Key takeaways
- Power scale is the defining requirement: OpenAI’s Stargate campus in Abilene, Texas, is planned at about 1.2 GW and Meta’s Louisiana campus at up to 5 GW.12
- Campus-scale loads usually connect at transmission voltage; Entergy’s plan for Meta’s campus adds 500 kV lines and a new 500 kV substation.6
- Training sites can be remote if power is available, but McKinsey expects inference to overtake training by 2030, pulling new capacity toward metro areas.75
- Liquid cooling raises density and changes water planning; closed-loop designs can limit on-site water mostly to an initial fill.8
- Grid operators now treat large AI loads as a reliability risk, so expect ride-through and modeling requirements during interconnection.9
- Land for substations, generation and future phases matters as much as land for buildings.
01How AI data centers differ from traditional ones
AI workloads run on dense clusters of accelerators that draw far more power per rack than traditional servers. Uptime Institute’s 2026 survey put the average rack at 7.8 kW and the most common at 11 kW, while a single NVIDIA GB200 NVL72 rack is specified at 132 kW of total power.43 That density changes the building, the cooling and, above all, the scale of power a site must deliver. A traditional colocation building might need tens of megawatts. A large AI training campus can need hundreds of megawatts at first and plan for a gigawatt or more over its life. Our guide to high-density GPU racks covers the building-level effects.
Fig. 1Rack power: industry survey vs. one AI rack
- Average rack, 2026 survey7.8
- Most common rack, 2026 survey11
- GB200 NVL72 rack (vendor spec)132
kW per rack
There are two broad types of AI facility. Training campuses build and refine models and can run anywhere the power and network backbone reach. Inference facilities serve models to users and need to be closer to population centers and network hubs, more like cloud or colocation sites. Our guide to AI training vs. inference site needs goes deeper.
| Factor | Cloud and colocation | AI training campus | AI inference |
|---|---|---|---|
| Typical scale | Tens of MW per building; hundreds of MW for hyperscale campuses | Hundreds of MW to several GW | Wide range; often colocation to hyperscale scale |
| Land | Tens of acres; hundreds for hyperscale campuses | Hundreds to thousands of acres | Similar to colocation or hyperscale |
| Location driver | Users, networks, power | Power first | Users and networks, then power |
| Latency sensitivity | Moderate to high | Low | Moderate to high |
| Rack density | Mostly under 10 kW per rack4 | About 100 kW or more with current GPUs3 | High |
| Cooling | Mostly air, some liquid | Mostly liquid | Mix, trending to liquid |
02Power: scale, voltage and timing
Power is the defining requirement. Loads of this size usually connect directly to high-voltage transmission, with one or more on-site substations. Entergy’s plan to serve Meta’s campus in Richland Parish, Louisiana, for example, includes a new 500 kV substation and two 500 kV lines of about 25 miles each.6 Lower transmission voltages can serve smaller AI deployments, but a full campus usually needs the higher-voltage network. Our guide to transmission voltage for data centers explains why.
Few utilities have hundreds of megawatts sitting idle at a single location. Serving an AI campus often requires new substations, new or upgraded transmission lines and sometimes new generation in the region. In 2025, Louisiana regulators approved three new gas plants as part of Entergy’s plan to serve Meta’s campus.10 That makes timing the core question: when can the first block of capacity arrive, and how fast can it grow?
Some developers bridge the gap with on-site generation, often fueled by natural gas, while waiting for grid service, or plan long-term on-site generation alongside the grid. Epoch AI found that at least three of the seven Stargate sites use on-site gas to avoid long grid-connection queues.11 That shifts site needs toward gas pipeline access, air permitting and extra land. See on-site generation and bridge power.
Load behavior matters too. In July 2024, a fault on a 230 kV line in Northern Virginia was followed by about 1,500 MW of data center load dropping off the grid at once; NERC found the facilities’ own protection and controls had shifted them to backup power.12 In May 2026 NERC issued a Level 3 alert, its highest level, calling on grid planners and operators to take essential actions on large computational loads, including ride-through.9 Expect detailed questions about load profile and ride-through during interconnection; our guide to NERC reliability and large loads explains what is changing.
Redundancy assumptions can differ too. Some AI training facilities accept less electrical redundancy than a mission-critical colocation building, because workloads can checkpoint and restart. Others are built to similar standards. The design choice affects how many utility feeds, substations and backup generators the site must accommodate.
