Data Center Site Selection: Criteria, Data Layers, Process

Data center site selection is the process of finding land for a facility of tens to hundreds of megawatts. On suitable land, the facility can be powered, connected, cooled, and permitted, and kept clear of natural hazards. Analysts screen candidate land against grid capacity, transmission, fiber, water, hazards, zoning, and latency to users. They then rank the land that passes on cost and time to build. This article tests one claim: power limits where a data center can go, and the rest of the screen finds suitable land near power.

Most of those criteria are geographic. The screen combines map layers in a geographic information system (GIS) and ranks land with multi-criteria decision analysis (MCDA), a method that turns several criteria into one score.

Key takeaways

• Power is the first screen for a data center site: a substation with spare capacity and a short path to high-voltage transmission.
• The criteria can conflict: cheap land sits far from fiber, water sits in flood plains, and spare grid capacity sits far from users.
• The standard method is GIS-based MCDA. Constraints remove land, factors rank what remains, and weights combine the factors into one score.
• An illustrative screen of 1,808 H3 hexagonal grid cells in eastern Loudoun County, Virginia, runs as one SQL query in Wherobots in 13.9 seconds; 41 cells score full marks.
• A parcel screen of 52 West Texas counties finds 3.9% of buildable land within an estimated 5 km line route of 345 kV. Two weightings share 9 of their top 50 parcels.

What is data center site selection?

Hyperscalers (the largest cloud operators), colocation operators (who rent out data center space), and their advisors run these screens for new campuses. Utilities run them in reverse to find where they can serve a large new load. Investors use them to check portfolios for hazard exposure.

The stakes come from demand. Masanet and colleagues estimated that global data centers used about 205 TWh (terawatt hours) in 2018, around 1% of world electricity. The estimate appears in Recalibrating global data center energy-use estimates (Science, 2020). The Lawrence Berkeley National Laboratory 2024 United States Data Center Energy Usage Report by Shehabi and colleagues puts U.S. data center use at 176 TWh in 2023. That was about 4.4% of U.S. electricity. The report projects 325 to 580 TWh by 2028. The International Energy Agency's Energy and AI report projects global use rising from about 415 TWh in 2024 to about 945 TWh in 2030.

Data center site selection criteria

Each criterion maps to data layers and to either a hard filter (a constraint that removes land) or a score (a factor that ranks it):

Criterion	Typical data	Hard filter or score
Power capacity and interconnection	Utility load forecasts, interconnection queue records, substation capacity	Hard: no firm capacity by the target date; score: available megawatts
Transmission and substations	Transmission lines by voltage, substation locations	Hard: distance to 230 kV or higher; score: distance
Fiber	Long-haul and metro routes, carrier hotels, internet exchanges	Score: distance and route diversity
Water and cooling	Municipal capacity, water rights, aquifers, watershed stress, climate	Score, or hard where water is unavailable
Land and zoning	Parcel size and shape, zoning, ownership, slope	Hard: under the minimum acreage, wrong zoning, steep slope
Natural hazards	Flood zones, seismic hazard, wildfire, severe wind	Hard: mapped flood zone, high seismic or wildfire hazard
Latency to population	Population and enterprise density, network distance	Score
Tax incentives	Sales and use tax exemptions, abatement and opportunity zones	Score
Land cost	Assessed values, recent sales	Score

Power capacity and interconnection queues. A hyperscale campus needs firm capacity, power the utility commits to deliver at all hours. The grid connection usually sets the schedule. Generators also wait in interconnection queues, the lists of projects waiting to connect to the grid. Lawrence Berkeley National Laboratory's Queued Up: 2024 Edition by Rand and colleagues counted the queues at the end of 2023. They held nearly 2,600 GW (gigawatts; a gigawatt is a million kilowatts) of generation and storage. The typical project built in 2023 took nearly five years from request to commercial operation.

Transmission lines and substations. Distance to a high-voltage line and a substation with spare capacity sets the cost of the connection. New lines take years to build.

Fiber. Data centers need diverse fiber routes to carriers and other data centers. Fiber maps are mostly commercial, so open-data screens use highway and rail corridors as a rough stand-in.

