Data releases

Changelog

Every dataset version since first publication, including versions that predate frozen files.

1. 2026-10-03Method v4

   Corrected figures and primary sources for the mega projects held from the 2026-10-02 release, the SpaceXAI cluster counted once, and two projects added.

   • Project Tembo (Google, Laramie County): investment $50B to $17B, as Google presented the campus to Laramie County in September 2026 (Wyoming Tribune Eagle, 2026-09-10), and status under construction to announced, since site work starts late in 2026. Google has not restated a capacity, so the Crusoe-era 1.8 GW is now marked as a Malachi.Energy estimate, with low confidence.
   • Nexus Hubbard (Anthropic): status announced to under construction, and investment $15B to $16B, the full first-phase project finance package at final investment decision (W.Media, 2026-08-20). The $15B was the bank debt within it. The figure is financing, so it stays press reported.
   • Galaxy Digital Helios: capacity 1,630 MW to 800 MW and investment $1.4B to $1.7B, per Galaxy’s SEC 8-K of 2025-10-21. $1.7B is the Phase I build cost ($1.4B was the debt facility that funds it) and 800 MW is the approved power CoreWeave has committed to. The further 830 MW ERCOT approved in January 2026 has no tenant and is noted in the description.
   • AVAIO Digital Leo: capacity 1,070 MW to 1,000 MW and first capacity 2027 to 2028. The $21B full-development investment now cites the Arkansas Economic Development Commission release of 2026-01-12, which states both figures, instead of AVAIO’s own page, which states 1,070 MW but not the $21B.
   • Project Bunkhouse (Digital Realty): capacity 1,830 MW to 1,300 MW and investment $19B to $18B, per Bartow County’s August 2026 policy statement, replacing an undated aggregator source. Capacity basis is now a government filing, investment basis developer-stated campus cost, and confidence low to medium.
   • AWS New Carlisle: capacity 2,200 MW to 2,250 MW to match the source’s 2.25 GW. The $11B investment stays: AWS’s July 2026 statement that it has already invested $13.8B gives no capacity, so it is noted in the description.
   • SpaceXAI Colossus (Memphis) and MACROHARDRR (Southaven): both listed the full 2 GW cluster, so it was counted twice. Each now carries its own site: Colossus 2,000 MW to 1,500 MW (Colossus 1 about 500 MW plus Colossus 2 about 1 GW, Introl, 2026-01-03), with confidence medium to low because its $18B is the reported GPU cost, not construction; MACROHARDRR 2,000 MW to 250 MW of IT compute and status under construction to operational, per Measured AI (2026-06-25), which also states the $20B Mississippi investment and replaces an aggregator source.
   • Novva Borealis: the source moves from an aggregator to Novva’s 2024-08-21 announcement, which states the 300 MW and $3B. Figures unchanged.
   • Project Camellia (OpenAI): figures unchanged at 3,200 MW, the Georgia Power contract, and $20B. The description adds the 3,210 MW peak load from Georgia’s regional impact review and the September 2026 site plan postponement.
   • Added TeraWulf Justified (Hawesville, KY), leased to Anthropic: 482 MW under a Kentucky PSC-approved power agreement and $4.0B, the low end of TeraWulf’s $4.0B to $4.5B first-phase estimate (TeraWulf, 2026-08-24).
   • Added Kingston Technology Hub (Thor Equities, Bartow County, GA), proposed: 150 MW and $1B (Coosa Valley News, 2026-09-08), placed at the Kingston city centroid until a site is published.
   • Transmission proximity was measured for the two added projects. The OpenStreetMap layer was not refreshed and no coordinate moved. Methodology version 4.

2. 2026-10-02Method v4

   Meta El Paso investment restated to the developer-stated total development cost, and the Baxtel national and state counts refreshed.

   • Meta El Paso: investment $10B to $14B (developer-stated campus cost, buildings plus long-lived power, cooling, and connectivity infrastructure), per Meta’s 2026-07-28 release. Capacity stays at 1 GW.
   • Baxtel national count: 5,207 to 5,274 facilities and 793 to 804 operators, as listed on Baxtel’s United States page on 2026-10-02.
   • Baxtel state counts: Virginia 786 to 793 and Texas 774 to 779; California is unchanged at 312. State estimates for the other jurisdictions rescale so every estimate still sums to the national count.
   • The OpenStreetMap layer was not refreshed and no project was added, removed, or moved. Methodology version 4.

3. 2026-09-25Method v4

   Street addresses for 362 more OpenStreetMap facilities, from public-domain address points reviewed by hand, and an address_basis column that says where every facility address comes from. The OpenStreetMap layer and the mega projects are unchanged.

