A BlackForge Ventures company
BlackForge Data Centers

Site strategy by data center type

Site Needs by Data Center Type: Edge, Colocation and Hyperscale

Edge sites are small, from tens of kilowatts to a few megawatts, and are chosen for proximity to users.1 Colocation sites range from single buildings to wholesale campuses near metro fiber.2 Hyperscale campuses need hundreds of megawatts and hundreds of acres with transmission-level power.3 The largest AI campuses go further, to a gigawatt or more on sites of 1,000 acres or more, and trade latency for power.45 The type of facility decides which site criteria come first.

Last reviewed · 10 min read · BlackForge Data Centers

Key takeaways

  • Edge sites prioritize location and latency; loads run from tens of kilowatts to a few megawatts and are often served at distribution voltage.1
  • Colocation sites prioritize metro access and fiber diversity; tenants lease from a few hundred kilowatts to multi-megawatt blocks, with market rents benchmarked to N+1, Tier III-type space.2
  • Hyperscale sites prioritize transmission-level power, large contiguous land and room to phase; hyperscale operators now hold close to half of the world’s data center capacity.6
  • AI training campuses prioritize power above almost everything else and can sit farther from users, while inference is pulling capacity back toward metros.7
  • A site that is excellent for one type can be a poor fit for another.

01How the main types compare

Data centers are not one product. The same parcel can be a strong candidate for one type and unusable for another. The table below summarizes typical ranges and priorities.

Typical site profile by data center type
TypeIT powerLandPower deliveryTop siting priority
EdgeTens of kW to a few MW1A small building, module or pad; often existing spaceDistribution, sometimes existing serviceProximity to users and networks
Enterprise and colocationUnder 1 MW to tens of MW per building; wholesale campuses far more2Tens of acres for a building, more for a campus8Distribution or transmission, often a dedicated substation at the upper endMetro access and fiber diversity
Hyperscale campusAbout 100 MW to more than 1 GW3Hundreds of acres3Transmission, with one or more on-site substationsPower capacity and timing, contiguous land
AI and gigawatt campus1 GW and up for the largest41,000 acres or more for the largest5High-voltage transmission, sometimes with on-site or bridge generation9Large blocks of power, delivered on schedule

These ranges are typical, not limits. A constrained hyperscale site can fit more power on less land with multistory buildings, and the lines between types are blurring as wholesale colocation providers build hyperscale-sized campuses for cloud tenants. Digital Realty’s Kansas City site, for example, covers 1,440 acres, with a first phase of nine buildings on 280 acres and a power agreement that starts at 600 MW and can scale to 2 GW.10 Land figures also depend heavily on how much of a tract is actually usable, which is covered in how much land a data center needs.

02Edge data centers

Edge data centers place computing close to the people and devices that use it, to cut latency and reduce backhaul traffic. They support content delivery, network functions, local cloud services and applications that need fast response. Because they serve a local market, the location is chosen first and everything else is fitted to it.

Size varies widely. Research Institutes of Sweden, for example, groups sites under 300 kW, which often sit in a company’s own building or at the outer edge of a network, separately from 0.3–1 MW sites at internet exchange points and telecom cores.1 At the larger end, American Tower’s edge facility in Raleigh, North Carolina, opened at 1 MW in about 4,000 square feet and is designed to grow to 4 MW in about 16,000 square feet.11

  • Location: inside or at the edge of the population center served, often near network aggregation points or wireless infrastructure.
  • Power: modest loads that a utility can often serve at distribution voltage, sometimes from existing service.
  • Land: small sites, existing industrial buildings or pads in business parks; reuse is common.
  • Fiber: strong local fiber and access to carrier networks matter more than long-haul routes.
  • Entitlement: closer to homes and businesses, so noise, setback and generator standards and appearance rules can be stricter even at small scale.

Our guide to edge data center site selection goes deeper on how close is close enough, and on the latency each workload can tolerate.

03Enterprise and colocation data centers

Colocation facilities lease space, power and connectivity to many tenants, from enterprises to cloud and network providers. Retail colocation serves smaller deployments in shared halls, while wholesale colocation leases larger blocks of power, often whole suites or buildings. Their value depends heavily on being where customers want to be and where networks interconnect. Enterprise facilities built for a single company share many of the same needs.

  • Location: within or near established metro markets, within practical reach for customer staff and with acceptable latency to the business center.
  • Fiber: multiple carriers and physically diverse routes are core requirements, since interconnection density is part of what tenants buy. See fiber connectivity for data center sites.
  • Power: from a few hundred kilowatts per tenant to multi-megawatt blocks. CBRE tracks asking rents for requirements from 250 kW to more than 10 MW, and in mid-2026 vacancy in primary North American markets was a record-low 1.4%, with more than 80% of capacity under construction already preleased.2
  • Land: tens of acres for a single building, often in suburban industrial or business park settings where land competes with logistics and other uses; one industry estimate puts a typical facility’s footprint near 40 acres.8
  • Redundancy: tenants expect high availability, and market rent benchmarks are quoted for N+1, Tier III-type space, which can be maintained without taking IT offline.2 See data center tier levels and redundancy.

