How to Evaluate a Data Center Site for AI Computing

Updated 9 November 2026 · Standard formula, shown below

Use this guide to decide whether a proposed data center site can support AI workloads reliably, safely, and at a manageable scale. Before starting, define the expected GPU or accelerator capacity, workload mix, deployment timeline, and acceptable downtime; obtain preliminary utility, site, and network information, and involve qualified power, cooling, structural, and environmental specialists.

What You’ll Need

A target capacity and growth plan, including accelerator count or IT power in megawatts (MW), rack density, and expected utilization.
Workload requirements: training versus inference, peak demand, operating schedule, latency, data location, and availability targets.
Site documents: utility service details, electrical single-line diagrams, land and building plans, geotechnical and hazard reports, and cooling or water information.
Network-provider options, fiber routes, construction estimates, permitting information, and a preliminary operating-cost model.
Access to relevant specialists and the authority to request written confirmations from utilities, providers, and permitting agencies.

Step-by-Step

Turn the workload into measurable requirements. Separate training, fine-tuning, and inference: their utilization patterns, latency needs, and tolerance for interruption can differ. Estimate initial and future IT load, peak and average demand, rack power density, and required network throughput. Include storage, CPUs, memory, and cooling equipment; accelerator nameplate power alone is not the facility load. Ask the equipment and infrastructure teams to validate assumptions.
Screen out unsuitable locations early. Check land use, ownership, access, easements, nearby sensitive uses, and the likely path to permits. Review authoritative flood, wildfire, extreme heat, storm, seismic, and other relevant hazard maps, then commission site-specific studies where needed. Look for physical constraints such as limited construction access, poor drainage, contaminated land, or insufficient room for substations, generators, cooling plant, and future expansion.
Verify power availability, not just proximity to a grid. Ask the electric utility in writing about deliverable capacity, connection point, voltage, available redundancy, upgrade work, schedule, and dependencies. A nearby transmission line or substation does not prove that capacity is available. Confirm whether the site can meet the required load now and at planned expansion, and assess outage history and restoration arrangements. Model the consequences of utility delays and supply interruptions.
Assess electrical resilience and cost. Have electrical engineers review the proposed service, transformers, switchgear, backup generation, uninterruptible power supply, protection, grounding, and maintenance access. Define what failures the design must withstand and how long critical loads need support. Check fuel or other backup-energy logistics and applicable emissions and permitting requirements. Compare total delivered electricity cost and likely upgrade or demand charges using documented assumptions, not an unverified headline rate.
Prove the cooling approach against local conditions. Obtain equipment-level heat output and allowable inlet conditions, then model the facility at the site’s expected peak temperatures and humidity. Evaluate air, liquid, or hybrid cooling as appropriate; account for pumps, heat rejection, redundancy, maintenance, and partial-load performance. Determine water demand, source, seasonal reliability, treatment, discharge route, and permits if water is used. A design that works in mild weather may fail during a heat wave or water restriction.
Check network capacity and route diversity. Confirm available fiber, providers, bandwidth, latency to users and data sources, installation lead time, and service-level commitments. Ask providers to document whether routes share conduits, bridges, or other single points of failure. For distributed training, estimate communication needs between systems and sites; for inference, map the locations of users and dependent services. Two providers do not guarantee diverse paths.
Review the building and land for high-density equipment. Structural engineers should verify floor loading, vibration, clear heights, equipment access, and the route for bringing in heavy components. Check space for racks, power distribution, cooling, storage, fire protection, maintenance, and staged expansion. Evaluate fire detection and suppression with qualified designers for the selected equipment and occupancy; do not assume a generic office or warehouse system is suitable.
Build an operating and compliance plan. Identify staffing, physical security, maintenance coverage, spare-parts logistics, monitoring, incident response, and replacement plans. Confirm applicable building, electrical, fire, environmental, water, and land-use approvals with the relevant authorities and qualified advisers. Map data residency, customer, and contractual requirements for the intended deployment. Treat compliance as a site-specific review, not a checklist copied from another project.
Compare lifecycle cost, schedule, and risk. Estimate land, construction, utility connections, equipment, energy, water, network, staffing, maintenance, and eventual expansion. Show assumptions, ranges, and contingencies. Set decision gates for capacity confirmation, permits, network commitments, and construction milestones. Compare the site with alternatives using the same workload, reliability, and financial assumptions.
Make the decision conditional on evidence. Classify each critical requirement as verified, estimated, or unresolved. Assign an owner and deadline to every open issue. Do not approve a full deployment based only on preliminary utility or provider statements; require written commitments and engineering validation before major irreversible spending. If a requirement cannot be met, reduce the planned capacity, change the design, select another site, or stop.

