Euphemios
Back to insights

Report

Where AI data centers and crypto mining sit in the U.S.

Public records locate known U.S. computing sites and show where construction, electricity sales and commercial real estate balances are changing.

Euphemios10 min read

Executive summary

The stated operating rosters locate 124 known U.S. sites across AI or cloud computing, commercial colocation and crypto mining. The available rosters yield no state with a defensible data-centers-per-mining-site ratio. Virginia is part of the Richmond Fed's Fifth District, which had 8 data-center occurrences across three Beige Book editions. Within that district, Virginia held 392 of 450 mapped features in the separate OpenStreetMap context layer.

The Census estimate of private data-center construction reached $96.46 billion in September 2026, or 66.2% of private office construction. Public bank filings give the headquarters of lenders that carried commercial real estate balances. The records reviewed do not identify loan destinations or the amount outstanding for data centers.

Key findings

  • The operating site roster includes 56 AI or cloud, 34 colocation and 34 mining facilities from named sources. The state map uses all eligible sites. Its pins show mining sites with verified positions.
  • Virginia accounted for 392 of 450 mapped features in the Richmond Fed footprint. The feature counts measure mapping coverage and can include more than one feature per site.
  • Census puts the national data-center construction pace at $96.46 billion in September 2026, up from $50.70 billion a year earlier. The available annual figures show $27.18 billion in the South in 2025, the largest regional total.
  • Retail electricity sales rose 6.9% in Virginia, 5.5% in Texas and 2.1% in Georgia from 2024 to 2025. These figures include retail customers across the states.
  • Atlanta recorded 11 data-center occurrences, Dallas 10, and Richmond 8 across the three reviewed editions. For literal AI language, their respective totals were 11, 8, and 6. These are occurrences in the district chapters.
  • In 2026 Q2, banks headquartered in Texas reported $35.64 billion in construction and land development loans outstanding. Banks headquartered in Virginia reported $12.75 billion, and Georgia-headquartered banks reported $7.11 billion.

Where known facilities sit

The map counts a physical facility once when the sources establish its primary use. In Texas the eligible records contain 9 AI or cloud, 0 colocation and 7 mining sites. Virginia had 7 AI or cloud, 13 colocation and 0 mining sites. Zero means the sources yielded no eligible record.

Known operating facilities by state

Use the arrow keys to explore the chart’s regions in alphabetical order.

Loading map boundaries…

Total known operating facilities
020
  • Insufficient known facilities
  • AI or cloud
  • Colocation
  • Crypto mining

Federal Reserve districts, literal AI across three editions

  • BostonJuly 15: 1, September 2: 0, October 14: 0
  • New YorkJuly 15: 1, September 2: 4, October 14: 2
  • PhiladelphiaJuly 15: 1, September 2: 0, October 14: 1
  • ClevelandJuly 15: 1, September 2: 7, October 14: 2
  • RichmondJuly 15: 1, September 2: 5, October 14: 0
  • AtlantaJuly 15: 5, September 2: 3, October 14: 3
  • ChicagoJuly 15: 0, September 2: 0, October 14: 1
  • St. LouisJuly 15: 0, September 2: 1, October 14: 1
  • MinneapolisJuly 15: 2, September 2: 0, October 14: 1
  • Kansas CityJuly 15: 1, September 2: 0, October 14: 0
  • DallasJuly 15: 1, September 2: 1, October 14: 6
  • San FranciscoJuly 15: 3, September 2: 3, October 14: 1
Map data tables
Alabama01200MissingInsufficient
Alaska02000MissingInsufficient known facilities
Arizona04300MissingInsufficient
Arkansas05000MissingInsufficient known facilities
California06011011OK
Colorado08000MissingInsufficient known facilities
Connecticut09000MissingInsufficient known facilities
Delaware10000MissingInsufficient known facilities
District of Columbia11000MissingInsufficient known facilities
Florida12000MissingInsufficient known facilities
Georgia1330710OK
Hawaii15000MissingInsufficient known facilities
Idaho16100MissingInsufficient
Illinois17030MissingInsufficient
Indiana18200MissingInsufficient
Iowa19200MissingInsufficient
Kansas20000MissingInsufficient known facilities
Kentucky21002MissingInsufficient
Louisiana22000MissingInsufficient known facilities
Maine23000MissingInsufficient known facilities
Maryland24000MissingInsufficient known facilities
Massachusetts25000MissingInsufficient known facilities
Michigan26000MissingInsufficient known facilities
Minnesota27100MissingInsufficient
Mississippi28203MissingInsufficient
Missouri29000MissingInsufficient known facilities
Montana30000MissingInsufficient known facilities
Nebraska31401MissingInsufficient
Nevada32100MissingInsufficient
New Hampshire33000MissingInsufficient known facilities
New Jersey34160MissingInsufficient
New Mexico35100MissingInsufficient
New York36001MissingInsufficient
North Carolina37100MissingInsufficient
North Dakota38100MissingInsufficient
Ohio39402MissingInsufficient
Oklahoma40100MissingInsufficient
Oregon41200MissingInsufficient
Pennsylvania42103MissingInsufficient
Rhode Island44000MissingInsufficient known facilities
South Carolina45100MissingInsufficient
South Dakota46000MissingInsufficient known facilities
Tennessee47305MissingInsufficient
Texas4890716OK
Utah49100MissingInsufficient
Vermont50000MissingInsufficient known facilities
Virginia51713020OK
Washington53011MissingInsufficient
West Virginia54000MissingInsufficient known facilities
Wisconsin55100MissingInsufficient
Wyoming56102MissingInsufficient
Kentucky Facility owned siteCrypto miningRiot Platforms2114537.0894-88.70288Source for Kentucky Facility owned site2026-03-02
Corsicana FacilityCrypto miningRiot Platforms4834931.96935-96.51472Source for Corsicana Facility2026-03-02
Rockdale FacilityCrypto miningRiot Platforms4833130.57323-97.0801Source for Rockdale Facility2026-03-02

