Evidence first. Keep the record precise.

Evidence entry types and selection

Evidence & answers contains three visibly labeled kinds of entry. A Question reviewed entry is framed by DataCenterData and does not imply that an unnamed person made the statement; a reviewed question can remain unresolved (Unknown). A Public statement checked entry preserves an identifiable speaker and statement locator. A Documented fact entry reproduces a bounded fact from an official record without extending it beyond the stated unit, place or period.

We prioritize consequential questions tied to a specific project or place, demonstrated public interest or confusion, material effects on power, water, jobs, taxes or project status, timeliness, and source evidence capable of resolving at least part of the answer. Geographic balance is a secondary safeguard, not a quota. The number of entries is never a measure of how truthful a company, state or community is.

An entry can be selected from an attributable public statement, a reader or correction submission, a material change in an official record, or a recurring question documented during project research. DataCenterData-authored questions are labeled as such. When one tests a common misreading, its result applies to that misreading: Contradicted there means the misreading is wrong, not that a particular person said it.

Evidence results

The same eight labels appear on every badge, page and download, and only Supported is shown in green.

A verdict applies only to the wording shown.

Company claims are shown as stated; they are not findings.

Policy tracker method

The policy tracker holds 567 sourced records (state, local, utility territory and regional grid). Each record is one instrument with its issuing authority, jurisdiction, action type, status, dates and primary source. Action type is the legal character of the instrument, whatever its outcome; it is not a score. Status is recorded as of the tracker date: in effect means the instrument had started and had not expired, been repealed or been blocked. Current, pending and historical records stay separate, and federal and multistate instruments are stored once and are not counted in any state's records, in-effect totals or map shading.

Review dates show when each record's controlling source was last read against the record; DataCenterData's research is AI-assisted. Fetching a page is not editorial verification, and an earlier review date alone does not mean a record is incorrect or overdue. Policy sources are re-read by the automated source checks; a detected change never edits a published fact by itself. A published record for a state is not a complete inventory, and the absence of a record is not a finding that no policy exists. Download the policy records (JSON).

Facility register method

The facility register holds 2,650 current register records, joined from DataCenterData reviewed records, U.S. EPA Facility Registry Service records, state or local registry records, Compute Atlas licensed discovery rows and operator-listed captures. Each facility page joins every source layer DataCenterData holds for one site unit. Each value says which layer stated it and when; layers are shown side by side and never reconciled into one status, and capacities of different kinds are never added together.

Campus, building, suite and presence records can overlap, so the register is not a unique physical-site census. When two records may describe the same site, an identity review stays open until a reviewed decision with captured evidence resolves it; it is never resolved by age. Each record shows when it was last checked and when its next check is due. How complete the facility register is · Download the register.

Source hierarchy

1. Controlling agreements, regulatory orders, permits and audited government records.
2. Government operating and financial records; EPA, DOE, Berkeley Lab and similar technical sources.
3. Company disclosures, clearly labeled as company claims.
4. Responsible secondary reporting used to locate or contextualize primary evidence, not replace it.

Rules that prevent false precision

• A permit is permission, not proof of use.
• A projected job is not observed employment.
• Construction workers, job-years, permanent employees and contractors remain separate.
• Announced investment is not completed investment.
• Withdrawal, consumption, discharge, make-up water and avoided potable use remain separate.
• Developer, owner, operator, tenant, subsidiary and guarantor are not silently merged.
• Local fiscal return, statewide economic effects and neighborhood/resource effects use separate ledgers.

Index and size rules

• Every total on the site, what it counts and where it appears are listed in What the counts include.
• A campus, a building, a physical facility and a provider presence are different units.
• Project-lead counts and facility-research counts are separate, overlapping inventories and are never added.
• Anticipated site demand is not measured demand, contracted demand or critical IT load.
• Project size bands use disclosed anticipated site demand. Facility bands use disclosed critical IT load only when a source identifies that scope. Numeric bands are navigation aids, not business-model labels or regulatory definitions.
• A source-listed lead is not independently verified. It enters the verification queue with its provenance intact.
• Licensed directory rows remain in a separate index with their upstream ID, edition, attribution and changes-made marker. They never become canonical records or totals without independent identity and evidence review.
• Withdrawn and expired projects remain in the historical index instead of disappearing.

Lifecycle states

Every status on a DataCenterData reviewed record is a dated evidence label, not a prediction: proposed, approved, under construction, preleasing, operating, expanding, paused, withdrawn, expired, closed, demolished or unknown. An approval is not construction; construction is not operation; and an unknown is not a zero or an absence.

Maps, the facility register and listings from other sources show the status each source reported, labeled as source-reported, plus three labels of their own: EPA registry record means the site appears in a federal registry, and a permit or registry record is not an operating facility; Lifecycle not stated means the source gives no status; Lifecycle varies means listings for one site disagree or are partly unstated.

