Data center site selection is the process of scoring candidate locations against the requirements of your workload, and for HPC and AI the order of those requirements matters more than the length of the list. Power comes first: how much you can get, how soon, and at what price. Fiber, land, water, taxes and approval risk follow. A site that fails on power cannot be rescued by strength elsewhere.
This guide walks through the criteria in the order that works, a scoring method, the extra demands of dense GPU and HPC designs, and the risks that surprise teams late in a project.
TL;DR
- Screen for power timing first, then price, land, fiber, water, taxes and approval risk. That is the order a 2026 Lawrence Berkeley National Laboratory report advises developers to use.
- Define success before you look at sites: latency, scalability, sustainability or cost. Then build a weighted scoring model and stress-test it against scenarios such as heat waves and grid outages.
- HPC and AI change the math. Rack densities of 40 to 140 kW today, and liquid cooling above roughly 50 kW per rack, push water, floor loading and electrical design into the site decision.
- Plan logistics, tariffs and vendor lead times at the same time as the site, not after you sign.
Table of Contents
- Start with what the site has to do
- The criteria, in the order that matters
- Build a scoring model you can defend
- Designing for HPC and AI changes the site decision
- Onsite, colocation or new build
- Risks teams find too late
- Where delivery and logistics enter the plan
- Perspective
- How we can help
- FAQ
- Sources
Start with what the site has to do
Before a single location is compared, write down what the facility must deliver. A Felt guide on site selection suggests defining success first, whether that is latency, scalability or sustainability, then building a trade-off model across performance, cost and risk.
For most HPC and AI buyers the requirement sheet should hold:
- Initial and final power draw in kW or MW, and the phase-in schedule.
- Rack density now and in the next hardware generation.
- Target date for first workloads.
- Latency tolerance, which depends on the job. Training and batch work barely care where the building sits. Real-time inference does.
- Data residency or export-control constraints.
- Whether you will own, lease or colocate.
Dropbox describes the same discipline in its own site selection process: capacity requirements first (power, cabinet space, lease start date), then candidate locations, then requests for proposals and a technical questionnaire, and finally a stack-ranking on weighted parameters.
The criteria, in the order that matters
1. Power availability and timing
Power is the gating factor. The Lawrence Berkeley report places power timing at the top of the screening order, and an industry executive roundtable is not needed to see why: a building with no energized capacity is a lease on an empty shell. Ask utilities concrete questions. How many megawatts are available at the substation today? What is the interconnection timeline for the load you need? Is there a second feed? What would an upgrade cost, and who pays?
A Hengeler guide to data center developments in Europe makes the same point about interdependence: a missing grid connection can make a site unviable regardless of its other strengths.
2. Price and price stability
Enverus frames siting as a balance of power, price and land, with price meaning electricity that is competitive and stable over time. This is not abstract. Flexential, citing JLL data, reports that commercial electricity rates rose nearly 30% since 2020, averaging 9.7 cents per kWh in the first half of 2025. For a facility running dense GPU clusters around the clock, a one-cent difference per kWh is a line item that finance will notice.
Look at the tariff structure, not only the headline rate: demand charges, minimum take obligations, and whether contracts allow renewable sourcing.
3. Land and expansion room
TechTarget's seven considerations include an abundance of land, with hyperscale projects needing hundreds of acres, and note that rural areas often offer that land but can bring talent and community-acceptance challenges. Most enterprise and startup projects are far smaller, but the lesson scales down: check that the parcel or the building can absorb a second and third phase without a new permit cycle.
4. Fiber and connectivity
Count the carriers that physically enter the building, and ask about diverse paths. Latency is physics: according to Noded, light travels about 200 km per millisecond through fiber, and the same guide gives practical round-trip targets, such as under 30 ms for real-time trading or gaming and under 150 ms for snappy web applications. Match the target to your actual users. If your workload is asynchronous or batch, connectivity quality and cost matter more than distance.
5. Water and climate
A cooler climate lowers cooling costs, a point TechTarget lists among its considerations. For liquid-cooled AI halls, water also matters as a permitting issue. The Lawrence Berkeley study notes that environmental permits for data centers primarily cover backup generator air emissions and water use and discharge. Ask early whether the local utility can supply water at the volume you need, and whether closed-loop designs are accepted by the authority.
