# Eco TerraVault intelligence atlas: sources and limits

Reviewed September 27, 2026. Two independent source layers describe different things. **Their counts must not be added together.** This atlas provides research context, not a complete census, live operating-status service or verified inventory of Eco TerraVault projects.

## Source-reported facility records

Source: **Compute Atlas by Edward Kubiak**, version `v1.34.0`, published September 21, 2026. [Pinned dataset](https://raw.githubusercontent.com/ek33450505/compute-atlas/c8ead6584b0553ad09aa1ac769f12456e29b4f5b/data/facilities.json) · [source methodology](https://github.com/ek33450505/compute-atlas/blob/c8ead6584b0553ad09aa1ac769f12456e29b4f5b/docs/methodology.md) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

Eco TerraVault filters the 2,025-record release to its 1,605 records marked `data_center`, excluding 188 crypto-mining records and 232 power-generation records. It preserves source IDs, names, operators, city/jurisdiction, reported status, confidence, update dates and citations. A record may refer to a campus, an individual site, or grouped locations, not one building. Campus-level and individual-site records can overlap; these are source records, not a deduplicated count of unique physical facilities. The current status labels are source-reported: operational 747, proposed 424, under construction 238, permitted 108 and cancelled 88. They have not all been independently checked by Eco TerraVault. Eighteen records have the source confidence `rumored`; confidence is displayed rather than silently upgraded. Unknown facts remain unknown.

A cancelled proposal is not a decommissioned operating facility. This source does not supply dormant or decommissioned categories. Status histories are incomplete and can concern expansions; this release does not derive national trends or original opening dates from them. Neither `lastUpdated` nor a source retrieval date is an event date. Record counts are not power capacity, water consumption or environmental scores. Optional technical quantities are excluded from this public derivative.

The catalog has records in 55 of the 56 displayed jurisdictions. American Samoa has none in this snapshot; this does not prove there are no data centers there. Source coverage is partial and favors publicly documented facilities. Current live-source counts can differ because this atlas uses a reproducible frozen release.

Source SHA-256: `deca4eea8d71e904681f4a1d9ae2c64103157cdd69ee1462c4912e5ea74f3b7f`.

## Physical setting: reviewed evidence

Eco TerraVault review date: **2026-09-27**. The physical-setting evidence register addresses **9 selected catalog records**, of which **8 are classified**. The complete 1,605-record catalog contains 4 above-ground, 4 underground, 0 mixed and 1597 unclassified records. This small, purposefully selected review is not representative coverage and does not establish national setting totals or market shares.

- **Above ground**: a reviewed primary source supports above-ground data halls for the identified record and stated scope. A multistory building may still be above ground.
- **Underground**: a reviewed primary source supports data halls below ground or inside a subterranean facility for the identified record and stated scope.
- **Mixed**: primary evidence explicitly establishes both above-ground and underground data halls within the same catalog record's scope. Conflicting or ambiguous descriptions do not establish mixed use.
- **Unclassified**: available reviewed evidence does not establish a physical setting, or the record has not received a setting review. Missing evidence never defaults to above ground.

The register `scripts/assets/intelligence-setting-evidence.json` is bound to the pinned catalog's SHA-256 and joins by exact catalog ID and jurisdiction. A new source snapshot requires explicit review of record identities and scope before the evidence register can be reused. The offline build rejects duplicate or unknown IDs, jurisdiction mismatches, unsupported categories, incomplete review metadata and missing or malformed citations. It copies only the explicitly reviewed category; it does not scan facility names or source text for keywords. References to underground utility feeds, fiber, water pipes, cooling loops or power generation do not establish underground data halls. Each classified record carries its review date, summary, limitations, evidence scope and primary-source links. Unreviewed records carry a null setting-review date and no setting citations.

Setting is independent of the inherited Compute Atlas lifecycle. A source may describe a proposed building or one part of a campus; the evidence scope and limitations must remain visible. The review does not update lifecycle status, prove current operation, certify a design, establish a national underground inventory, or infer environmental performance. Records and campus-level overlaps remain counted as source records. National and jurisdiction setting totals and setting-by-lifecycle tables reconcile exactly to their existing catalog totals; building and event layers are unchanged.

Evidence register SHA-256: `8ec99401eb846d5770b0784930f1bba1a70694539e086503db69226a070930e9`. Changes to this editorial register require a new evidence review and deterministic regeneration alongside the original source snapshots.

## Mapped building records

Source: **PNNL-led IM3 Open Source Data Center Atlas**, February 9, 2026 release. [Publication and DOI](https://doi.org/10.57931/3017294). The retained GeoPackage is from authors' repository commit `74ab37d5b9d200400a01639f9ffc3c3a8b716314`.

The building layer has 1,239 rows. Two county-boundary duplicate rows are removed by exact building ID, yielding **1,237 mapped building records** across the 56-jurisdiction display area. This comprises the existing 1,236 records in the 50 states and D.C. plus one Puerto Rico record (`00365941471`). Other territory zeros mean no records in this layer. Campus and point layers are excluded. We do not infer operational status, ownership, capacity, event history or site suitability from a mapped building.

Derived building data is available under the [Open Database License 1.0](https://opendatacommons.org/licenses/odbl/1-0/). Contains information from the IM3 atlas derived from © OpenStreetMap contributors. The complete derived building-ID table, source notices, license and aggregation code accompany this release. The facility compilation's CC BY license does not replace this layer's ODbL terms. Jurisdiction downloads identify both layers and licenses separately.

Source SHA-256: `1c0d8c206eb2070785e594784fda90f615e6ed7fd9646d67e1a9de237b8cc9f4`.

## Downloads and reproducibility

- The full normalized facility catalog is held in `lib/server/intelligence-facilities.json`, outside the public download directory. The website retrieves bounded pages of at most 24 records through `/api/intelligence/catalog`, using jurisdiction, reported lifecycle, reviewed physical setting, text search and offset filters. Every result retains source citations and snapshot/license metadata. No street coordinates, technical capacity or raw narrative notes are included.
- The former full-catalog JSON and CSV downloads are retired. Bounded browsing is a presentation and delivery choice, not authentication, a paid-access entitlement or a restriction on the original source's CC BY 4.0 reuse rights. The attributed upstream source remains linked above. Eco TerraVault's professional services concern scoped analysis and advice, not exclusive ownership of this public-source dataset.
- `summary.json` and `jurisdictions.csv`: all 56 jurisdictions, with separate facility/building totals, facility lifecycle counts and four physical-setting categories. Summary JSON also includes physical-setting-by-lifecycle cross-tabulations.
- `building-records.csv`: all 1,237 building IDs and their source jurisdiction.
- `compute-atlas-license.txt`, `ODbL-1.0.txt`, `im3-source-notices.txt`: source-specific notices.
- `build-intelligence.py`: maintained normalization and export code. Reproduction also requires the pinned source files retained in the project repository.

CSV text beginning with spreadsheet formula characters is prefixed with an apostrophe; this is an export safety transformation, not source content. JSON preserves the original text. Year/month/day precision is retained where supplied; no missing dates are filled in.

The Python standard-library pipeline verifies source checksums, uses SQLite to normalize facilities, citations, jurisdictions and building IDs, then emits a deterministic server-side catalog and separate public summary, building and license files. Its local SQLite workspace is not a public cloud database. `--check` compares rebuilt outputs without writing files and fails if the retired public full-catalog files exist. Human review precedes a new source release. A previous snapshot remains recoverable through version control.

Reported events, where separately shown, form a curated source register. They are not generated from lifecycle transitions, nearby natural hazards or missing facility records. Neither no recorded event nor no record establishes absence.