Useful evidence starts with honest boundaries.

Three layers. Three different meanings.

The tool keeps facility catalog records, mapped buildings, and reviewed public reports separate. These datasets have different definitions, dates, and coverage. They are not added together, joined by a guessed location, or used to calculate incident rates.

Facility catalog

1,605 records classified as data centers in Compute Atlas v1.34.0, released September 21, 2026. Crypto-only and dedicated power-generation records are excluded. A source record can represent a facility, project, or campus; it is not necessarily one building. The snapshot is licensed under CC BY 4.0.

Mapped buildings

1,237 distinct building IDs in the PNNL-led IM3 / OpenStreetMap February 9, 2026 snapshot, including one Puerto Rico record. Repeated county-boundary rows are deduplicated. Separate point and campus layers are excluded. This incomplete building inventory is available under ODbL 1.0, with attribution to OpenStreetMap contributors.

Reviewed reports

A small, manually curated collection of historical operator postmortems, government records, and facility observations. Each keeps its own date, geographic scope, source link, and limitations. A regional service event is not assigned to a building. A group report is not silently converted into multiple geolocated facilities.

Above ground and underground: classify the evidence.

Physical setting describes the documented location of data-center space relative to ground level. It is a separate Eco TerraVault review layer, not a field supplied by Compute Atlas and not a measure of safety, sustainability, or operational performance.

Above ground

A facility-specific source documents data-center space in an above-ground building. This does not establish that every ancillary room or utility is above grade.

Underground

A facility-specific source documents data-center space below grade, including a mine, cavern, or underground bunker. An underground power feed, cable, reservoir, or parking area alone is insufficient.

Mixed

A source explicitly documents both above-ground and underground data-center space within the same record’s scope. Conflicting evidence alone does not justify this category.

Not classified

The reviewed evidence does not support a category yet. The record may be unreviewed, ambiguous, or insufficiently documented; this does not imply above ground or no facility.

The reproducible algorithm joins the reviewed evidence register to exact catalog IDs and jurisdictions. Only supported classifications are applied; all remaining records default to not classified. It then aggregates mutually exclusive categories by jurisdiction and reported lifecycle. The pipeline rejects unmatched IDs, duplicate assignments, invalid categories, missing citations, and jurisdiction mismatches. It does not predict building type from names, imagery, location, or keywords.

This initial review covers 8 of 1,605 catalog records. Classified examples are a targeted research sample, not a representative census. A jurisdiction with no documented underground example may still contain underground data centers. Sites absent from this catalog are not silently added to its totals.

Setting and lifecycle filters intersect on the same records. A proposed setting describes an intended design, not a completed facility. Counts retain the catalog’s site/campus scope and possible overlaps. Open a facility’s evidence panel for the source, review date, scope, and limitations; review dates do not imply a site inspection or live verification.

A reported status is not a live measurement.

Operational, under construction, permitted, proposed, and cancelled are classifications inherited from the facility catalog. They describe the source’s account at a point in time. “Operational” does not mean independently confirmed uptime today, full campus completion, spare capacity, or that every planned phase is running.

Campus records and individual facility records can overlap in the catalog. The totals count source records, not unique physical facilities. The tool preserves source identities rather than applying an unsupported proximity-based deduplication rule.

Closed and decommissioned are distinct findings. Closure alone does not establish physical dismantling, remediation, or demolition. A dormant classification requires explicit evidence; silence, an old webpage, or a cancelled proposal is not enough. This release has no verified dormant-site example.

Confidence and publisher labels are inherited from the source. A publisher label does not guarantee an operator-owned destination; links show their actual domain. These are not Eco TerraVault certifications. The proposed Eco TerraVault assessment framework is separate from this public research tool.

Territories belong in the research.

The workspace includes 50 states, the District of Columbia, Puerto Rico, Guam, the U.S. Virgin Islands, American Samoa, and the Northern Mariana Islands: 56 jurisdictions. All remain selectable even where a source has no records.

U.S. Minor Outlying Islands are outside these source inventories and boundary layers. We do not invent shapes or counts for them. Generalized Census boundaries are intended for display, not parcel analysis, engineering, or navigation. Alaska, Hawaii, and the territory cards use insets; the territories are independently scaled and are not shown at a common physical scale.

A zero in a source means no records in that source and scope. It does not establish that no data centers, incidents, or opportunities exist in the jurisdiction.

Evidence before interpretation.

The tool does not infer environmental performance, cybersecurity posture, energy availability, water demand, operating capacity, or ethical quality from facility counts. Capacity estimates from the catalog are deliberately omitted: records can mix expansion increments, ultimate buildout, and operating capacity, making simple totals misleading.

Event counts reflect editorial coverage rather than comparative reliability. Sources may be incomplete, outdated, or revised. Historical reports do not imply current unresolved problems, legal fault, or endorsement by the named organizations. Follow the original evidence and commission project-specific assessment when a decision depends on it.

Reproducible publication and corrections.

Python validates pinned source files, identity rules, counting units, URLs, and geographic scope. A local SQLite research database retains normalized records and their provenance. Reviewed snapshots are exported to the public website with source dates and licenses; public reads do not query a live commercial directory.

Updates require a source review, regeneration, tests, and a versioned release. The review date is not the date of every underlying observation. This is not a live status monitor or an automatically refreshed inventory.

Use the contact form to suggest a public source or request a correction. Include the record name, the claim to review, a public source URL, and its date. Please do not submit confidential facility information or credentials.