Known facilities only, not a complete inventory.

Census cartographic boundaries, 2024, 1:20m

© OpenStreetMap contributors · Data available under the Open Database License (ODbL). Share-alike centroid extract.

Source · SEC EDGAR annual reports; Euphemios analysis · Latest original annual filings through November 20, 2026 for the stated issuer list · As of

Source · Epoch AI data on AI data centers; Euphemios analysis · November 18 U.S. records and operating dates, with separately verified geographic precision · As of

Source · Equinix colocation availability; Euphemios analysis · Named U.S. sites matched to specific OpenStreetMap features · As of

Source · OpenStreetMap contributors, via the PNNL IM3 Open Source Data Center Atlas (v2026.11.02); Euphemios analysis · Published centroid layer v2026.11.02, spatially joined to 2024 Census counties · As of

Source · U.S. Census Bureau, cartographic boundary files; Euphemios analysis · County equivalents and FIPS from the 2024 cartographic boundary vintage · As of

State totals include sites located only to the state. The methods give the leading county counts.

The November 2 PNNL atlas yields 1,846 mapped features across 294 county equivalents. The atlas can include several buildings on one campus. Census County Business Patterns counts employer establishments in data processing, hosting and related services. Both measures provide context outside the map's facility totals. PNNL atlas · Census County Business Patterns

What we cannot measure yet. The roster misses undisclosed facilities, individual sites hidden within portfolios and facilities outside the source lists. Within an area, the ratio formula divides the combined AI or cloud and colocation count by the mining count. A displayed figure would describe those selected records. Texas has zero colocation sites in this roster because the Texas listings had no unique atlas feature with a matching facility code, and a fuller roster would need more operators to publish site names, locations, and operating status that can be independently checked. Comparing the broader industries needs a common and more complete observation of each category.

What capacity means

Megawatts measure power. Megawatt-hours measure energy over time. Adding a lease's critical IT load to the campus's developed power could double-count capacity.

Filed exampleOperatorCapacityDefinitionStage
Salt CreekHut 863 MWTotal site capacity for the filed ASIC compute siteASIC compute
CorsicanaRiot Platforms400 MWDeveloped capacity at the siteMining operating and further development planned
River BendHut 8330 MWTotal utility capacity for the projectAI infrastructure under construction
Rockdale AI leaseRiot Platforms25 MWInitial critical IT load under the leaseContracted initial delivery

Critical IT load serves computing equipment. Utility power can also serve cooling and other systems. Developed or energized capacity describes a physical stage. Contracted capacity describes an agreement and may be delivered later.