Water and cooling. Evaporative cooling consumes water, and climate sets how many hours a year outside air can cool the servers. In dry regions water becomes a hard constraint.

Land and zoning. A campus needs a large, flat, contiguous parcel zoned for the use, or a realistic path to rezoning. Slope comes from a digital elevation model, wetlands and forest from land cover, and the legal unit is the parcel.

Natural hazards. The Federal Emergency Management Agency (FEMA) National Flood Hazard Layer maps the 1% and 0.2% annual chance flood zones that most screens exclude (see flood risk data). Seismic hazard comes from the U.S. Geological Survey (USGS) 2023 National Seismic Hazard Model described by Petersen and colleagues (Earthquake Spectra, 2024). Wildfire and wind layers complete the hazard screen.

Latency to population. Latency is the delay for data to travel between users and servers. A common rough rule divides the straight-line distance times 2.1 by the speed of light in fiber. Bozkurt and colleagues cite it in Dissecting latency in the Internet's fiber infrastructure (2018). Workloads that need fast responses stay near users and internet exchanges. AI training can sit farther away.

Tax incentives and land cost. State sales and use tax exemptions on equipment, local property tax rates, and land prices enter as scores.

How the data center site selection process works

The data center site selection process turns a region into a ranked shortlist:

1. Define the study area and the unit of analysis. A region, a grid of cells (such as H3 hexagons), or parcels.
2. Apply constraints. Remove every cell or parcel that fails a hard rule: in a flood zone, too steep, too small, or too far from transmission.
3. Score the factors. Convert each criterion to a common 0 to 1 scale, such as 1 within 2 km of a substation and 0 beyond 5 km.
4. Weight and combine. Multiply each factor by its weight and add the results. This is the weighted linear combination, also called a weighted overlay.
5. Test and refine. Vary the weights, check which cells stay at the top, then send the shortlist to due diligence and utility discussions.

Each layer enters through a spatial join, which matches records by location, the core operation of geospatial analysis for siting. Distance to transmission is a distance join. Flood exposure is a point-in-polygon test. Slope comes from zonal statistics, a summary of elevation values inside each cell or parcel.

Weights from the Analytic Hierarchy Process. Thomas Saaty introduced the Analytic Hierarchy Process (AHP) in A scaling method for priorities in hierarchical structures (Journal of Mathematical Psychology, 1977). Decision makers compare criteria two at a time on a 1 to 9 scale. A matrix calculation, the principal eigenvector, turns the comparisons into weights. Saaty's rule of thumb is to revisit any matrix whose consistency ratio, a check for contradictory judgments, is above 0.10.

Data center site selection examples

• Northern Virginia power constraint, 2022. Loudoun County calls itself the data center capital of the world. In July 2022, Dominion Energy told data center customers that power for some new facilities in eastern Loudoun County would be delayed for years. The cause was transmission that had not kept pace with demand, as Data Center Dynamics reported.
• Texas statewide suitability. Arzumanyan and colleagues weighted criteria with machine learning and multi-criteria decision making. Their paper is Geospatial suitability analysis for data center placement: a case study in Texas, USA (Sustainable Cities and Society, 2025). Energy access and infrastructure carried 31% of the combined weight and environmental hazards 35%. The Dallas, Austin, Houston, and San Antonio metros ranked most suitable.
• United Kingdom regional screen. Hussain and colleagues built a national suitability map from excluded land, AHP weights, and weighted linear combination. Their paper is GIS-driven regional assessment for sustainable data center siting in the United Kingdom (Land, 2026).
• Australia on Wherobots. A Wherobots user ranked 17 candidate industrial sites across all eight Australian states and territories. The ranking drew on 15.91 million geometries from 16 national and state portals. The results are in a national AI data center suitability report with adjustable weights.

Data center site selection in Wherobots

WherobotsDB runs a first-pass screen as one SQL query over open layers in the Havasu catalog. The worked example is an illustrative screen of eastern Loudoun County, Virginia. The study box runs from 77.75°W to 77.30°W and 38.85°N to 39.15°N. It covers Leesburg, Ashburn, Sterling, Dulles Airport, and Chantilly. The unit of analysis is the H3 resolution 8 hexagon, under a square kilometer and close to the size of a campus parcel. ST_H3CellIDs returns the 1,808 cells whose centers fall in the box.

