Sources & methodology

Every value on this site is a claim: one source asserting one value for one attribute of one facility, with a confidence level, a date, and — for hand-researched facts — a verbatim quote. Nothing is overwritten when a new source arrives; the displayed value is computed from all claims, and disagreements stay visible.

Data built 2026-09-29 · OpenStreetMap snapshot 2026-09-27 · Epoch AI snapshot 2026-09-27

Data layers

Confidence levels

Confirmed

Stated by a primary source (the company, a filing) or strong evidence per the dataset.

Likely

Reported by a reputable secondary source or dataset, or estimated with a documented method.

Unverified

Inferred by our pipeline or community-contributed without corroboration.

Speculative

Some reason to believe it, but no direct record of it.

Values also carry a method: reported (stated by the source), derived (computed by our pipeline from source data, e.g. footprint from an outline) or estimated (a model estimate, e.g. Epoch’s power or capital cost).

How conflicts are resolved

1. Within one source, only its most recent claim for a field counts.
2. Across sources, the value backed by the highest-confidence claim wins.
3. Ties go to the higher-priority source (curated research › research datasets › community maps), then the more recent claim, then the value more sources agree on.
4. All other values are shown under the winning value as “Also reported”, with their sources.
5. Ownership, customer and utility relationships are multi-valued: every asserted party is listed with its own confidence.

Three kinds of property values

• Official assessments: the county assessor's value for a tax year, shown with the tax year and parcel.
• Recorded transactions: sale prices and dates from deed records.
• Model estimates: our comparable-assessment model (median assessed $/sq ft of footprint among ≥5 assessed data-center facilities in the same county). Always labeled, rated speculative, and never an appraisal.

Time to power & project verification

Each facility gets a verification ladder, a power-reality comparison and, where evidence allows, a time-to-power window. All are computed by explicit rules in src/lib/pipeline/power from the facility's claims, queue records and timeline. Nothing is predicted by an AI model.

Stages and evidence

Stages: Announced → Land acquired → Planning / zoning → Permitting → Interconnection application → Feasibility study → System impact study → Facilities study → Interconnection agreement / service contract → Financing → Under construction → Energization → Operational. A stage counts as reached only with dated evidence. Projected milestones and speculative claims never count. Earlier development stages can be shown as implied by a later one, labeled as such. Grid-study stages are never implied. Evidence is classified as: Community map, Model estimate, News report, Company announcement, Utility filing / announcement, Government record, Permit / site plan, ISO/RTO queue, Interconnection agreement / service contract, Satellite imagery, Construction evidence.

• Verified operating: Operation documented by imagery, government records or the operator.
• Grid-backed: A documented interconnection agreement or utility service commitment, or construction together with a grid record.
• Under construction: Construction documented; no public grid agreement found yet.
• Documented development: Land, zoning, permits or an interconnection application on record; no construction yet.
• Announced only: Only announcements or news reports; no filings, permits or grid records found.
• Unverified: Only community-mapped or modeled data.

Power reality

Announced is the largest MW reported by a company, utility, government or press source. Interconnection is MW in active load requests. Contracted / approved covers requests past an executed interconnection agreement, or a reported utility or power contract. Under construction is phases observed under construction. Operational is capacity online. On-site and dedicated generation is listed separately and never added to grid capacity. A nearby EIA plant counts only as unverified co-location.

Grid path

Facility → substation → transmission → utility → ISO. Each hop states how it is known. Documented means named in a source. Queue POI means the point of interconnection on the site's own load request. Satellite imagery means an on-site substation outline. Proximity means nearest mapped asset, always unverified. Service territory is based on utility or ISO polygons. Substation capacity and congestion are not yet sourced and are shown as such.

Time-to-power window

The window starts from the furthest documented stage. It then uses editorial ranges for how long large loads (roughly 100 MW and up) have typically taken to reach first power from that stage. These are judgments drawn from public reporting on large-load interconnection and hyperscale construction, not a fitted statistical model:

• Under construction with a known start date: the remainder of a 1.25–2.5 year build. Expansions are timed from the current phase's start.
• No public grid record before construction: one year is added to the upper bound.
• A stated date (energization, the queue's requested in-service date, or expected completion) anchors the window. It allows slippage of 2 years before an interconnection agreement, 1.25 years after one, and 0.75 years once under construction. Passed dates without confirmation are flagged. Modeled dates only widen the window.
• Confidence is High only for sites under construction with a grid record and a consistent stated date. It is Low for announced-only or unverified projects.
• The map and dashboard bucket sites by the window's midpoint. Expansions show the timing of their remaining capacity.

How AI is used

The language model sits on top of the database; it is never a source of facts. We use free models via OpenRouter for:

• Source extraction: proposing facts and relationships from a document. A fact is kept only if its supporting quote appears verbatim in the stored document text, and its facility is matched with ≥75% confidence (or a human accepts it). Such claims are marked AI-extracted, rated “likely” at most, and cite the document.
• Entity resolution: judging whether records from different sources are the same site; decisions are stored with confidence and reasoning and can be overridden.
• Routing explanations: rephrasing our routing estimate; rejected if it mentions a region or facility not in the estimate.
• Research assistant: planning database queries; the answer may cite only rows the queries returned, and figures are checked against them.
• Power evidence briefs: summarizing a facility's evidence and explaining its time-to-power window. The model sees only a numbered evidence pack. The brief is rejected if it cites an id not in the pack, uses a number, year or window not in the pack, names a substation or facility not in the pack, or claims certainty. The model explains the window; it never sets it.
• Grid and power extraction (queue IDs, serving substation, expected energization, on-site generation, financing) also requires the extracted MW, date or queue ID to appear in the fact's own quote.

History

Every build appends tracked values (capacity, status, size, assessments, permits…) to an append-only ledger with first- and last-seen dates, and dated claims (e.g. each observed construction milestone) are kept rather than overwritten. Facility pages show what changed over the last three years.

Known limitations

• Coverage follows what’s mapped or researched; many enterprise and smaller colocation sites are missing, and some OSM entries are small server rooms.
• Most facilities don’t disclose power; totals only sum facilities with a figure, and most figures for AI campuses are modeled estimates.
• Tenants of colocation facilities are rarely public; the tenant field is empty unless a source states it.
• Company names from OSM that couldn’t be matched to a reviewed company appear as “unreviewed” records.
• County parcel and assessor data covers seven counties so far (Loudoun, Fairfax, Prince William, Maricopa, Franklin, Licking, Ellis); Tarrant and Dallas (TX) were tried but lack reliable public parcel services.
• Interconnection queues list applications, not commitments; only NYISO publishes project-level large-load requests, so most data centers have no site-specific queue record.
• Utility territory polygons are approximate and overlap; the utility shown is the most likely retail provider and is marked unverified when territories overlap.
• Photos are matched by camera location and title keywords and may show a neighboring building.

Source registry

Sources that contribute claims, followed by documents cited as supporting references.

Cited references (217)