What the database measures
The canonical dataset tracks publicly documented U.S. data center developments. The July snapshot contains 1,048 records across 50 states: 952 active and 96 canceled. A record represents a named campus, facility, or disclosed development program—not an individual building unless the source treats that building as a separate project.
Discovery and review
Candidate changes are discovered state by state from company releases, government records, utility and regulatory filings, local planning materials, and reputable reporting. Automated discovery produces candidates; it does not edit the canonical CSV. A separate classification and review pass matches candidates to existing rows, identifies possible duplicates or renames, and proposes explicit actions. Actions are applied by exact project name and state, then the resulting dataset must pass schema, uniqueness, state, status, numeric, and provenance checks.
Inclusion, status, and cancellation
The working inclusion threshold is a reported investment of at least $10 million or capacity of at least 10 MW, with exceptions for projects important to policy or community history. “Announced” and “planned” do not mean a facility will be built. “Under construction” requires evidence of site or facility work, not only permitting. “Operational” requires evidence that service has begun. Canceled projects remain in the dataset with a canceled status so public promises and reversals are not erased.
Reported values, estimates, and missing data
Numeric fields are stored in U.S. dollars, megawatts, square feet, jobs, or tax-revenue dollars as identified in the data dictionary. Blank means the value was not located or was not safely comparable; it does not mean zero. Some records contain clearly labeled estimates derived from disclosed area, IT load, or comparable construction costs. The notes field records the basis. Ranges and theoretical maximums are not silently converted into committed capacity.
Headline denominators
The site’s “all tracked” figures include active and canceled records and sum nonblank stored values. Its active-only figures exclude canceled records. Neither total imputes every missing value. The working paper may differ because it can use a different frozen snapshot, exclusion rule, reported-only subset, or estimate-complete analytical table. Each publication must state its snapshot and denominator rather than treating those totals as interchangeable.
Names, identities, and aliases
Every canonical row receives a deterministic ID and route. A display-name change does not create a second project: reviewers match it to the existing development, retain the current public name, and preserve former names in notes or redirects. Rich editorial profiles overlay the canonical record only when their name and state match the reviewed identity.
Dates and sources
The dataset snapshot date describes when the release was assembled. Each row’s lastUpdated field records
its own last substantive review and may be earlier. Policy and power pages likewise show both the site snapshot date
and the date of their underlying research when those differ. Source URLs are retained in project notes and surfaced
on basic project pages where possible.
Reproducibility
The canonical input is all_projects.csv. The site generator produces the JSON, state aggregates, ranked extracts, project routes, and central statistics from that file. The snapshot manifest publishes SHA-256 hashes and byte sizes for release files. See downloads and the data dictionary for formats and units.