Key takeaways
- A single large campus spreads fixed infrastructure over more megawatts, but needs a grid that can deliver all of them; gigawatt campuses such as Meta’s Hyperion have required new utility generation.3
- Several smaller sites can often be powered sooner, because each asks less of the local system.
- Concentration risk covers more than outages: one utility, one hazard, one community and one approval all sit behind a single campus, and local opposition blocked or delayed an estimated $64 billion of projects over roughly ten months.4
- AI training has traditionally favored one large site, but Microsoft now links Fairwater sites about 700 miles apart for single training jobs.2
- Clustered load is a grid issue too: in July 2024, roughly 60 data center loads in Virginia dropped off the grid at once after a single line fault.56
- A regional cluster can capture much of the benefit of both approaches.
01What the choice really is
The question usually comes up when a developer or operator has a capacity target, say 600 MW, and has to decide how to get there. One option is a single campus of 1,000 acres or more with one or two large substations. Another is several sites of 50 to 200 MW each, spread across a region or across several regions.
Scale has moved quickly. Lawrence Berkeley National Laboratory estimated that U.S. data center electricity use rose from 58 TWh in 2014 to 176 TWh in 2023 and could reach 325 to 580 TWh by 2028.7 Single campuses now reach gigawatt scale. Oracle describes the Abilene, Texas, Stargate campus as eight buildings on about 1,100 acres, and Meta’s Hyperion campus in Richland Parish, Louisiana, is planned on about 2,250 acres.89
Neither approach is right by default. The answer depends on what the grid can deliver at each location, what the workload needs from the network, and how the owner thinks about concentration. Our guide to site needs by data center type covers the land and power profile of each scale, and what a gigawatt-scale AI campus requires covers the largest end.
Fig. 1How big single campuses have become
02One campus and several sites compared
| Factor | One large campus | Several smaller sites |
|---|---|---|
| Power availability | Needs one location with a very large, often multi-source connection | Each site asks less of its local grid; more locations qualify |
| Time to first capacity | Often gated by major transmission upgrades or new generation3 | Smaller loads may fit existing capacity sooner |
| Infrastructure cost per MW | Lower; substations, fiber and roads are shared | Higher; fixed costs repeat at each site |
| Land assembly | Large contiguous tract, often from many owners | Several smaller tracts, each easier to assemble |
| Entitlement | One major approval, high visibility | Several approvals, each more modest |
| Concentration risk | One utility, one hazard profile, one community | Spread across utilities, hazards and jurisdictions |
| Operations | One staff base, one security perimeter | Multiple teams or remote operations |
| Network fit | Suits tightly coupled workloads such as AI training | Suits distributed cloud regions, inference and colocation |
03The case for one large campus
Scale economics are real. A large substation, a fiber build, a water main extension and an access road cost roughly the same whether they serve 200 MW or 400 MW on the same site. Spreading that fixed cost over more load lowers the cost per megawatt. Operations benefit too: one security team, one maintenance base, one set of spares.
- Tightly coupled workloads. Large AI training clusters work best when the compute sits together on a short, high-bandwidth network. Microsoft built its Fairwater sites as two-story buildings partly to pack compute closer together and reduce latency.2 See AI data center site requirements.
- One negotiation. A single development agreement, incentive package and utility service agreement can cover the full build.
- Phasing flexibility. A large campus can add buildings as demand arrives without finding a new site each time. See phased vs. full build-out.
- Grid planning. A utility planning one large load can sometimes justify a backbone-scale investment that several scattered smaller loads would not. In Louisiana, regulators approved three Entergy gas plants totaling about 2.3 GW in August 2025 tied to Hyperion, and Entergy later sought approval for seven more.310
That last point cuts both ways. A campus that needs its own generation fleet inherits that fleet’s regulatory process, cost recovery questions and construction schedule; Entergy expected the first two Hyperion plants in late 2028 and the third by the end of 2029.3
04The case for several smaller sites
The main argument is power. Locations that can serve 500 MW or more on a useful timeline are scarce. Locations that can serve 50–150 MW are more common. Splitting the target across sites widens the search and often brings first capacity online sooner. In the busiest market, the Data Center Coalition told Virginia regulators in 2026 that Dominion Energy’s proposed large-load queue could take up to 15 years to cycle through.11
- Resilience. Separate sites on separate utility systems, with different flood, storm and seismic exposure, are less likely to fail together.
- Entitlement risk. A very large campus draws more attention, and a single denial or delay stops the whole program. Data Center Watch counted $18 billion of projects blocked and $46 billion delayed by local opposition between May 2024 and March 2025, across 24 states.4 Several modest approvals spread that risk.
