How the dataset is assembled and updated

1. Unit of observation and scope

A record is a data center facility, campus, or named build phase that a public source identifies as existing, under construction, or planned. Campuses and the buildings and phases inside them are held as separate records connected by an explicit hierarchy link, never silently merged — the same infrastructure described at two levels stays visible at both, and aggregation logic (section 4) prevents it from being counted twice. Closed or decommissioned sites are retained with a closed status and excluded from every forward-looking aggregate.

2. Sources: hierarchy and conflict resolution

Every fact traces to a retrievable public page, retained with its URL and access date. When sources disagree, resolution follows a fixed ranking rather than editorial judgement: operator disclosures and official filings outrank planning and government documents, which outrank utility and regulatory materials, which outrank press coverage and directory listings. Within a rank, the more specific and more recent statement wins. Facts extracted for the same field from different pages are held as competing candidates and resolved by these rules; the losing candidates are retained internally so a resolution can be revisited when new evidence lands.

Every quantitative value carries the field-specific source URL and the date that page was read, published beside the value rather than as one citation for the whole record. A facility can take its power from an operator press release and its cooling description from a sustainability report, and the catalogue keeps them apart, so a reader can judge each value against the page that states it instead of trusting an aggregate quality claim.

3. Field definitions and normalisation

Facility power is the facility or campus power under the source’s definition, in MW. It is not annual electricity consumption. Values reported in other electrical units are converted explicitly — apparent power (MVA) is converted only where the conversion basis can be stated, and the conversion is recorded with the value. A reported range is reduced to a stated midpoint and marked as range-derived. IT load is held in its own field and never substituted into facility power; where a source states only IT or critical load, facility power stays blank.

Announced full build-out capacity (“designed for up to X MW”) is held in a separate ultimate-capacity field with the stage of the un-built remainder, and is never added to facility power or to any stage total: expansion phases routinely exist as their own records, and summing a parent’s announced ceiling as well would double-count the pipeline. Values that fail physical plausibility screens — unit slips, transposed figures — are withheld from publication and logged for review rather than published or silently corrected. Missing values remain blank and are never converted to zero.

4. Facility hierarchy and aggregation

Aggregation follows each hierarchy branch and uses the highest available facility-power value on that branch. When a parent campus states no power, the method moves down the branch to the nearest available values. A parent and its children are therefore never added together, and a campus with no stated total still contributes its children’s values. Country and company aggregates sum over these hierarchy-deduplicated values only from records that state a power value — they are floors supported by the public record, not estimates of total installed capacity.

5. Development status

Every record is classified as operating, construction, planning, closed, or status-not-specified, from the same ranked sources as its other facts. Planned and construction capacity is not operating capacity, and every published comparison keeps the stages separate. A status change — a campus entering service, a project cancelled — is applied through the corrections process (section 7) with the page that proves it, so the stage split can be traced and checked release over release.

6. Coverage

The current release catalogues the corpus described on the coverage page: the facility, country, city, and operator counts, the stage split, and how many facilities record each attribute. Public disclosure differs by operator and by market, so attribute availability varies across the corpus; the coverage page states it per field so the dataset’s reach can be assessed before purchase.

7. Corrections and adjudication

Corrections are applied through reviewed overlays, never by editing the corpus in place: each correction names the record, the field, the action, and the public page that proves it, and ships in the next versioned release. Values proven wrong are replaced or suppressed; a source URL shown to describe a different facility is unbound from the value it was wrongly attached to. Independent verification sweeps re-verify the highest-impact records against their sources, and their findings enter the same overlay process. A release that has been sold is never modified — see versioning.

8. Derived values are quarantined

A facility that states no power value is never given one. Nothing on this site, in any total, ranking, aggregate, chart or exported figure, is derived from a peer average, a floor-area rule of thumb, or any other proxy. Where an internal size estimate exists it stays internal and is excluded from every published and purchased figure, and automated tests enforce that boundary against the published field names on every release. The coverage page states how many facilities record each attribute, so the size of what is unknown is visible rather than filled in.

9. Reserved: electricity and emissions

The commercial release publishes no electricity-consumption or greenhouse-gas figures and no emissions factors. The calculation layer is specified and reserved for a future versioned release: electricity derives from load or capacity with explicit PUE, utilisation, and stage assumptions; emissions require a declared accounting boundary, and on-site or contracted supply changes the applicable energy mix, not the demand. When this layer is released it will carry its own method version, input-release lineage, and scenario definitions, and this page will be extended before any such figure is sold. No number from this reserved layer appears anywhere in the current product.

10. Versioning and updates

The corpus is rebuilt monthly: new facilities, status changes, corrections, and new evidence are compiled, checked by automated release gates, reviewed by a person, and only then published under a new immutable release ID with a fresh checksum. A release that has been delivered is never edited in place. Interim months without material change publish no new release. The release notes record what changed in each release.

11. Limitations

Public facility information is incomplete, uneven across markets, and changes over time. Aggregates cover only facilities whose public record states a power value, and source definitions can differ — campus power, utility connection, and project scale are not always the same quantity, which is why the source definition travels with the value. The dataset records what the public evidence supports at access time; it is not investment, engineering, or regulatory advice.

12. Proposing a correction

Send the facility ID, the affected field, the proposed value, a public source URL, and a short explanation through the contact page. Substantiated corrections enter the overlay process and ship in a later release; the reporter is not identified in the data.