Methodology
Every number on this site can be traced: to a source for measured facts, to a published formula for derived indicators. Start here.
The dataset
Indicator register
Social friction index
Water stress index
Disclosure coverage
Conflicts, pipeline, AI figures
Data dictionary
Release notes
One row per physical asset, deduplicated across sources, with every substantive value wrapped as { value, source_url, confidence, source } in the full dataset, which is not distributed (the free sample flattens this to one src_<field> source column per value and has no confidence levels). Power and area figures are measured: never estimated: coverage is honest but sparse, and absence is not proof a facility doesn't exist.
A two-layer pipeline keeps it reproducible: a deterministic layer (open bulk data → geometry → dedup → permanent IDs) and an agent layer that only adds measured metadata on top, keyed by the permanent dc_uid. The agent layer never edits the deterministic layer.
Ingest
Fetch OSM (Overpass) + PeeringDB; build footprint polygons.
Resolve
GADM point-in-polygon country (→1 km→10 km buffer); Wikidata operator QIDs.
Dedupe & merge
Match country + coords (~150 m) + name; hard-join on shared IDs; mint permanent dc_uid.
Classify
Separate data center vs network POP vs unknown (footprint → operator type → interconnection density).
Enrich
Agents harvest measured MW / area / PUE / certs from operator pages, each datum source-linked.
Validate
Review un-merges wrong joins and merges missed duplicates; a documented correction layer fixes what remains (public log on /releases).
Internal output: a GeoPackage (points · footprint polygons · best-of layer) and a JSONL record store with per-field provenance, a 0–100 Data Quality Score, a geo-flag and an asset-type on every row. Neither file is distributed: see what you can get.
Every value is source-linked and carries the licence of its origin. The dataset is OSM-derived, so ODbL governs redistribution (attribution + share-alike). Hover a licence to see its main rules.
Re-verify per-source terms before redistribution or commercial use, a research compilation, not legal advice.
| Source | Licence | Provides |
|---|---|---|
| OpenStreetMap | ODbL 🟢 | Locations + building-footprint polygons, names, operators |
| PeeringDB | AUP 🟡 | Facility identity, address, interconnection counts |
| Wikidata | CC0 🟢 | Canonical operator names, inception years |
| GADM 4.1 | ref-only | Country assignment (the boundaries are not redistributed; the country name it gives is) |
| Operator sites | site terms 🟡 | Measured capacity (MW), floor area, PUE, cooling, certs |
| SEC / filings | public | Measured facility facts, each with the filing URL |
- MW coverage is sparse (~299 of 6,790 data centers). Treat missing as unknown, not zero.
- Coverage is uneven by region/operator; OSM & PeeringDB completeness varies.
- Discovery coordinates may be area/city centroids (medium confidence).
- OSM-derived → ODbL applies to redistribution; PeeringDB under its AUP.
The underlying dataset is not exposed for bulk download or scraping. The map is clustered server-side and only ever returns aggregated zones or a bounded, capped set of points; full per-facility metadata is served one asset at a time on click. Re-verify per-source terms before any redistribution or commercial use, this is a research compilation, not legal advice.