- Land. Assembling 1,000 contiguous acres can mean dealing with many owners. Several 100–200 acre tracts are easier to tie up. See parcel assembly for data centers.
- Market reach. Colocation and cloud providers often want capacity near several customer bases rather than in one place. See colocation site selection.
05How the grid usually settles the question
In theory, the choice is a strategic one. In practice, the utility’s answer often makes it. A utility asked to serve 600 MW at one point may respond with a phased plan that reaches full load only after major transmission work. The same utility, or a neighboring one, may be able to serve 150 MW at each of several substations much sooner, because each request fits within existing capacity or needs only local upgrades.
That is why the comparison should be run on dated megawatts, not end-state megawatts. Ask what each option delivers in each year of the ramp. A single campus that reaches 600 MW in year eight may be worth less than three sites that reach 450 MW combined in year four, even if the single campus is cheaper per megawatt at full build.
Fig. 2Dated megawatts: one campus vs. three sites
Illustrative- One campus, year 4150
- Three sites, year 4450
- One campus, year 8600
- Three sites, year 8600
MW online
- Large single loads are more likely to need new high-voltage lines, substations or generation, which carry long lead times for equipment, routing and permits.
- Smaller loads spread across substations are more likely to fit within existing transformer and line capacity.
- Large loads face new operating rules. In Texas, Senate Bill 6 applies large-load standards at 75 MW and gives ERCOT authority to curtail large loads during grid emergencies.12 See the Texas SB 6 guide.
- Grid operators are also studying concentrated data center load. On July 10, 2024, a fault on a 230 kV line in Virginia led data centers at about 60 load points across 25 substations to switch to backup power at the same time.56 When a similar fault hit the same area in April 2026, after some data centers had adjusted their protection settings, sustained data center load loss was about 500 MW.13 In July 2026, FERC directed NERC to develop mandatory reliability standards for data centers and other computational loads.14 See NERC reliability and large loads.
06The middle path: a regional cluster
Many large portfolios end up as clusters. Several sites sit within the same metro or region, close enough to be linked by owned or leased fiber and to share an operations team, but far enough apart that a single substation outage, flood or local decision does not affect them all.
The definition of “close enough” is stretching. Microsoft’s Fairwater sites in Wisconsin and Atlanta sit roughly 700 miles apart and are linked by a dedicated AI wide area network so that several sites can support one training job; Microsoft says it has deployed 120,000 miles of dedicated fiber for the network.2 That is a hyperscaler-scale investment, but it shows that the network, not just geography, now sets how far apart cooperating sites can be. Our guides to fiber route diversity and latency and AI training vs. inference cover the limits.
A cluster has its own checks. Are the sites served by different substations, or ideally different transmission paths? Are they in the same jurisdiction, and would a local restriction hit all of them? Can fiber between them take diverse routes? The Virginia load-loss events showed that many data centers on one part of the grid can respond to the same disturbance at once.5 The answers decide whether the cluster actually reduces risk or just looks like it does.
Fig. 3Single campus, regional cluster, dispersed sites
Single campus
- Lowest cost per MW at full build
- Needs one very large power path
- One approval, one community
- Best for tightly coupled AI training
Common middle path
Regional cluster
- Shared operations team and fiber
- Can use several substations
- Some shared hazard and policy exposure
- Works for cloud regions and phased AI
Dispersed sites
- Most resilient to local events
- Fixed costs repeat at each site
- Separate utilities and approvals
- Fits colocation, inference and edge
07How to decide: a short checklist
- 01Define the total capacity target and the ramp schedule, not just the end state. See load ramp schedules and phased power.
- 02Decide whether the workload needs to sit in one place, or can run across sites linked by fiber.
- 03Screen for locations that can deliver the full target. If none can on a useful timeline, the decision is made for you.
- 04Compare infrastructure cost per megawatt for one campus against the repeated fixed costs of several sites.
- 05Map concentration risk: utility, transmission path, hazards, jurisdiction and community for each option.
- 06Weigh entitlement exposure: one large, visible approval against several smaller ones. Opposition has been bipartisan, so don’t assume local politics protect a site.15
- 07Keep alternates in the pipeline until the primary site’s power path is confirmed.
When we screen candidates for a portfolio, we compare dated megawatts and open risks for each configuration rather than the end-state headline. If you want a set of parcels compared that way, get a site reviewed.
Common questions
Is it cheaper to build one large data center campus or several smaller ones?
Per megawatt, one large campus is usually cheaper to build, because substations, fiber, roads and water infrastructure are shared across more load. That advantage can disappear if the large site needs major transmission upgrades or new generation the customer must fund, or if delays push revenue years out. Several smaller sites cost more per megawatt in fixed infrastructure but can reach first capacity sooner, which matters when grid waits run four years or more.1
How far apart should data center sites be for resilience?
