Download all 2,650 U.S. data-center facilities

Free facility data

Free to download, with per-field reuse terms. DataCenterData fields are licensed CC BY 4.0: credit DataCenterData and link to datacenterdata.com/data/. Fields that come from Compute Atlas keep Compute Atlas’s CC BY 4.0 license and its credit. U.S. government work fields carry no copyright. Values the facility register marks reference only are left out of the files, and each row’s withheld-field list names the fields they belong to; in a list such as operators only the withheld entries are left out, so other entries can remain. Exact column names are in the data dictionary.

Every facility in the register — 2,650 register records — in three formats with the same rows:

• datacenters.csv — spreadsheet, one row per facility (2.4 MB)
• datacenters.geojson — map layer; facilities without a published coordinate have no geometry (5.6 MB)
• datacenters.json — full records with sources, statuses by layer and stated capacity (5.4 MB)
• datacenters-dictionary.json — data dictionary: every field, its reuse terms, attribution, counts and these checksums

Suggested citation: DataCenterData. U.S. data center facility register, free download (data as of 2026-10-18). https://datacenterdata.com/data/. DataCenterData fields CC BY 4.0; Compute Atlas fields CC BY 4.0 (Compute Atlas by Edward Kubiak).

Which fields are CC BY

• DataCenterData fields, CC BY 4.0, every row: dcd_id, public_id, site_class, site_class_label, evidence_level, evidence_tier, record_unit, county_fips, cbsa_code, cbsa_name, coordinate_precision, coordinate_source, location_precision, source_families, first_listed, latest_source_date, field_origins, each row’s Compute Atlas field list and withheld-field list, plus any other field whose field_origins value is dcd_original.
• Compute Atlas fields, CC BY 4.0, credit Compute Atlas: the fields named in each row’s Compute Atlas field list (1,728 rows).
• U.S. government works: fields whose field_origins value is public_record (U.S. EPA Facility Registry Service fields and Census Geocoder points).
• Links only: source_urls and site_class_sources; the linked documents keep their own rights.

Rows carrying each source-dependent field, by origin, and rows where it is withheld:

Labels and evidence

• Crypto-mining and AI compute campuses confirmed from permitted primary sources are included and labelled in site_class, so you can keep or drop them; none is labelled yet, so every row is unclassified (not evidence either way).
• evidence_level gives the register’s evidence tier a plain name: dcd_reviewed (a DataCenterData reviewed record in the verified inventory layer), multiple_source_families (two or more source families joined; their independence is not tested), single-family records and third_party_listing_only discovery listings, so a lead is not mistaken for a reviewed site.
• county_fips is given only where the exported county matches the pinned Census files (1,504 rows). Coordinates appear only where the facility page publishes one (559 rows).

What is not included

• Every value the facility register marks reference only: the operator-listed layer (names, addresses, stated operators, statuses and figures captured from operators’ pages) and DataCenterData reviewed-record fields that have not been cleared field by field. Each row’s withheld-field list names the fields they belong to (in a list such as operators, only those entries are left out); the facility page still shows them with their terms. One exception is visible: dcd_url is the address of the facility’s public DataCenterData page, built from its published page title, so for a row whose name is withheld that address can still contain a form of the name. It is a link, not a licensed value.
• Verbatim quotations from operator pages and other sources (held in capture records; never republished).
• Linked permits, filings, operator pages and other documents (linked, not republished; they keep their own rights).
• Selected factual fields cited to state and local agency records (shown on facility pages with their citation; not relicensed).
• From a Compute Atlas row that cites OpenStreetMap: its coordinate, street address and postal code. Other Compute Atlas values (for example names, cities and capacity text) pass through under Compute Atlas’s CC BY 4.0 terms; DataCenterData has not traced whether Compute Atlas drew any of them from OpenStreetMap or its other inputs. No value from DataCenterData’s own private OpenStreetMap or PNNL/IM3 lead lists is included.
• Coordinates the site itself does not publish (below street precision or withheld after review).
• Contact routes, private captures and internal review notes.

Compute Atlas credit

Values marked Compute Atlas (listed in each row’s Compute Atlas field list) carry this credit: “Compute-infrastructure data from Compute Atlas by Edward Kubiak”, licensed CC BY 4.0. Changes made: Filtered to data-center records in the 50 U.S. states; selected fields normalized for search and display; DataCenterData directory IDs added. In this download: values organized by DataCenterData facility id; coordinates, street addresses and postal codes from rows that cite OpenStreetMap removed. Keep this credit when you reuse those values.

Dates and updates

Data as of 18 October 2026: the newest input date of the facility register, not the build time and not a fresh check of every facility. The files are regenerated with every site release; changes are listed on Changes and in the Atom feed. Counts are register records (campuses, buildings, suites and unresolved presences), not unique sites and not a census.