For historical context, the U.S. Energy Information Administration's February 2024 analysis identified 137 U.S. crypto mining facilities. It mapped 52 with both location and capacity information and estimated maximum capacity of 10,275 MW across the 101 with that measure. That inventory predates the filings used here. EIA historical analysis

What we cannot measure yet. These examples leave actual energy consumption and the nationwide split of operating megawatts by workload unresolved. They also leave the completion dates of planned power additions uncertain.

Where construction and electricity sales are changing

The Census Bureau's preliminary September 2026 estimate puts private data-center construction at $96.46 billion at a seasonally adjusted annual rate, compared with $50.70 billion in September 2025. That is a 90.3% rise on the same basis. The September 2026 private office total, which includes data centers, was $145.63 billion at an annual pace. Data centers represented 66.2% of that total, against 48.6% in September 2025. Census monthly private construction

U.S. private data-center construction

Construction spending (USD billions at annual rate)
Use the arrow keys to explore chart values.
Month
U.S. private data-center construction (USD billions at annual rate)
MonthPrivate data centers
2025-09-0150.7
2025-10-0154.26
2025-11-0153.04
2025-12-0155.46
2026-01-0156.21
2026-02-0157.36
2026-03-0158.12
2026-04-0161.86
2026-05-0165.69
2026-06-0172.7
2026-07-0180.72
2026-08-0188.82
2026-09-0196.46

Nominal dollars at seasonally adjusted annual rates, latest 13 months. The underlying workbook gives earlier observations.

Source · U.S. Census Bureau, Construction Spending; Euphemios analysis · Monthly private office and data-center construction at seasonally adjusted annual rates · As of

In calendar 2025, private data-center spending was $49.74 billion nationwide, or 47.7% of private office spending for the full year. The South accounted for $27.18 billion, followed by the Midwest at $13.12 billion, the West at $8.92 billion, and the Northeast at $0.52 billion. The regional series measures work put in place during the calendar year. Census regional construction workbook

Retail sales grew faster in Virginia and Texas than in Georgia. The statewide figures cover residential, commercial, industrial and transportation customers. Weather, population, pricing and many industries can affect sales. EIA historical state sales workbook

Statewide retail electricity sales changes

State or district
Virginia
Texas
Georgia
2025 versus 2024 (percent)
Use the arrow keys to explore chart values.
Statewide retail electricity sales changes (percent)
State or districtChange in retail sales
Virginia6.91
Texas5.51
Georgia2.11

All retail customers in the three comparison states. The methods give all states and D.C.

Source · U.S. Energy Information Administration; Euphemios analysis · Annual statewide retail electricity sales, all sectors, full-year 2024 and 2025 · As of

What we cannot measure yet. The Census files do not publish data-center construction dollars by state or county. Electricity sales do not identify specific operators or allocate demand to data centers, crypto miners or other customers. They also leave total consumption behind the meter outside the retail measure.

What the Federal Reserve districts mention

Atlanta led the three-edition literal AI count with 11 occurrences. Cleveland had 10, Dallas 8 and Richmond 6. In Richmond's October chapter there were 4 data-center occurrences and 0 literal AI occurrences.

AI and data-center occurrences by Federal Reserve district

  • Literal AI
  • Data centers
Federal Reserve district
Boston
New York
Philadelphia
Cleveland
Richmond
Atlanta
Chicago
St. Louis
Minneapolis
Kansas City
Dallas
San Francisco
Reviewed occurrences
Use the arrow keys to explore chart values.
AI and data-center occurrences by Federal Reserve district (occurrences)
Federal Reserve districtLiteral AIData centers
Boston11
New York73
Philadelphia25
Cleveland109
Richmond68
Atlanta1111
Chicago16
St. Louis25
Minneapolis34
Kansas City15
Dallas810
San Francisco72

Totals for July 15, September 2 and October 14, 2026, excluding the separate national summaries.