Latency is left out on purpose: every cell sits inside the Washington metro area, so it barely varies between cells. Population from Census block groups and American Community Survey (ACS) estimates belongs in a regional screen that compares metros.

Overture power, building, and road layers feed the four factors, and the Overture water layer feeds the constraint. Each Overture table is stored in Havasu, Wherobots' spatial table format built on Apache Iceberg. Row counts cover each table's search box in the query:

Source	Format	Rows the query reads	What the query takes from it
H3 grid from ST_H3CellIDs	Generated in the query	1,808 cells	Unit of analysis
Overture infrastructure (base_infrastructure)	Havasu table	494 power lines and substations tagged 230 kV or more	Power: distance to the nearest substation and line. Voltages come from OpenStreetMap voltage tags. The study box holds 422 lines and 113 substations. Of these, 349 lines and 79 substations are 230 kV or more, including 44 lines and 7 substations at 500 kV
Overture buildings (buildings_building)	Havasu table	196,661 building footprints	Land: built area per cell, each building counted in the cell that holds its centroid
Overture transportation (transportation_segment)	Havasu table	2,318 motorway and trunk segments	Access: distance to the nearest highway, a rough stand-in for fiber rights-of-way
Overture water (base_water)	Havasu table	97 rivers and lakes	Flood proxy: cells to exclude

The water layer stands in for FEMA flood zones, which a production screen loads from the National Flood Hazard Layer (NFHL) through the ArcGIS Feature Service reader.

The weights are illustrative. A real study sets them with stakeholders, for example through AHP, and tests the ranking's sensitivity to them.

The inventory query counts the power features that intersect the box, grouped by each feature's highest tagged voltage. It ran in 91 seconds on a Small SQL session:

-- Power lines and substations in the study box, by highest tagged voltage
SELECT 'power' AS layer, class,
       CASE WHEN kv >= 500000 THEN '500 kV' WHEN kv >= 230000 THEN '230 kV'
            WHEN kv IS NULL THEN 'untagged' ELSE 'under 230 kV' END AS band,
       COUNT(*) AS n
FROM (SELECT class, array_max(transform(split(source_tags['voltage'], ';'), v -> CAST(v AS INT))) AS kv
      FROM wherobots_open_data.overture_maps_foundation.base_infrastructure
      WHERE bbox.xmax > -77.75 AND bbox.xmin < -77.30 AND bbox.ymax > 38.85 AND bbox.ymin < 39.15
        AND subtype = 'power' AND class IN ('power_line', 'substation'))
GROUP BY 1, 2, 3
ORDER BY 2, 3;

Each criterion becomes a column per cell, and the score is the weighted sum:

-- Illustrative multi-criteria screen for data center sites on H3 resolution 8 cells
WITH cells AS (
  -- 1,808 cells over eastern Loudoun County and its neighbors, centers projected to UTM zone 18N meters
  SELECT /*+ REPARTITION(8) */ h3,
         ST_Transform(ST_Centroid(ST_H3ToGeom(array(h3))[0]), 'EPSG:4326',
                      '+proj=utm +zone=18 +datum=WGS84 +units=m +no_defs') AS cm
  FROM (SELECT explode(ST_H3CellIDs(ST_PolygonFromEnvelope(-77.75, 38.85, -77.30, 39.15), 8, false)) AS h3)
),
hv AS (
  -- Power lines and substations tagged 230 kV or more (OpenStreetMap voltage tags, in volts)
  SELECT class, ST_Transform(geometry, 'EPSG:4326', '+proj=utm +zone=18 +datum=WGS84 +units=m +no_defs') AS gm
  FROM wherobots_open_data.overture_maps_foundation.base_infrastructure
  WHERE bbox.xmax > -77.82 AND bbox.xmin < -77.23 AND bbox.ymax > 38.80 AND bbox.ymin < 39.20
    AND subtype = 'power' AND class IN ('power_line', 'substation')
    AND array_max(transform(split(source_tags['voltage'], ';'), v -> CAST(v AS INT))) >= 230000
),
roads AS (
  SELECT ST_Transform(geometry, 'EPSG:4326', '+proj=utm +zone=18 +datum=WGS84 +units=m +no_defs') AS gm
  FROM wherobots_open_data.overture_maps_foundation.transportation_segment
  WHERE bbox.xmax > -77.82 AND bbox.xmin < -77.23 AND bbox.ymax > 38.80 AND bbox.ymin < 39.20
    AND subtype = 'road' AND class IN ('motorway', 'trunk')
),
wet AS (
  -- Constraint: every cell a mapped river or lake touches (a flood-plain proxy that stands in for FEMA flood zones)
  SELECT DISTINCT explode(ST_H3CellIDs(geometry, 8, true)) AS h3
  FROM wherobots_open_data.overture_maps_foundation.base_water
  WHERE bbox.xmax > -77.75 AND bbox.xmin < -77.30 AND bbox.ymax > 38.85 AND bbox.ymin < 39.15
    AND subtype IN ('river', 'lake')
),
built AS (
  -- Building footprint area per cell, each building assigned to the cell holding its centroid
  SELECT ST_H3CellIDs(ST_Centroid(geometry), 8, false)[0] AS h3, SUM(ST_AreaSpheroid(geometry)) AS m2
  FROM wherobots_open_data.overture_maps_foundation.buildings_building
  WHERE bbox.xmax > -77.76 AND bbox.xmin < -77.29 AND bbox.ymax > 38.84 AND bbox.ymin < 39.16
  GROUP BY 1
),
-- Distance from each cell center to the nearest feature, with the small side broadcast
d_sub  AS (SELECT /*+ BROADCAST(h) */ c.h3, MIN(ST_Distance(c.cm, h.gm)) AS m FROM cells c JOIN hv h
             ON h.class = 'substation' AND ST_DWithin(c.cm, h.gm, 5000) GROUP BY c.h3),
d_line AS (SELECT /*+ BROADCAST(h) */ c.h3, MIN(ST_Distance(c.cm, h.gm)) AS m FROM cells c JOIN hv h
             ON h.class = 'power_line' AND ST_DWithin(c.cm, h.gm, 3000) GROUP BY c.h3),
d_road AS (SELECT /*+ BROADCAST(r) */ c.h3, MIN(ST_Distance(c.cm, r.gm)) AS m FROM cells c JOIN roads r
             ON ST_DWithin(c.cm, r.gm, 5000) GROUP BY c.h3),
f AS (
  SELECT c.h3, ds.m AS sub_m, dl.m AS line_m, dr.m AS road_m,
         COALESCE(bu.m2, 0) / ST_AreaSpheroid(ST_H3ToGeom(array(c.h3))[0]) AS built_share,
         w.h3 IS NOT NULL AS wet
  FROM cells c
  LEFT JOIN d_sub ds ON ds.h3 = c.h3
  LEFT JOIN d_line dl ON dl.h3 = c.h3
  LEFT JOIN d_road dr ON dr.h3 = c.h3
  LEFT JOIN built bu ON bu.h3 = c.h3
  LEFT JOIN wet w ON w.h3 = c.h3
),
s AS (
  -- Factor scores: 1, 0.5, or 0
  SELECT *,
         CASE WHEN sub_m  <= 2000 THEN 1 WHEN sub_m  <= 5000 THEN 0.5 ELSE 0 END AS s_sub,
         CASE WHEN line_m <= 1000 THEN 1 WHEN line_m <= 3000 THEN 0.5 ELSE 0 END AS s_line,
         CASE WHEN built_share < 0.02 THEN 1 WHEN built_share < 0.05 THEN 0.5 ELSE 0 END AS s_land,
         CASE WHEN road_m <= 2000 THEN 1 WHEN road_m <= 5000 THEN 0.5 ELSE 0 END AS s_road
  FROM f
)
-- Illustrative weights: substation 0.30, 230 kV+ line 0.25, open land 0.30, highway 0.15
SELECT lower(hex(h3)) AS h3, ROUND(sub_m) AS sub_m, ROUND(line_m) AS line_m, ROUND(road_m) AS road_m,
       ROUND(built_share, 4) AS built_share, wet, s_sub, s_line, s_land, s_road,
       CASE WHEN wet THEN NULL
            ELSE ROUND(0.30 * s_sub + 0.25 * s_line + 0.30 * s_land + 0.15 * s_road, 3) END AS score
FROM s
ORDER BY score DESC NULLS LAST, sub_m;

