inframap.DATA CENTERSInfraMap Edu ↗

METHOD · VERSION 0.2

Every claim needs
its evidence.

Our goal is a public record of US data centers at building resolution. This research release establishes the discovery pipeline. It does not yet establish comprehensive coverage.

Find candidates through different sources

Community maps identify visible sites. Air permits, planning applications, utility filings and operator disclosures identify projects and phases. Each source has blind spots, so we keep its identity, retrieval date and supported fields.

The current release ingests OpenStreetMap data center tags, Virginia DEQ's air permit register and Google's US location directory. Two document dossiers also demonstrate extraction from an issuer-hosted SEC exhibit, an operator announcement and a facility specification. These examples preserve dates and quantity definitions. The research pipeline also captures state filings, hyperscaler documents, operator specifications and government geospatial records. These source records remain separate until identity and scope review resolves overlaps. There is no general SEC crawler.

Keep buildings and campuses separate

An OSM object can represent a building, a campus or a facility point. We retain that distinction. We only draw a building footprint when the source identifies it as a building. An adjacent warehouse, cooling building or office does not become a data center because it is nearby.

Distance alone does not merge records. A future match needs shared identifiers or a documented identity review. Stable IDs use the source and source identifier, so correcting a name or coordinates does not break its link. Campus records and their buildings must not be summed as distinct facilities.

Correlate county structures with building polygons

Loudoun County identifies addressable structures and their land use. Our accepted associations require an exact normalized street address, one county data-center structure inside one building polygon, and no competing mapped point, overlapping polygon or duplicate county structure at that address. Ambiguous rows stay unresolved. The county's combined existing-or-permitted status does not prove operation, and construction year does not establish opening year.

Separate power quantities and phases

IT capacity measures critical computing load. Total facility power also includes cooling, lighting and losses. Backup generation, contracted utility service, ultimate campus plans and live IT load are different quantities. Each needs its own scope and effective date.

We keep missing values empty. A standard PUE assumption does not fill a reported field. A generator permit does not establish grid demand. Footprint area is ground coverage, not building floor area.

Read permits as evidence, not completion

The Virginia register groups entries by registration number and preserves permit revisions. It does not establish current construction or operation. Locality-only records stay in the research queue. Operator directory pins can mark a city center, so those also stay out of the map until we confirm a site location.

Use dated imagery to check construction

The imagery pipeline searches Microsoft Planetary Computer for Sentinel-2 and NAIP scenes. Published assistant reviews record visible conditions and roof-appearance bounds, with source resolution and limitations. They do not count as an independent accuracy audit. For the initial pilot, it samples one location per state, prioritizes source-tagged construction, and selects lower-cloud scenes by month. Every image links to its acquisition metadata.

Sentinel-2's 10 m pixels can help identify large changes. They often cannot resolve individual small structures. Scene-wide cloud cover does not prove that the site itself is cloud-free. NAIP offers finer aerial detail, but survey dates can be years apart. These sources support review; neither proves that servers are operating.

A reviewer must record which dated scenes were examined, what is visible, the construction stage, limitations and their name before a record can claim imagery verification. Collecting a timeline alone never changes status.

Show uncertainty and measure errors

We publish unknown fields and the source behind known fields. No percentage score claims certainty. Independent verification is separate from source attribution. The coverage page compares field completeness with a dated Cleanview snapshot and explains the difference in counting units.

Before claiming matching accuracy, we need a stratified sample review that checks identity, coordinates, status and capacity against independent primary evidence. The current release has no completed accuracy audit.

Sources and reuse

PeeringDB and the retired Edu inventory are excluded from public data. PeeringDB's data terms require separate review; the old inventory has unresolved provenance.

Reproducibility and corrections

Downloads include stable IDs and source links. The release manifest records file hashes. The build checks source permissions, coordinates, identities and capacity provenance before exporting. Source refreshes run on demand; no unattended update schedule is active.

Use a record's correction link to send a public source and the proposed change. The link opens an email draft. It does not edit the published map automatically.