6. Natural hazard and resilience risk
Check flood, wildfire, hurricane, seismic and extreme-heat exposure. Felt notes that in seismic regions, architectural design must match local building standards and soil conditions. Run the scenarios: what happens to the site in a multi-day heat wave or a grid outage? Score the answer instead of assuming it.
7. Workforce
Intel's site selection approach recommends researching both the availability of skilled electrical and mechanical workers and the history of regional labor stoppages. Construction and commissioning of liquid-cooled halls depend on trades that are in short supply in many markets.
8. Taxes, incentives and approval risk
Taxes and incentives vary by state and by country, and approval risk comes last in the Berkeley screening order for a reason: you can only judge it once the other factors narrow the field. Zoning, community opposition and local moratoria still decide whether a shortlisted site is buildable. Ask the permitting office for the typical timeline, not the statutory one.
Build a scoring model you can defend
A scoring model turns a debate into a record. Intel argues that quantitative analysis of factors and trade-offs turns raw data into insight for informed decisions. A simple version works:
- Set pass/fail gates. Minimum MW by your go-live date, at least two fiber carriers, no unmitigable hazard. Sites that fail a gate are out, whatever their score.
- Weight the rest. For an AI training facility you might weight power cost and timing heavily and latency lightly. For inference serving a city, reverse that.
- Score each site 1 to 5 on every criterion, with evidence attached: a utility letter, a carrier map, a hazard report.
- Test the ranking. Change the weights by 10 to 20 percent. If the winner changes, the decision is fragile and you need more data on the close criteria.
- Run scenarios. Heat wave, outage, a 12-month delay in energization. Pick the site that degrades most gracefully.
A sample layout:
| Criterion | Type | Weight | Evidence to collect |
|---|---|---|---|
| Energized MW by go-live | Gate | n/a | Utility will-serve letter |
| Power price and stability | Scored | 25% | Tariff schedule, contract terms |
| Fiber carriers and paths | Gate and scored | 15% | Carrier maps, route diversity |
| Cooling and water | Scored | 15% | Utility water letter, climate data |
| Land or floor expansion | Scored | 10% | Phase plan, zoning |
| Hazard exposure | Scored | 15% | Flood, seismic and heat reports |
| Workforce | Scored | 10% | Trade availability, labor history |
| Taxes and approvals | Scored | 10% | Incentive terms, permit timeline |
The weights are an example, not a standard. Set yours from the requirement sheet.
Designing for HPC and AI changes the site decision
The search for HPC data center design often starts after the site is chosen. That is backwards, because density decides what a site must supply.
An Ethernet Alliance white paper advises planning AI facilities for rack power of 40 to 140 kW today, with whole sites reaching hundreds of megawatts to several gigawatts. It states that air cooling is generally insufficient above 50 kW per rack, where rear-door heat exchangers become preferred, and systems above 120 kW per rack need direct-to-chip cooling. It also describes four separate networks in an AI system: compute at 400G to 800G, storage at 200G to 400G, in-band management at 100G and out-of-band management below 10G.
AMD's data center design guide goes further, citing densities of 150 to 400 kW per rack for modern AI and HPC workloads and noting that at 300 to 400 kW traditional air cooling is not optimal or even feasible, so direct liquid cooling or immersion is needed.
What that means for the site:
- Electrical. Medium-voltage service, transformer space and, for new buildings, 415V distribution to the rack, per the Ethernet Alliance paper.
- Cooling. Liquid cooling needs water or heat-rejection capacity, plumbing paths and floor loading that older buildings may lack.
- Network. Dense fiber and structured cabling between halls, which the Ethernet Alliance paper prefers for larger or expandable networks.
- Growth. If your next hardware generation doubles density, the same building and utility feed must handle it.
NVIDIA's older GPU-ready data center overview from 2018 shows how fast the ceiling moved: it treated 32 kW up to 50 to 60 kW racks as high density. Treat any design guide's numbers as a snapshot, and confirm against the specifications of the hardware you actually plan to buy.
Onsite, colocation or new build
An onsite data center, often called on-premises, is a group of servers you privately own and control, housed in your own facility. Site selection is different there: you rarely pick a location, you evaluate the one you have. The questions become whether the building's utility service, structure and cooling can carry the target density, and whether expansion is possible.