03Land and campus layout
AI campuses need large, contiguous and regular sites. Meta’s Hyperion campus in Louisiana started on a site of about 2,250 acres and later added roughly 1,400 acres as its plans grew toward 5 GW.2 That acreage reflects more than buildings. The site must also hold one or more high-voltage substations, possible on-site generation, cooling plants, water storage, setbacks and buffers, and room for future phases.
- Buildable acreage after floodplain, wetlands and easements, in a shape that fits long rectangular buildings.
- Space for substations at the edge closest to transmission, with room for new line entries.
- Separation from homes for noise from cooling equipment and any generation.
- Heavy-haul access for large transformers and equipment.
- Structural capacity: very dense racks and liquid cooling equipment are heavy, which favors slab-on-grade buildings on good soils.
Fig. 2One AI campus at gigawatt scale
- planned campus capacity
- 5 GW
- original site, before later purchases
- 2,250 acres
- new lines and substation to serve it
- 500 kV
- new gas plants approved in 2025
- 3
04Cooling and water
High rack densities push AI facilities toward liquid cooling, such as direct-to-chip cold plates or immersion, usually alongside some air cooling. Liquid cooling moves heat more efficiently and allows warmer coolant temperatures, which can widen the range of climates where dry or hybrid heat rejection works. NVIDIA’s chief executive has said its Vera Rubin systems can be cooled with 45°C water, without chillers.13
Water needs depend on the heat rejection design. Evaporative systems use more water and can lower energy use; closed-loop systems with dry coolers or air-cooled chillers use far less water but more power in hot weather, an energy penalty one engineering analysis puts at roughly 10–20% for dry cooling.14 The Stargate campus in Abilene sits at the low-water end: developer Crusoe says its closed-loop system uses no water for heat rejection after a one-time fill of about one million gallons per building.8 Sites should be evaluated against the specific design, not a generic figure. Our guides to data center water requirements and closed-loop and waterless cooling cover the choices.
For scale, Berkeley Lab estimated that U.S. data centers directly consumed about 17 billion gallons of water in 2023, while cautioning that limited disclosure makes such estimates uncertain.15 Climate still matters too. Cooler, drier climates give more hours of efficient heat rejection, though power availability usually outweighs climate in AI site selection.
05Fiber and network
Training campuses need high-capacity, diverse fiber to move large datasets and connect to other facilities, but they are less sensitive to latency to end users. That is why they can sit in rural areas with strong power. Inference facilities need low latency to users and so favor sites near metro areas and established network hubs.
The balance is shifting toward inference. McKinsey projects global inference load rising from about 21 GW in 2025 to about 93 GW in 2030, more than training,7 and it sees new builds moving toward metro and near-metro areas, where round-trip times are short and network interconnection is dense.5 Sites with both strong power and good metro access are the scarcest. See latency requirements by workload.
In both cases, the site needs at least two physically diverse fiber paths. Building new long-haul fiber to a remote site is possible but adds time and cost.
Fig. 3Where AI workloads tend to locate
Training campus
Remote, power-rich land; hundreds of MW to GW.
Metro inference hub
Large loads near users; the hardest sites to find.
Batch and fine-tuning
Fits existing colocation where power allows.
Edge inference
Small sites close to users and networks.
Low ← Latency sensitivity → High
06Entitlement and community at gigawatt scale
The scale of AI campuses draws attention. Large buildings, new transmission lines, on-site generation and water use all raise local questions. Zoning that suits a single data center building may not contemplate a multi-building campus with power generation, so approvals can be more involved.
- Confirm whether on-site generation is allowed under zoning and what air permits it would need.
- Check whether new transmission lines to the site would need state siting approval.
- Plan buffers and setbacks for noise from cooling and generation equipment.
- Engage early with local officials on water, traffic, noise and tax effects.
Power supply decisions can become public proceedings of their own. In April 2026, Louisiana regulators voted to put Entergy’s request for seven more gas plants for Meta’s campus on an accelerated review schedule; the plants themselves still needed approval.16 Expect similar scrutiny of who pays for new generation and transmission lines wherever an AI campus needs them, and prepare clear answers before hearings begin.
07Screening sites for AI campuses
When we screen for AI campuses, we start with the grid: where high-voltage transmission and substations could plausibly support hundreds of megawatts, and where the utility or grid operator signals room to grow. Land comes next, filtered for large contiguous buildable acreage near those corridors. Water, gas, fiber and zoning then separate the strong candidates from the rest.
- Deliverable megawatts, and the date the first block can be energized.
- Room for a large substation beside the transmission corridor, with space for new line entries.
- A credible path to on-site generation if grid service is late.