Source · Board of Governors of the Federal Reserve System, Beige Book; Euphemios analysis · Reviewed district and national chapters published July 15, September 2 and October 14, 2026 · As of

Richmond. Virginia is in the Fifth District, which also covers Maryland, North Carolina, South Carolina, the District of Columbia and most of West Virginia. Virginia accounted for 87% of the district's 450 mapped features. The Virginia Joint Legislative Audit and Review Commission separately estimated about 150 sites and 340 buildings in the state in 2024. It put Northern Virginia at 13% of reported global operational capacity in that vintage. Richmond's September chapter describes moratorium concerns in the Washington metropolitan area. Its October chapter discusses building inputs, ports and manufacturers serving data-center customers across the district. District footprint · Virginia study · Richmond · Richmond

Atlanta. The Sixth District covers Alabama, Florida, Georgia and defined portions of Louisiana, Mississippi and Tennessee. The atlas contains 120 mapped features within that footprint. The October district chapter reports that construction and land development lending expanded amid data-center buildouts. It also describes construction activity and demand for freight serving those projects. The contacts did not identify lenders, borrowers or amounts. District footprint · Atlanta

Dallas. The Eleventh District covers Texas, northern Louisiana and southern New Mexico. Its footprint had 225 mapped features, including 220 in Texas. The October Dallas chapter described a Texas permit pause and reported staffing and building-cost pressures. District footprint · Dallas

What we cannot measure yet. Beige Book chapters vary in length, contacts and editorial emphasis. An occurrence is a measure of text coverage. The counts do not establish the prevalence of AI use, the size of local projects, or a causal link between facility density and how districts describe the economy.

Where lenders are headquartered

BankFind reports the loans outstanding on each institution's balance sheet. In 2026 Q2, Texas-headquartered institutions carried $99.20 billion in loans secured by nonfarm nonresidential properties. The corresponding balances were $63.60 billion for institutions headquartered in Virginia and $23.49 billion for those headquartered in Georgia. Construction and land development can include housing. Nonfarm nonresidential covers properties across industries. BankFind fields and definitions

Commercial real estate balances by bank headquarters

  • Construction and land development
  • Nonfarm nonresidential
Bank headquarters state
Virginia
Georgia
Texas
June 2026 balance (USD billions outstanding)
Use the arrow keys to explore chart values.
Commercial real estate balances by bank headquarters (USD billions outstanding)
Bank headquarters stateConstruction and land developmentNonfarm nonresidential
Virginia12.7563.6
Georgia7.1123.49
Texas35.6499.2

The two categories cover all borrowers in scope. The methods give four quarters.

Source · Federal Deposit Insurance Corporation, BankFind Suite; Euphemios analysis · Bank quarter balances and assets summed by headquarters state · As of

The Federal Reserve's H.8 shows a national commercial real estate loan balance of $3.16 trillion for the week ending 2026-11-11, seasonally adjusted. The national series has its own coverage and definitions, so it serves as a credit benchmark separate from these headquarters totals. The Census office share measures data-center spending as a fraction of private office construction, a flow. BankFind and H.8 measure credit outstanding, including loans from earlier periods. Federal Reserve weekly release

What we cannot measure yet. No public source reviewed provides comparable data-center loan amounts or the location of the underlying loans. The headquarters state identifies the reporting bank. It cannot show where a project is built, where a borrower operates, or who financed a particular data center. The construction share therefore cannot estimate a data-center share of bank lending.

Methodology and definitions

Select each stated issuer's latest original annual filing through November 20, 2026. Read the primary EDGAR document and separate current operations from later phases. Count one site or campus, using the stated primary use and the operating date in its evidence.

Use verified county or state evidence for area counts and verified coordinates for mining pins. The map shades states with at least 10 known sites. A ratio needs at least 3 mining facilities and comparable coverage of both data-center categories.

The separate feature layer uses the published PNNL IM3 Open Source Data Center Atlas centroids. Credit OpenStreetMap contributors, via the PNNL IM3 Open Source Data Center Atlas (v2026.11.02); Euphemios analysis (© OpenStreetMap contributors). The share-alike centroid extract and the accompanying county-joined extract osm-features-odbl.csv are available under the Open Database License. Credit the AI data centers dataset to Epoch AI under Creative Commons Attribution. Geographic joins and roster classification are adaptations. Dataset collector credits.

Geofabrik's November 20 Virginia extract had 432 distinct OSM objects matching the data-center tags, including 20 tagged only as proposed or under construction, compared with 392 centroids in PNNL's November 2 release. The raw count retains separate mapped objects, while PNNL removes points and campus polygons that overlap mapped buildings before making centroids from an earlier OpenStreetMap snapshot. Geofabrik Virginia extracts

Calculate Census shares from same-period office and data-center spending. EIA sales cover all retail customers statewide. Count Federal Reserve terms by district and edition, and group BankFind loans by reporting bank headquarters.