There is no single rule. The goal is that one event, such as a substation outage, flood, storm or local decision, does not take out more than one site. That usually means different substations or transmission paths and different hazard exposure. For workloads replicated across sites, the network design sets an upper limit on distance, so the right spacing depends on the application. See latency requirements by workload.
Why do AI data centers favor one large campus?
Large AI training jobs run across thousands of processors that exchange data constantly. That works best when the hardware sits together on a short, high-bandwidth network, so most training capacity is still concentrated. Some builders are now linking distant sites: Microsoft describes its Fairwater sites, about 700 miles apart, as a network of sites supporting one job.2 Inference is more flexible and can run across several sites closer to users.
How many acres does a large data center campus need?
It varies widely with building density, cooling design, substation and generation footprint, setbacks and how much land is lost to wetlands, floodplain or slopes. For scale, the Abilene Stargate campus is described as eight buildings on about 1,100 acres, and Meta’s Hyperion campus is planned on about 2,250 acres.89 Several smaller sites usually need more total acreage, because buffers and infrastructure repeat at each one.
Can a large campus change plans after it starts?
Yes, and large ones often do. At Abilene, OpenAI had planned to expand the site to about 2.1 GW but redirected that capacity elsewhere, according to Epoch AI’s 2026 tracking.16 That is one more reason to keep alternate sites in the pipeline.
Notes
- 1.Electrical Contractor, “How Long Does It Take to Power a Data Center?,” 2026. ecmag.com
- 2.Microsoft, “From Wisconsin to Atlanta: Microsoft connects datacenters to build its first AI superfactory,” 2025. news.microsoft.com
- 3.Gas Processing & LNG, “Entergy to build gas plants for Meta’s AI data center,” 2025. gasprocessingnews.com
- 4.Data Center Watch, “$64 billion of data center projects have been blocked or delayed amid local opposition,” 2025. datacenterwatch.org
- 5.North American Electric Reliability Corporation, “Incident Review: Considering Simultaneous Voltage-Sensitive Load Reductions,” 2025. nerc.com
- 6.Data Center Dynamics, “Virginia narrowly avoided power cuts when 60 data centers dropped off the grid at once,” 2025. datacenterdynamics.com
- 7.Lawrence Berkeley National Laboratory, “2024 United States Data Center Energy Usage Report,” 2024. energyanalysis.lbl.gov
- 8.Oracle, “Oracle OAI Data Center Fact Sheet,” 2025. arrington.house.gov
- 9.Data Center Dynamics, “Entergy reports surge in interest from data centers seeking to build in Louisiana,” n.d. datacenterdynamics.com
- 10.The Lens, “Meta triples order of gas-fired power plants for its AI campus in Louisiana,” 2026. thelensnola.org
- 11.Virginia Mercury, “New state law mandates review of Dominion’s load forecasting, as data centers raise concerns,” 2026. virginiamercury.com
- 12.Foley & Lardner, “Senate Bill 6 Positions Texas to Better Manage Large Electrical Loads and Strengthen Grid Reliability,” 2025. foley.com
- 13.North American Electric Reliability Corporation, “Reliability Insights: Large Computational Load Risks Due to Voltage Sensitivity,” 2026. nerc.com
- 14.POWER Magazine, “FERC Orders Mandatory NERC Reliability Standards for Data Center and Other Computational Loads,” 2026. powermag.com
- 15.Virginia Mercury, “Report highlights community pushback stalling $64 billion in data center development nationwide,” 2025. virginiamercury.com
- 16.Epoch AI, “OpenAI Stargate: Where the U.S. Sites Stand,” 2026. epoch.ai
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.
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 Decisions and trade-offs
- Buy, Option or Lease: How Developers Secure Data Center Land
- Phased vs. Full Build-Out: How to Stage a Data Center Campus
- New Site or Expand an Existing Campus? How to Decide
- Grid Power vs. On-Site Generation: How to Decide
- Near the Metro or Near the Power? Choosing a Data Center Location
- Build, Lease or Colocate? How to Source Data Center Capacity
- Regulated Utility vs. Competitive Market Sites for Data Centers
- Rural vs. Suburban Data Center Sites: Land, Power, Labor and Local Politics
- Tax Incentives vs. Power Costs: Which Matters More for a Data Center Site?
- Utility-Owned vs. Customer-Owned Substations for Data Centers
- Converting a Warehouse or Factory to a Data Center vs. Building New
- How to Choose a Data Center Site Selection Firm
- Site Selection Consultant vs. Broker vs. Engineer: Who Does What
Guide 5 of 14 in Decisions and trade-offs · all 242 guides · glossary · calculators