- Water for the chosen cooling design, or a design that needs little.
- Two diverse fiber routes, and latency that suits the target workload.
For a broader comparison of how requirements change by facility type, see site needs by data center type, and for the largest projects, what a gigawatt-scale AI campus requires. To test a specific parcel against these needs, get a site reviewed.
Common questions
How much power does an AI data center need?
It varies widely by size. Smaller AI deployments inside colocation or cloud facilities may need a few to tens of megawatts. The largest training campuses are planned in gigawatts, such as the roughly 1.2 GW Stargate campus in Abilene, Texas, and are usually built in phases.1 The scale depends on the number and type of accelerators, cooling design and the campus plan. Power availability at that scale is the main constraint in site selection.
How much land does an AI data center campus need?
Large AI campuses are typically planned on hundreds to a few thousand acres; Meta’s Louisiana campus started on about 2,250 acres and later added more.2 That land holds multiple buildings plus high-voltage substations, cooling plants, possible on-site generation, buffers and future phases. What matters is buildable, contiguous acreage after floodplain, wetlands and easements are removed, in a shape that fits large rectangular buildings.
Do AI data centers need to be near cities?
Training campuses generally do not. They are less sensitive to latency for end users, so they can locate where large amounts of power and land are available, as long as high-capacity fiber can reach them. Inference facilities, which respond to user requests, benefit from being closer to population centers and network hubs, and McKinsey expects more new capacity in metro and near-metro areas as inference grows.5
Do AI data centers use more water?
Not necessarily. Water use depends on the cooling and heat rejection design, not on whether the workload is AI. Liquid cooling at the chip can pair with closed-loop dry coolers or air-cooled chillers that use little water, or with evaporative systems that use more water and less energy. Dry systems shift the burden to electricity, which matters most in hot weather.14 The right choice depends on climate, water availability, power cost and local rules.
Notes
- 1.R&D World, “Stargate’s $500B bet could force data center and 1.2 GW grid rethink,” n.d. rdworldonline.com
- 2.Data Center Dynamics, “Meta expands data center campus in Richland Parish, Louisiana to 5GW, will invest $50bn,” 2026. datacenterdynamics.com
- 3.Supermicro, “SuperServer SRS-GB200-NVL72,” n.d. supermicro.com
- 4.Uptime Institute, “Uptime Institute Global Data Center Survey 2026,” 2026. intelligence.uptimeinstitute.com
- 5.McKinsey & Company, “The next big shifts in AI workloads and hyperscaler strategies,” 2025. mckinsey.com
- 6.Entergy Louisiana, “NELA Transmission Lines 500kV,” n.d. entergylouisiana.com
- 7.McKinsey & Company, “The future of AI workloads,” n.d. mckinsey.com
- 8.Crusoe, “An Inside Look at the Abilene AI Data Center,” n.d. crusoe.ai
- 9.Morgan Lewis, “NERC Alert Signals Growing Reliability Focus on Data Centers,” 2026. morganlewis.com
- 10.Louisiana Illuminator, “Entergy Louisiana gains regulatory approval for massive Meta power project,” 2025. lailluminator.com
- 11.Epoch AI, “OpenAI Stargate: where the US sites stand,” 2026. epoch.ai
- 12.North American Electric Reliability Corporation, “Incident Review: Considering Simultaneous Voltage-Sensitive Load Reductions,” 2025. nerc.com
- 13.Facilities Dive, “No chillers needed for Vera Rubin server racks: Nvidia CEO,” 2026. facilitiesdive.com
- 14.GHD, “Cooling Hysteria: Reframing Data Centre Water Use,” n.d. nexus.ghd.com
- 15.Lawrence Berkeley National Laboratory, “2024 United States Data Center Energy Usage Report,” 2024. eta-publications.lbl.gov
- 16.FOX 8 (WVUE), “Seven new power plants for Meta data center in northern Louisiana fast-tracked,” 2026. fox8live.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.
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
- Edge Data Center Site Selection: Location, Power and Zoning
- Colocation Data Center Site Selection: Markets, Connectivity and Multi-Tenant Needs
- Converting Bitcoin Mining Sites to AI Data Centers: Power Advantage and Gaps
- Modular and Prefabricated Data Centers: Speed, Footprint and Permitting
- Government, Defense and Sovereign Data Center Siting: Security, Federal Rules and Data Residency
- Enterprise Data Center Site Selection: Distance From HQ, DR Separation and Hybrid Cloud
Guide 2 of 11 in Site strategy by data center type · all 242 guides · glossary · calculators