Common Mistakes

Confusing grid access or a nearby substation with confirmed, deliverable capacity and a realistic connection date.
Using accelerator power as the whole facility load and overlooking cooling, conversion losses, storage, and growth.
Choosing a cooling design without testing hot-weather operation, water limits, and equipment-specific requirements.
Counting multiple network providers as resilient without checking whether their physical routes are independent.
Comparing sites by land price or electricity rate while excluding connection upgrades, delays, operating costs, and permitting risk.
Assuming a site that suits conventional server racks will automatically support the weight, heat, power density, and service needs of AI systems.

Tips

Use one documented assumptions sheet for load, growth, redundancy, and workload; update it when equipment plans change.
Request written, site-specific information from utilities and network providers, including exclusions and prerequisites.
Run sensitivity cases for delayed grid upgrades, peak weather, lower-than-planned utilization, and faster growth.
Keep spare physical capacity and maintainable layouts in the plan, but quantify the cost of that headroom.
Use independent engineering reviews at concept design and before final investment approval.

Troubleshooting

The utility cannot confirm capacity or timing: Ask what studies or upgrades are outstanding and request a written sequence of decision dates. Compare phased deployment, another connection point, or another site; do not treat a verbal estimate as a commitment.

Cooling projections fail during hot or dry conditions: Recheck design weather, heat rejection, water availability, and equipment inlet limits with a qualified engineer. Evaluate a different cooling configuration or reduce the initial density, then validate the revised design under peak conditions.

Fiber appears available but resilience is unclear: Request route maps or written diversity confirmation from each provider. If paths converge, add a genuinely independent route where feasible or document the resulting outage risk and decide whether it meets the workload target.

Permits or construction costs threaten the schedule: Identify the specific approval, study, or long-lead component causing the delay. Rebuild the schedule with the responsible authority or supplier, add realistic contingencies, and compare against a phased or alternative-site plan before committing further.

Different teams produce conflicting capacity estimates: Align units, boundaries, and assumptions—for example, IT load versus total facility load, average versus peak, and current versus future capacity. Have the relevant engineers reconcile the numbers in a single signed-off design basis.

Frequently Asked Questions

How much power does an AI data center need?
There is no universal figure. It depends on accelerator count and power, rack configuration, workload, cooling, redundancy, and planned growth. Estimate the full facility demand with equipment-specific data and engineering review.

Is cheap electricity enough to make a site suitable?
No. Confirm deliverable capacity, connection timing, reliability, upgrade costs, cooling conditions, network access, permits, and operating costs as well as the energy price.

Does AI computing always require liquid cooling?
No. Cooling needs depend on the selected hardware, rack density, ambient conditions, and design. Verify equipment requirements and have the complete cooling system assessed for peak conditions and maintainability.

How important is water availability?
It depends on the cooling design and local conditions. Establish expected demand, seasonal supply, treatment and discharge needs, and applicable approvals; water use can constrain some sites even when power is available.

Can an existing data center be converted for AI workloads?
Possibly, but verify electrical capacity, cooling, structural loading, equipment access, fire design, network, and expansion space. A site designed for lower-density racks may need major upgrades.

What should be confirmed before buying or leasing a site?
At minimum, validate land rights and permitted use, utility capacity and schedule, hazards, cooling resources, network routes, construction feasibility, and a realistic total-cost and delivery plan. Keep unresolved conditions explicit in the decision.

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.