Maps and 3D

• One map mark or 3D block is one source listing, not one verified site, and listings can repeat the same place.
• A listing is drawn only where its coordinate has a recorded source, such as the listing itself, a reviewed record or an official address match. Listings without one are counted but not drawn.
• Dashed marks have lower-precision coordinates. Marker size never indicates capacity.
• Nearby points or a shared street do not make two listings the same site; that takes an identity review.

Company pages

A company page shows the source listings DataCenterData holds for that company, with their stated statuses. Company counts may overlap, and combined totals count each source listing once. Shared-building and project relationships are individually cited with the listings. Like the facility register, a company page is not a census of a company’s sites.

Ballot results

Election office counts are from county and municipal pages, linked on each row. A press report says what that outlet reported and is not a count DataCenterData has checked, so results known only from press reports are labeled that way.

How the Indiana and Ohio indexes compare

Indiana and Ohio use the same field standard, but their published rows still have different units and depths. They are useful inventories, not additive statewide totals.

National aggregation rules

• A state page is not automatically a statewide census. Each page shows its independently source-linked records, attributed or unresolved rows, dated review window, source families checked and remaining gaps.
• Unresearched counts remain blank. They never enter a national sum as zero.
• Regional and operator totals only combine records with compatible entity units, time periods and measurement scopes.
• A disclosed-capacity total must show how many active records have a comparable figure; unknown capacity remains in the denominator.

Metropolitan-area assignments

• Metro pages use full-resolution U.S. Census Bureau TIGER/Line 2024 Core Based Statistical Area boundaries reflecting OMB Bulletin 23-01 delineations.
• A licensed row is assigned only when its source coordinate falls inside an official metropolitan boundary. City names, county names and marketing labels do not decide membership.
• Boundary points count as inside. If a point ever matches more than one boundary, an interior match wins; otherwise the lowest five-digit CBSA code wins and the row is flagged for review.
• Micropolitan areas and rows outside every CBSA remain counted in the assignment audit but do not become metro pages. Invalid, missing, non-finite, out-of-range and Null Island coordinates fail closed.
• Metro grouping is navigation over licensed discovery rows—not independent verification, a parcel determination, a capacity total or a census.

Complete-inventory protocol

1. Build a discovery union from operator lists, state and local records, utility and commission dockets, permits, tax and incentive records, historical archives, and commercial directories used only as leads unless licensed.
2. Resolve each lead to a project, campus, physical building, provider presence or historical event. Unresolved leads stay visible and out of confirmed totals.
3. Check current and former names, subsidiaries, acquisitions, closures and withdrawals so a rename neither deletes nor duplicates a site.
4. Publish a statewide total only after the named source families and research leads have been reconciled to the same record unit.
5. Keep monitoring after publication. A census is a dated snapshot, not a permanent claim.

Document packets and watch points

A dossier’s source packet lists the controlling or evidentiary records reviewed and explicitly names missing agreements. Page or section locators identify the relevant provision when possible. A watch point may be a date, phase trigger, annual filing, docket event or replacement proposal; an unknown date remains unknown rather than becoming an invented deadline.

Public and social statements

A public post proves that its author made a statement; it does not prove the underlying statement is true. Published entries keep the original platform, URL, author role, publication and capture dates, linked project, exact assertion, independent evidence and remaining unknowns. Official company, government, utility and regulatory accounts are the initial publication scope.

Ordinary people’s posts may inform the private research queue, but DataCenterData does not republish their handles, screenshots, comments or distinctive verbatim wording merely because a name can be hidden. What’s being said lists news headlines found automatically through the GDELT Project and posts from a checked list of official Bluesky accounts, each confirmed by a link on its office’s own government website. A news row shows the headline and a link to where it was published, with no quotation, image or summary; an official post shows its own wording. Where a story names a place DataCenterData covers, the site’s records sit beside it and do not assess it. Stories that may name a private person are held back, rows are rechecked when the list is refreshed, a post its author deleted loses its wording, and any row can be taken down on request.

Automated source checks

DataCenterData re-reads the public sources behind its published records on a schedule so that a changed, moved or withdrawn document reaches a person for review. The checks identify themselves as DataCenterData-SourceCheck (the recurring monitor), DataCenterData-PolicyMonitor (manual policy-source checks) and DataCenterData-Research (research retrieval directed by an editor), read each site’s robots.txt before requesting a page, and stop at logins, paywalls, CAPTCHAs and access controls. They request single documents, not bulk collections, and the copies are used only to verify and monitor our records, not to train models. A detected change never edits a published fact by itself. Site operators can contact us through the corrections page.

Corrections and commercial independence

We publish the source, source period, last-reviewed date and uncertainty with each finding. Better primary records should lead to a visible correction. Sponsorship, if ever accepted, cannot buy a verdict, select only favorable projects, delay a finding or suppress a correction. Editorial findings and commercial relationships remain separate and disclosed.

Featured claims get a second source check: a separate software-assisted review that did not write the original record re-reads every cited source. It is a workflow check, not a human expert review and not an outside audit. “Second review complete” on a claim means that check has finished.