The three paths compare like this:
| Path | Site decision | Main constraint | Fits when |
|---|---|---|---|
| Onsite | Assess an existing building | Utility capacity, floor loading, cooling retrofit | Data must stay close, load is modest |
| Colocation | Choose a provider and market | Available powered capacity at your density | You need speed and shared infrastructure |
| New build or powered land | Choose land, utility and permits | Interconnection timeline, approvals | Large, long-horizon loads |
The criteria in this guide apply to all three. Only the order of effort changes.
Risks teams find too late
Energization slips. A building can be finished while the utility upgrade is not. Tie your lease or purchase milestones to the utility's committed date.
Hardware lead times outrun construction. If servers, GPUs and cooling components arrive months before or after the hall is ready, you pay for storage or idle space. Sequence procurement with the build.
Cross-border costs. Equipment moving into a new country brings duties, tariffs and customs paperwork. These affect the total cost of a location, so model them in the scoring table instead of discovering them at the port.
Community opposition. Particularly in rural or residential-adjacent sites. Engage early and ask what the permitting body has done with comparable projects.
Underused capacity. Flexential reports average server utilization of 12% to 18% of capacity. Size the first phase to realistic utilization and leave room to grow instead of building for peak on day one.
Where delivery and logistics enter the plan
The site choice sets how equipment gets there. Port access, customs, road access and delivery windows should be part of the criteria, and our guide to data center logistics covers moves, delivery and lifecycle planning in detail. If you are relocating existing capacity to the site you choose, the data center migration planning series covers sequencing, and our piece on global vendor management explains how to keep multiple suppliers aligned to one schedule.
Perspective
We have worked in infrastructure since 2011 and in HPC and AI since 2017, and the pattern we see most is teams choosing a site and a hardware design in separate conversations. They meet later, and the meeting is expensive.
Our view is that the site, the rack design, the supply chain and the delivery route are one decision. A location that looks cheap on land and power can lose its advantage once tariffs, shipping lead times and a liquid-cooling retrofit are counted. A slightly dearer site with a clear energization date, a workable port route and room for the next hardware generation often wins on total cost and time.
That is why we handle design, sourcing, deployment, logistics, tax and tariff questions and ongoing support together. Our network of vendors and logistics partners is designed to give clients sound pricing and delivery timelines, and a single project team keeps the site scoring honest, because the people scoring the options also know what it takes to land the equipment there.
Dormedia
How we can help
If you are comparing sites for an HPC or AI deployment, we can help you turn the requirement sheet into a scored shortlist, align it with a hardware configuration, and plan sourcing, shipping and compliance around it. Our services span servers, components, customized solutions, data center solutions, global shipping, tax and tariff support, project management and strategic advisory, followed by ongoing support after go-live.
If you already know your hardware, you can also browse our available inventory to size power and cooling needs against real configurations.
FAQ
How do I select a data center site?
Define what the facility must deliver, then set pass/fail gates for power by your go-live date, fiber and hazard exposure. Score the remaining sites on weighted criteria, test the ranking with scenarios such as a heat wave or grid outage, and pick the site that holds up best.
What factors matter most in data center site selection?
Power availability and timing matter most, followed by electricity price, land, fiber, water, taxes and approval risk, in the order the Lawrence Berkeley report suggests. Climate, workforce and natural hazards refine the shortlist. A failure on power can make a site inviable regardless of the rest.
How do data centers decide where to build?
Operators define capacity needs, find locations that can supply them, request bids and rank facilities on weighted parameters, as Dropbox describes. Power timing and cost usually narrow the field first. Land, connectivity and permitting then separate the finalists.
What is an onsite data center?
It is a set of servers you own and run in your own facility, commonly called on-premises. Planning focuses on whether your existing building can supply the power, cooling and space you need, rather than on choosing among markets.
Sources
- Data Center Siting: Criteria, Policies, and Recent Trends (Lawrence Berkeley National Laboratory)
- Key data center site selection factors for success (Felt)
- How the data center site selection process works at Dropbox
- Power and Place: A Regulatory Guide to Data Centre Developments in Europe (Hengeler)
- Data Center and Large Load Siting Guide (Enverus)
- Data Center Location Strategy: A Leader's Framework (Flexential)
- 7 considerations for data center site selection (TechTarget)
- How to Choose a Data Center Location (Noded)
- Selecting a Data Center Site: Intel's Approach
- Evolving infrastructure requirements for AI data centers (Ethernet Alliance)
- Data Center Design Guide (AMD)
- Considerations for Scaling GPU-Ready Data Centers (NVIDIA)
- What is On-Premises vs. Cloud? (HPE)