The distance joins copy the small layers (substations, lines, highways) to every worker. EXPLAIN FORMATTED shows three such broadcast index joins and no Cartesian product, the slow all-pairs comparison. Buildings and water join on the H3 cell ID, so no polygon is intersected with another. The one query built the grid, read four tables, and ran eight joins: three distance joins and five joins on the H3 cell ID. It scored all 1,808 cells in 13.9 seconds on a Small SQL session.

To count cells per factor score, and to test the weights, the summary query reuses the main query. It keeps every named subquery, or common table expression (CTE), through s and adds a comma after it. It then replaces the final SELECT with the scored CTE and SELECT below. It ran in 12.5 seconds:

scored AS (
  SELECT *,
         CASE WHEN wet THEN NULL ELSE 0.30 * s_sub + 0.25 * s_line + 0.30 * s_land + 0.15 * s_road END AS score,
         CASE WHEN wet THEN NULL ELSE 0.25 * (s_sub + s_line + s_land + s_road) END AS score_equal
  FROM s
)
SELECT crit, lvl, COUNT(*) AS cells
FROM (SELECT stack(8, 'substation', CAST(s_sub AS DOUBLE), 'line_230kv', CAST(s_line AS DOUBLE),
                      'open_land', CAST(s_land AS DOUBLE), 'highway', CAST(s_road AS DOUBLE),
                      'near_water_excluded', CAST(IF(wet, 1, 0) AS DOUBLE),
                      'score', ROUND(CAST(score AS DOUBLE), 3), 'score_equal', ROUND(CAST(score_equal AS DOUBLE), 3),
                      'both_085', CAST(IF(ROUND(score, 3) >= 0.85 AND ROUND(score_equal, 3) >= 0.85, 1, 0) AS DOUBLE)) AS (crit, lvl)
      FROM scored)
GROUP BY crit, lvl
ORDER BY crit, lvl;

Loudoun results with 395 cells excluded and 41 at full marks

Each criterion removes a different part of the map:

Criterion	Scores 1	Scores 0.5	Scores 0
Substation, 230 kV or more	535 cells within 2 km	699 within 5 km	574
Power line, 230 kV or more	407 cells within 1 km	547 within 3 km	854
Open land	986 cells under 2% built	234 under 5%	588
Highway access	854 cells within 2 km	548 within 5 km	406

The river and lake constraint gives 395 cells a null score, which leaves 1,413 scored. 41 cells score 1.0, full marks on every factor. The top cell, 882aaabb19fffff near Arcola, sits 61 m from a 230 kV substation, 117 m from a 230 kV line, and 253 m from a highway. Buildings cover 0.57% of it.

The open-land factor penalizes the existing cluster. Cells in the substation ring around Ashburn and Sterling score mid-range because buildings already cover them. The Dulles runways count as open land. The shortlist goes to parcel, zoning, and ownership checks next.

The ranking holds up when the weights change. With equal weights of 0.25 on every factor, 112 cells score 0.85 or more under both weightings, and the same 41 cells keep full marks. The step scores of 1, 0.5, and 0 cause ties, which the fuzzy scoring of Jiang and Eastman (cited below) reduces.

A production screen replaces each stand-in with the authoritative layer. Parcels come from a File Geodatabase, and slope from GLO-30, the Copernicus global surface model with 30 m cells through RS_ZonalStats. The Site Selection use case shows the pattern, and an AI agent can run it through the Wherobots Model Context Protocol (MCP) Server. These joins add facts about physical places to geospatial AI models trained on text, documents, databases, and the internet.

