Methodology

1Coverage

The data covers the equipment used to build and power US data centers, and the listed companies that make it. At the site level that means more than 500 US data center projects and every on-site generating unit in their filings: backup diesel generators, gas turbines, reciprocating engines and fuel cells, including generation meant to run behind the meter. Each unit carries its count, kW rating, manufacturer and model, fuel, emissions tier and permitted operating hours.

At the company level it covers listed makers of transformers, switchgear, cable, fuel cells and memory, and the operating figures they disclose: order backlog, new orders, plant utilization, production capacity, facility investment, R&D spending, export share and individual large sales contracts.

2Sources

The data is built primarily from public sources: regulatory filings by listed companies and records published by government agencies. Site equipment comes from permit filings. Company figures come from quarterly and annual reports and from event notices, which disclose individual contracts with their value, term and often the counterparty.

Business entity registrations, local government records and company announcements are used only to establish who is behind a site (section 5), never as a source of equipment or financial figures.

3Collection and extraction

Sources are checked daily for new documents and amendments. Each document is stored as received, with its retrieval time and a content hash, and a document re-posted with different contents is processed again.

Scanned site filings go through OCR and table extraction to produce one row per unit. Every extracted value keeps a reference to its document and the page or table it came from. Where a document states its own totals, extracted values are checked against them, for example unit ratings against the facility total in a permit, or segment figures against the company total. A value that does not reconcile is held until it has been reviewed.

4Translation

Most company filings in the data are not in English. We translate the passages that values are taken from, such as business descriptions, order and capacity tables and contract notices, and read the numbers from the original-language document.

Product, equipment and accounting terms go through a fixed glossary that maps each source term to one English term and one field. A record with a term not yet in the glossary waits until an analyst adds it. Company names use the company's registered English name where one exists. The original text is kept with every translated passage.

5Unmasking site owners

Site filings are usually made in the name of a single-purpose project LLC or the engineering consultant that prepared the application. The developer or operator is often not named at all. We identify that company from public evidence, typically:

• officers, registered agents and mailing addresses in business entity registrations, which often lead to a parent company
• other records for the same parcel or address, such as zoning, utility, tax incentive and building permit records
• earlier filings by the same applicant or consultant that name the operator
• company announcements and local government meeting records about the project

Developer and operator are separate fields, since a developer often builds a site that another company leases. Each link lists the documents it rests on and is marked direct, when a document about the site names the company, or inferred, when it rests on indirect evidence such as a shared officer or address. If the evidence supports neither, the field stays empty and the site is marked unresolved. Links are checked again whenever a new document for the site arrives.

6Normalization and deduplication

Filings state output in kW, MW, kVA or horsepower. Ratings are converted to kW, and the reported value and unit are kept beside the converted figure. A conversion that needs an assumption, such as a power factor, is flagged. Manufacturer and model strings are standardized, collapsing abbreviations and spelling variants into one manufacturer and one model, and each unit is assigned an equipment class and fuel.

Company figures keep their reported currency and basis (consolidated or standalone). Cumulative year-to-date figures are converted to discrete quarters by subtraction and marked as derived. Each company has a stable Meteor Labs identifier linked to its legal name, listed parent and ticker, and renames, mergers and spin-offs are recorded as dated changes to those links.

A project usually produces several documents: an application, draft and final permits, modifications, renewals, and sometimes a permit reissued after a change of ownership. These are matched on the identifier the issuing agency assigns to the facility where one exists, and otherwise on address, parcel, coordinates, applicant and equipment list, then merged into one site record with its document history attached. Buildings or phases on one campus stay separate projects under a shared campus identifier, since they are often permitted and built on different schedules.

Contracts are handled the same way. A contract that appears in an event notice, a later amendment and a periodic report's list of major contracts is stored once, with its amendments in order.

7Project changes

When a new document for a known site differs from the previous one, the difference is stored as a dated event: units added or removed, ratings changed, fuel switched, operating hours revised, a construction deadline extended, or the permit withdrawn, denied or allowed to lapse. A deadline extension or a reduced equipment list flags the site as possibly delayed, and a withdrawal, denial or lapse flags it as possibly cancelled. Each flag cites the event that set it.

8Aggregation

Unit rows are summed to site totals of unit count and kW by equipment class and fuel, and site totals are summed by developer, operator, equipment manufacturer and state. Documents are merged before summing, so each site counts once however many filings it has. Withdrawn, denied and lapsed sites are left out of active totals and kept in the data with their status.

Totals by equipment manufacturer show each maker's permitted units and capacity across tracked projects. All aggregates are computed from delivered rows and can be reproduced from them. When a row changes, the aggregates it feeds get a new version.

9Point-in-time history

Records are never overwritten. A new filing, an amendment, a restatement by the company, a corrected extraction or a changed ownership link each produces a new version of the record. Each version carries these fields:

To see the data as it stood on a past date, take the latest version of each record with recorded_at on or before that date. A correction made in March to a January filing has a March recorded_at, so a backtest as of February sees the uncorrected value. Use published_at for event studies around filing dates.

Versions and project events (section 7) are kept apart. An event records what changed at the site. A version records what we knew about the record and when.

Versions created by backfilling a newly added source carry their actual collection time in first_seen_at and recorded_at, so history assembled later is never shown as known on its publication date. The coverage log lists each source's start date and known gaps.

10Delivery

The data updates daily as documents are published. Delivery is by flat file, cloud data share or API. Each delivery contains the versions written since the previous one, and a full snapshot with all history is available on request. Corrections arrive in the next delivery as new versions with revision_reason set to correction. The data dictionary, company identifier tables and coverage log come with every delivery.

11Compliance

All data is derived from documents the issuing company or agency has made public, collected through the publisher's official channels and in line with its terms of use. We do not use material non-public information, surveys, channel checks, expert networks or contact with company employees.

Names of individuals in source documents, such as consultants' contacts or company officers, may be used as evidence of who is behind a site and are excluded from delivered data. We complete standard alternative data due diligence questionnaires on request.