Full methods and sources →

Limits

Filings vary in property detail and can describe status at fiscal year-end or in later updates. Some private facilities lack a public record at the needed precision. Acquisitions and leases can change the recorded operator.

Census regions are construction geographies. Federal Reserve districts are contact and publication geographies. States are used for utility sales and bank headquarters. An association across those geographies cannot establish ownership, use, financing or causality.

FAQ

What does the map count?

It counts a named U.S. physical facility or campus with evidence for its primary category and operating status. Its category counts come from the stated company and dataset selections, with geography verified to the level shown. Proposed and mixed phases can appear in the evidence without adding to the map.

What does an unshaded place mean?

The selection yielded too few known sites for shading, or the area had no classified observation. A state with no recorded facility can still contain facilities.

Why is the ratio sometimes missing?

The map suppresses ratios when there are too few known mining sites or too little comparable operating evidence for the other categories. A ratio would compare observed site counts. It would leave differences in power, floor area and computing capacity unresolved.

Why compare Virginia with Richmond?

Virginia is in the Richmond Fed's Fifth District. Its mapped feature concentration gives geographic context for Richmond chapter coverage. The district also contains several other places, so a district-wide contact report can describe activity outside Virginia.

Can the electricity chart show how much power the facilities used?

It shows annual retail electricity sold statewide, combining customer sectors. A site-specific answer would require measured deliveries or consumption for the named operators and a comparable time window.

Can the bank chart show who funded a data center?

It shows balances assigned to bank headquarters states. The public figures reviewed contain no loan-level facility identifier, borrower-level allocation or location for the loans underlying those balances.

Company and network names are trademarks of their owners. They are mentioned as facts about public filings and data, not as a relationship or an endorsement.

Sources

  • SEC EDGAR annual reports; Euphemios analysis
    Run Latest original annual filings through November 20, 2026 for the stated issuer list
    Data as of
  • Epoch AI data on AI data centers; Euphemios analysis
    Run November 18 U.S. records and operating dates, with separately verified geographic precision
    Data as of
  • Equinix colocation availability; Euphemios analysis
    Run Named U.S. sites matched to specific OpenStreetMap features
    Data as of
  • OpenStreetMap contributors, via the PNNL IM3 Open Source Data Center Atlas (v2026.11.02); Euphemios analysis
    Run Published centroid layer v2026.11.02, spatially joined to 2024 Census counties
    Data as of
  • U.S. Census Bureau, cartographic boundary files; Euphemios analysis
    Run County equivalents and FIPS from the 2024 cartographic boundary vintage
    Data as of
  • U.S. Census Bureau, Construction Spending; Euphemios analysis
    Run Monthly private office and data-center construction at seasonally adjusted annual rates
    Data as of
  • U.S. Census Bureau, Construction Spending by Region; Euphemios analysis
    Run Annual private office and data-center construction across four Census regions
    Data as of
  • U.S. Energy Information Administration; Euphemios analysis
    Run Annual statewide retail electricity sales, all sectors, full-year 2024 and 2025
    Data as of
  • Board of Governors of the Federal Reserve System, Beige Book; Euphemios analysis
    Run Reviewed district and national chapters published July 15, September 2 and October 14, 2026
    Data as of
  • Federal Deposit Insurance Corporation, BankFind Suite; Euphemios analysis
    Run Bank quarter balances and assets summed by headquarters state
    Data as of
  • Board of Governors of the Federal Reserve System, H.8; Euphemios analysis
    Run Seasonally adjusted commercial real estate loans at all commercial banks, November 20 release
    Data as of
  • U.S. Energy Information Administration
    Run Tracking electricity consumption from U.S. cryptocurrency mining operations, February 1, 2024
    Data as of
  • Virginia Joint Legislative Audit and Review Commission
    Run Data Centers in Virginia, chapter on industry concentration, December 9, 2024
    Data as of
  • U.S. Census Bureau, County Business Patterns; Euphemios analysis
    Run County employer establishments in Data Processing, Hosting, and Related Services, 2023
    Data as of
  • OpenStreetMap contributors, via Geofabrik; Euphemios analysis
    Run Virginia OSM regional extract, raw data-center tag count
    Data as of