Parcel-scale screen of 52 West Texas counties

The production version covered 52 counties in the Far West and West weather zones of the Electric Reliability Council of Texas (ERCOT), which runs most of the Texas grid. It ran as one Wherobots job run in 648 seconds on a Medium runtime. The unit is the parcel: 907,845 Regrid records reduce to 871,593 unique polygons. Setbacks remove land near water, pipelines, airports, schools, hospitals, and settled areas. A slope screen on GLO-30 removes land steeper than 5%. That leaves 153,910 parcels with at least 4 ha, and 30,690,000 buildable acres.

Source	Format	Size in the study area	What the job read from it
Regrid parcels	GeoParquet, one file per county	907,845 records	Parcel polygons and owners
Overture infrastructure	Havasu table	174 substations at 345 kV or more	Substations by voltage
Overture water	Havasu table	160 features classed as wastewater	Distance to a wastewater feature
Overture places	Havasu table	Schools, hospitals, and settled areas in the study box	300 m setbacks and distance from sensitive receptors, weighted 0.25 in the balanced score
Copernicus GLO-30	Elevation raster, 30 m cells	All 52 counties	Slope
ERCOT Transmission Project and Information Tracking (TPIT)	Excel workbook, July 2026	43 projects at 345 kV or more due by end of 2027	Planned in-service dates
ERCOT Generator Interconnection Status report	ERCOT report, September 2026	12,119 MW of signed generation due by end of 2027	Planned generation

The job read the parcels, the Overture tables, and the GLO-30 raster in WherobotsDB: raster and vector in one engine; most stacks pick one.

Distance to power uses an estimated line route, the straight-line distance times 1.32, the routing allowance in an independent balanced model built on Wherobots for Australia. The transmission cost estimation guide from the Midcontinent Independent System Operator (MISO) estimates line length the same way, adding 30% to the straight-line distance between substations. Within a 5 km estimated route, 5,570,879 buildable acres lie near substations of 138 kV or more. 1,192,666 acres, 3.9% of the buildable total, lie near 345 kV.

The balanced model puts 0.40 on distance to 138 kV or more and 0.25 on distance from sensitive receptors (schools, hospitals, and homes). It puts 0.20 on distance to a wastewater feature, and 0.15 on assemblage size (adjoining parcels with one owner). A power-first weighting puts 0.45 on 345 kV and 0.15 on 138 kV, and the other three factors share 0.40 in the same ratio. The two top 50 lists share 9 parcels. The balanced list spans 8 counties and the power-first list 21. Part of that shift is built in, since a heavier 345 kV weight lifts parcels near 345 kV. The Loudoun top set held under equal weights; the West Texas top 50 did not.

The wastewater layer needs a source check. The study box holds 160 Overture water features classed as wastewater, all unnamed OpenStreetMap polygons. In Reeves and Winkler counties they may be oilfield ponds. All 50 balanced top parcels sit within 10 km of one, and 32 sit in those two counties.

Planned power changes the ranking too. A timed score treats the end substations of planned lines and generators with signed agreements as power sources from their planned dates. At the end of 2027, it changes 9 of the power-first top 50. The dates are plans.

Site selection research and what changes at scale

Data center siting borrows from fifty years of land suitability research.

From map overlays to GIS-based MCDA

Ian McHarg's Design with Nature (1969) made the stacked map overlay the planner's tool. Collins, Steiner, and Rushman trace its path to GIS in Land-use suitability analysis in the United States (Environmental Management, 2001).

Steve Carver joined multi-criteria evaluation to GIS for siting in Integrating multi-criteria evaluation with geographical information systems (International Journal of Geographical Information Systems, 1991). Eastman, Jin, Kyem, and Toledano added constraint maps, factor maps, and weighted linear combination to raster GIS in Raster procedures for multi-criteria/multi-objective decisions (Photogrammetric Engineering and Remote Sensing, 1995). The same paper allocated land among competing objectives.

Jacek Malczewski wrote On the use of weighted linear combination method in GIS (Transactions in GIS, 2000). The paper found that analysts often applied the weighted overlay without a full understanding of its assumptions. Two of them shape the Loudoun screen. Every factor sits on the same 0 to 1 scale. Each weight sets how much of one factor can make up for a shortfall in another. His survey of GIS-based multicriteria decision analysis (International Journal of Geographical Information Science, 2006) classified more than 300 refereed articles published from 1990 to 2004.