   • Added address_basis and address_source to facilities.csv. address_basis is osm, address-point, campus, or researched; address_source names the dataset release or URL behind every address that is not tagged in OpenStreetMap.
   • Added street addresses, cities, and postal codes for 362 facilities that OpenStreetMap maps without an address: 325 from an address point inside the building outline or within 10 meters of it, and 37 from the nearest address point within 50 meters. 356 come from the USDOT National Address Database and 6 from public-domain county sources (Alameda, Santa Clara, Alachua, and Duval counties), all through Overture Maps release 2026-09-23.0. Facilities with a street address rose from 899 to 1,261 of 1,771.
   • Every candidate was screened, then reviewed by hand. Tested on the 855 facilities OpenStreetMap already addresses, with their tags hidden, screened candidates matched the tagged house number and street 90.6% of the time for in-outline points and 89.2% for points within 50 meters. Most mismatches were the building’s own number on the right street where the tag carries a campus mailing address. Candidates that a second address point, a PeeringDB listing, or the facility name contradicted were held back.
   • Cities from address points were checked against tagged facilities and other address points nearby, because some county records carry a neighborhood, an airport district, a county, the owner’s city, a placeholder such as “Incorporated”, or no city at all instead of the mailing city. 22 facilities had their city, and one its postal code, corrected by hand.
   • 510 facilities still have no street address and are listed with their coordinates only.
   • The OpenStreetMap layer was not refreshed: names, operators, coordinates, and counts are those of the 2026-09-14 extract. Mega project figures are unchanged, and their last_verified_at stays 2026-09-16, the last release in which every project was re-checked.
   • Methodology version 4.

4. 2026-09-16Method v3

   Location basis for every mega project, re-researched site coordinates where a public source identifies the campus, and the first derived spatial facts: distance from each site-located project to the nearest 230 kV or higher transmission line.

   • Added location_basis to the mega project exports: site (campus footprint identified), address (public address, geocoded), city-centroid, or county-centroid. The JSON export also carries locationSource where the location comes from a different public source than the figures.
   • Re-researched campus coordinates. Projects upgraded to site or address carry a new coordinate and a dated location source; centroid projects keep their previous coordinate.
   • Added transmission proximity, derived in PostGIS from the HIFLD Electric Power Transmission Lines layer (data current to 2023-09-05) filtered to 230 kV and above: nearest line distance, voltage, and owner per project, and the count of site-located projects within 5 miles. The derived values live in derived-facts.json with their own measurement date and are not part of the frozen exports in this release.
   • Only site- and address-located projects count toward any distance fact; centroid coordinates can sit several miles from the parcel, so the prose states the site-located denominator and excludes the rest.
   • Methodology version 3.

5. 2026-09-14Method v2

   Full OSM facility refresh, every mega project re-researched against a dated public source, per-figure basis and confidence columns, and state estimates reconciled to the national count.

   • Refreshed the OpenStreetMap facility layer through the Overpass API and reassigned every facility to a jurisdiction by point-in-polygon against Census boundaries.
   • Re-researched all tracked mega projects. Each now carries a public capacity figure, a public investment figure, a source (label, URL, date), a capacity basis, an investment basis, and a confidence rating. Projects with no public figure for either were retired rather than estimated.
   • Added the columns capacity_basis, investment_basis, is_estimate, confidence, included_in_aggregate, comparable_campus_investment, source_label, source_url, source_date, last_verified_at, and url to the mega project exports.
   • State estimates now reconcile to the Baxtel national count exactly (methodology version 2). This derivation landed on 2026-09-15, after the data refresh date, so estimated_total values in this release differ from the site as served on 2026-09-14 while osm_verified values do not.
   • Coverage is 46 states, the District of Columbia, and Puerto Rico.
   • First frozen release. Earlier versions were served only at the moving URLs and cannot be reproduced byte for byte.

6. 2026-03-11Method v1 · no frozen files

   Last data change before the 2026-09-14 refresh, as recorded in DATASET_UPDATED at the time. Not itemized.

7. 2026-03-09Method v1 · no frozen files

   First publication of the US Data Center Directory.

Methodology versions

The method version changes when the way a number is derived changes, independent of the data refresh date. Compare it before comparing estimates across releases.

Version 4since 2026-09-25

Every OpenStreetMap facility carries an address basis: osm when the address is tagged in OpenStreetMap, address-point when it comes from a public-domain address point inside or beside the building outline (the USDOT National Address Database and public-domain county sources, via Overture Maps), campus when it is the address of an adjacent building on the same campus, and researched when it comes from an operator page or public record. Only public-domain, CC0, or PDDL address points are used, so they can enter the ODbL layer. No address is published without a hand review, and an OpenStreetMap tag always takes precedence. The mega project last_verified_at column now dates the last release in which every project was re-checked, instead of the dataset version.

Version 3since 2026-09-16

Every mega project carries a location basis (site, address, city centroid, or county centroid). Transmission proximity is derived as the geodesic distance from each project coordinate to the nearest HIFLD transmission line of 230 kV or higher, computed in PostGIS on geography and committed as JSON at refresh time; only site- and address-located projects count toward the within-5-miles fact, and every derived sentence names the layer vintage and its denominator.

Version 2since 2026-09-15

State estimates reconcile to the Baxtel national count exactly: Baxtel state counts where published, the remainder allocated in proportion to OSM-verified count with a derived multiplier (largest-remainder rounding). Every mega project carries a capacity basis, an investment basis, a confidence rating, and a dated public source.

Version 1since 2026-03-09

State estimates were OSM-verified counts times a fixed multiplier (2.9 at the last version 1 refresh). Mega project figures were compiled without a per-figure basis or confidence rating.

Full method, reconciliation table, and investment split on the hub methodology block.