04Hyperscale campuses

Hyperscale campuses serve large cloud and internet platforms. They are built in phases, with several large buildings sharing on-site substations, and they are planned for growth over many years. Land and power are bought for the full build-out even if early phases use a fraction of it. Synergy Research Group estimates that hyperscale operators, counting space they lease as well as space they own, held close to half of worldwide data center capacity at the end of 2025, and it projects about two-thirds by 2031.6

  • Power: often 100 MW to more than 1 GW per campus,3 almost always at transmission voltage with one or more on-site substations. The utility’s ability to deliver capacity on the campus schedule is usually the deciding factor; JLL describes power, not land or capital, as the dominant constraint on site selection and build timelines.12
  • Land: large, contiguous, buildable tracts, with room for substations, setbacks, stormwater facilities and future phases. Data center land purchases averaged about 224 acres in 2024 (through October) as developers shifted to multi-building campuses.13
  • Fiber: diverse long-haul routes to other campuses and network hubs; latency to end users is less critical than for colocation.
  • Water: depends on cooling design; evaporative systems need reliable supply, while air-cooled and closed-loop designs need less.
  • Entitlement and incentives: the scale draws public attention, and CBRE’s mid-2026 report found community pushback and zoning delays slowing projects across North America.2 Tax treatment of equipment can matter a great deal over the campus life.

Our guide to how hyperscalers choose data center sites covers the selection process from the buyer’s side.

05AI and gigawatt-scale campuses

AI training clusters push density and total load well beyond traditional hyperscale, and the largest campuses are now planned in gigawatts.4 Two widely reported U.S. examples show the scale.

Fig. 1Two AI campuses at a glance

Hyperion’s planned capacity after Meta’s July 2026 announcement
5 GW
Hyperion’s original site, before about 1,400 more
2,250 acres
Stargate Abilene, across eight buildings
1.2 GW
Abilene’s total building area
~4M sq ft
As reported through mid-2026 for Meta’s Hyperion campus in Richland Parish, Louisiana,45 and the Stargate campus in Abilene, Texas.14 Plans, tenants and figures change.

Rack density is the other difference. Uptime Institute’s 2025 survey found average densities still rising slowly, driven by more racks in the 10–30 kW range,15 while an NVIDIA GB200 NVL72 rack draws about 120 kW and has to be liquid cooled.16 Training workloads are less sensitive to user latency than interactive services, so these campuses can locate where large blocks of power and land are available, even outside established data center markets. Inference is different: McKinsey expects it to overtake training as the main AI workload by 2030, with build-outs concentrating in metro and near-metro sites chosen for low latency and strong connectivity.7

Fig. 2Approximate rack power by hardware type

  • General-purpose CPU rackup to ~12
  • Air-cooled H100 rack~40
  • GB200 NVL36x2 rack~66
  • GB200 NVL72 rack~120

kW per rack

Approximate figures from SemiAnalysis. Racks well above about 40 kW are the main reason AI clusters need liquid cooling.16

The siting priorities shift accordingly: power capacity and delivery date first, then land, water or cooling strategy, and options for on-site or bridge generation if grid capacity arrives in stages. On-site and bridge generation bring their own entitlement risk. xAI’s Colossus data centers in Memphis have drawn on gas turbines across the state line in Southaven, Mississippi, where a state permit board approved 41 permanent turbines in March 2026, about three weeks after a hearing at which public comment was uniformly opposed.9 The AI data center site requirements and gigawatt-scale AI campus guides cover these in detail.

06How siting priorities change by type

Relative importance of key criteria
CriterionEdgeColocationHyperscaleAI campus
Proximity to usersCriticalHighModerateLow for training
Fiber diversityHigh (local)CriticalHigh (long-haul)High (long-haul)
Power capacityLow to moderateModerate to highCriticalCritical
Power timingModerateHighCriticalCritical
Contiguous acreageLowModerateHighCritical
Water and coolingLowModerateHighHigh
Tax incentivesLowModerateHighHigh
Entitlement pathHighHighHighHigh

Entitlement stays important across every type. Smaller facilities tend to sit closer to neighbors, and larger ones draw more scrutiny because of their scale. The two forces that most often separate the types are how much latency a workload can tolerate and how much power it needs, which the figure below frames as a simple grid.

Fig. 3Latency sensitivity vs. power need

Low ← Latency sensitivity → High

Edge

Small loads placed close to users and networks

Colocation and inference

Metro and near-metro sites with dense fiber

Backup and archive

Modest power and a flexible location

AI training

Power-rich sites, often far from metros

Lower ← Power need → Higher

A way to frame trade-offs, not a rule. McKinsey expects inference to pull more AI capacity toward metro and near-metro sites over time.7

07Matching a site to the right type

For landowners and portfolio holders, the useful question is not “Can this be a data center?” but “What kind of data center could this be?” A 15-acre industrial lot inside a metro with good fiber may be a strong colocation or edge site and of no interest to hyperscale buyers. A 900-acre rural tract near a 345 kV line may suit an AI campus but not a colocation operator. In a typical portfolio, a few properties fit one type well, a few fit another, and many fit none, which is useful to know before anyone spends money on marketing.