Jiang and Eastman replaced hard thresholds with fuzzy membership functions in Application of fuzzy measures in multi-criteria evaluation in GIS (2000). A fuzzy membership function scores land on a smooth curve from 0 to 1, which removes the sudden score drop at each threshold. Ligmann-Zielinska and Jankowski mapped where suitability depends on the weights in Spatially-explicit integrated uncertainty and sensitivity analysis of criteria weights in multicriteria land suitability evaluation (Environmental Modelling and Software, 2014).

Data center siting papers

Covas, Silva, and Dias applied MCDA to sustainable data center location in Multicriteria decision analysis for sustainable data centers location (International Transactions in Operational Research, 2013).

Spare substation capacity data

No public file gives the number site selection needs most: spare capacity at each substation by year. Open data covers lines, substations, and planned in-service dates. Spare capacity comes from the utility.

Water stress and cooling

Siddik, Shehabi, and Marston mapped the footprint in The environmental footprint of data centers in the United States (Environmental Research Letters, 2021). One-fifth of the direct water footprint of U.S. data center servers comes from moderately to highly water-stressed watersheds. Nearly half of the servers draw power fully or partly from plants in water-stressed regions. Evaporative cooling uses less electricity but consumes water, so the cheapest site to power can be the costliest to cool.

What scale changes

A county screen fits on a laptop. The West Texas screen joined 871,593 parcels to substations, receptors, and an elevation model. Jia Yu, Jinxuan Wu, and Mohamed Sarwat introduced spatial joins across many machines on Apache Spark with GeoSpark (ACM SIGSPATIAL 2015). Yu, Zhang, and Sarwat described its spatial partitioning and indexing in Spatial data management in Apache Spark: the GeoSpark perspective and beyond (GeoInformatica, 2019). GeoSpark became Apache Sedona, the engine under WherobotsDB. It splits those joins across the machines of a cluster.

Conclusions

Both screens show that data center site selection starts with power. Constraints, weights, and source layers shape the shortlist.

1. Start site selection with power, by voltage and by year. In Loudoun, the 230 kV line factor was the hardest to meet. In West Texas, 3.9% of buildable land lies within an estimated 5 km line route of 345 kV. Planned power due by 2027 changed 9 of the power-first top 50, so score the in-service year as well as the map.
2. Keep hard constraints apart from scores in a site screen. The Loudoun query returned all 1,808 cells. It gave the 395 near rivers and lakes a null score and sorted them last. A shortlist needs a filter that drops them.
3. Test the weights before trusting a site ranking. The same 41 Loudoun cells kept full marks under both weightings. Two West Texas weightings shared 9 of 50 parcels. If the top sites change with the weights, set them with AHP and map the sensitivity.
4. Check each weighted layer's source. Building footprints penalize existing campuses. Rivers stand in for flood zones. All 50 balanced West Texas leaders sit within 10 km of an unchecked wastewater layer, 32 of them in two counties. Swap in authoritative layers before the shortlist.
5. Drop site criteria that do not vary across the study area. Latency barely changes inside one metro, so it adds nothing to a one-metro ranking.
6. Score water as its own site criterion. The cheapest site to power can be the costliest to cool. In dry regions, make water a hard constraint.

Related terms

• Isochrone: travel-time areas that can replace straight-line distance in a screen
• Digital elevation model: the slope and drainage input to site screening
• Flood risk data: the hazard layer that removes the most candidate sites
• H3: the hexagonal grid used as the unit of analysis in the Loudoun screen
• Building footprint: the developed-land layer behind the open-land score
• Overture Maps: the infrastructure, roads, buildings, and places layers for scoring
• Parcel data: the unit a site is bought and permitted as
• Points of interest (POI) data: nearby businesses and services for context scoring
• Spatial join: the operation behind every criterion in a suitability screen

Read more from Wherobots

• Learning Wherobots by building a national AI data center suitability report
• Site selection use case
• Making drive time isochrones
• Energy & utilities solutions

Screen land against power, building, road, and hazard layers in WherobotsDB with a Wherobots free trial at cloud.wherobots.com.