid and substation guide covers that process in detail.

The diligence side uses established standards. Environmental site assessment follows the ASTM E1527 Phase I framework, with a Phase II where it flags a concern. Geotechnical investigation follows standard geotechnical engineering practice for foundation design. Flood risk is mapped against FEMA flood data, and seismic exposure against the published seismic hazard maps. For the data center itself, industry references such as the Uptime Institute Tier system, ASHRAE TC 9.9 thermal guidelines, and TIA-942 inform the design the site has to support. Treat every timeline and criterion in this guide as project-, utility-, and jurisdiction-dependent: the utility's interconnection answer, the local zoning and permitting rules, and the project's own basis of design control the decision, and all of them are moving in 2026.

Units and terms

A few terms and units carry the site-selection conversation, and they read differently across a utility study, a real estate listing, and an engineering report.

Capacity is stated in megawatts of electrical load, MW, sometimes split into IT load and total facility load. Cooling efficiency is PUE, a ratio with no units, and water use is WUE in liters per kilowatt-hour. Land is in acres, often expressed as acres per megawatt. Latency is in milliseconds of round-trip delay. The same site can read as strong in one report and marginal in another depending on which of these the report leads with.

Site selection

The structured evaluation of candidate sites against power, water, connectivity, land, climate, hazard, and policy

Interconnection queue

The ordered line of projects waiting for the utility or grid operator to study and approve a connection to the grid

Interconnection timeline

The time from queue entry through study and upgrades to actual energization, commonly years for a large load

MW capacity

Megawatts of electrical load a site can be delivered and the building can use, the headline number of a campus

Free cooling

Rejecting heat using cool outside air or a low wet-bulb instead of running mechanical chillers, which lowers power and water

Latency

Round-trip signal delay between users or sites, in milliseconds, scaling with distance and relevant only to some workloads

Entitlement / zoning

The land-use approvals that allow a data center on a parcel, by right or by special exception, granted by the jurisdiction

Due diligence

The confirmation studies, power, geotech, environmental, and title, that verify a shortlisted site before commitment

PUE / WUE

Power usage effectiveness, a facility-efficiency ratio, and water usage effectiveness in liters per kilowatt-hour

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• How does a data center get power?
• What is data center power density?
• What is a rack unit (U)?
• What is the difference between single-mode and multimode fiber?
• How do you size a data center generator?