When we screen a site or portfolio, we start by placing each property in the type it fits best, then test it against that type’s criteria. That avoids marketing land to the wrong buyers. The full framework is in data center site selection criteria, and if you want a parcel placed in the right category before you approach buyers, you can get a site reviewed.

CalculatorWhite space to IT load →CalculatorIT load to grid demand →CalculatorLand for a load →

Common questions

What is the difference between hyperscale and colocation data centers?

A hyperscale data center is built or leased for a single large cloud or internet operator, usually as a multi-building campus that draws hundreds of megawatts at full build-out. A colocation data center leases space, power and connectivity to many tenants, in blocks from a few hundred kilowatts to several megawatts or more.2 Colocation sites prioritize metro location and fiber diversity, while hyperscale sites prioritize large blocks of transmission-level power. The line blurs when a wholesale provider builds an entire campus for one hyperscale tenant.

How much land does an edge data center need?

Very little by data center standards. Many edge sites are single small buildings, modules or space in existing buildings; American Tower’s Raleigh edge facility opened at 1 MW in about 4,000 square feet.11 Their key requirement is location close to the users and networks they serve, plus good local fiber. Because they sit closer to homes and businesses, local noise and appearance rules can still shape the site.

Why are hyperscale data centers built in rural areas?

Because they need large blocks of power and land that are hard to find in metros. Data center land purchases averaged about 224 acres in 2024,13 and the largest AI campuses run to thousands of acres. Rural and exurban sites near high-voltage lines and substations can offer both, along with fiber routes and, in some places, favorable tax treatment.

How are AI data centers different from hyperscale data centers?

The largest AI campuses plan for a gigawatt or more, and their racks can draw around 120 kW each, which requires liquid cooling.16 Training workloads are less sensitive to latency, so they can go where power is available, while inference is expected to concentrate in or near metros.7 Power delivery timing dominates site decisions for both.

Can one site suit more than one type of data center?

Sometimes. A suburban tract with strong fiber and a nearby substation might suit colocation now and a larger campus later, and some wholesale providers buy land banks big enough for both. Most sites, though, are clearly better for one type, which is why screening should start by asking which type a parcel fits rather than whether it can host a data center at all.

Notes

  1. 1.RISE Research Institutes of Sweden, “The cloud resides in a data center on earth,” n.d. ri.se
  2. 2.CBRE, “North America Data Center Trends H1 2026: State of the Market,” 2026. cbre.com
  3. 3.Wayne Forward, posted by the City of Gibraltar, Michigan, “Wayne Forward Data Centers FAQ,” 2026. cityofgibraltarmi.gov
  4. 4.DatacenterDynamics, “Meta expands data center campus in Richland Parish, Louisiana, to 5GW, will invest $50bn,” 2026. datacenterdynamics.com
  5. 5.DatacenterDynamics, “Meta purchases additional 1,400 acres for Hyperion mega data center expansion,” 2026. datacenterdynamics.com
  6. 6.Synergy Research Group, “Hyperscale Operators to Account for 67% of all Data Center Capacity by 2031,” 2026. srgresearch.com
  7. 7.McKinsey & Company, “The future of AI workloads,” 2025. mckinsey.com
  8. 8.TechTarget, “The increasing concern of data center land acquisition,” n.d. techtarget.com
  9. 9.Mississippi Free Press, “Mississippi Permit Board Grants xAI’s Request for 41 Southaven Gas Turbines to Power Memphis Data Center,” 2026. mississippifreepress.org
  10. 10.Commercial Search, “Digital Realty to Enter Kansas City With 600 MW Facility,” n.d. commercialsearch.com
  11. 11.DatacenterDynamics, “American Tower opens edge data center in Raleigh, North Carolina,” n.d. datacenterdynamics.com
  12. 12.Data Center Frontier, “JLL’s 2026 Global Data Center Outlook: Navigating the AI Supercycle, Power Scarcity and Structural Market Transformation,” 2026. datacenterfrontier.com
  13. 13.Building Design+Construction, “Data center demand pivots to larger sites, emerging markets amid power and cost challenges,” n.d. bdcnetwork.com
  14. 14.DatacenterDynamics, “Crusoe tops out final building at OpenAI Stargate data center campus in Abilene, Texas,” 2025. datacenterdynamics.com
  15. 15.Uptime Institute, “Uptime Institute Global Data Center Survey 2025,” 2025. intelligence.uptimeinstitute.com
  16. 16.SemiAnalysis, “GB200 Hardware Architecture: Component Supply Chain & BOM,” 2024. newsletter.semianalysis.com

Have a site in mind?

Get a straight answer on your land.

Send a parcel number, an address, a map pin or a target load. We’ll tell you what it can support and what it would take.

Start a conversation →

This guide is general information about data center site selection. It is not engineering, legal, tax or investment advice. Requirements vary by state, utility and county, so confirm the specifics for any site with the relevant authorities and advisors.

Terms in this guide

Related guides

More in Site strategy by data center type

Guide 1 of 11 in Site strategy by data center type · all 242 guides · glossary · calculators