← Back to the U.S. map

How we build the data center map

The map helps readers explore publicly documented U.S. AI and large data-center sites. Coverage is incomplete, and a mapped location does not by itself establish that a facility is operating or running AI workloads.

What is on the map today

The interactive map contains 3,866 U.S. data-center records from Epoch AI, Compute Atlas, DataCentersExposed and OpenStreetMap. All 77 original Epoch entries remain. Coverage includes data centers whose AI use or size is unknown, and ambiguous matches may remain separate.

Coordinates may be approximate. Power figures retain the source’s description, including estimates or planned capacity; they are not live meter readings. Missing values do not mean zero. Each record links to its dataset page. The date above the map and in its attribution line comes from the latest recorded dataset retrieval or local import, shown in Mountain Time. It is not a claim that every facility was newly reviewed that day.

Expanded dataset and downloads

The map uses a compact index for search and clustered pins. Full source details load in geographic chunks as you explore at closer zoom or select a pin, rather than downloading the complete detailed database on arrival.

Download the combined database, inspect the current counts, source timestamps and chunk index, or read the changes log. All four sources are loaded. Local import times are recorded separately from upstream record dates; no automated download of the two uploaded files is claimed. OSM’s supplied boundary-derived state takes precedence over its address tag. Records without usable coordinates or with prohibited directory provenance are excluded and counted in the log.

Records with PeeringDB or commercial-directory provenance are excluded. The expansion contains documented data centers whose AI use or scale may be unknown; inclusion is not a claim that every site runs AI. Missing or conflicting power figures remain unavailable. The U.S. filter covers the 50 states and District of Columbia.

Sources for the expanded dataset

The interactive map and downloadable dataset use the following priority order.

1. Epoch AI, AI data centers — Creative Commons Attribution. Retain the existing 77 entries and their provenance.
2. Compute Atlas by Edward Kubiak — CC BY 4.0, with a public repository and bulk snapshot. Filter to data centers; exclude standalone crypto-mining and power-generation records.
3. DataCentersExposed contributors — ODbL 1.0. The locally supplied campus CSV includes OpenStreetMap-derived records; these are not independent corroboration of the same OSM object.
4. OpenStreetMap contributors — ODbL. The locally supplied point export contains U.S. objects tagged telecom=data_center or building=data_center. These tags do not establish facility size, operation or AI use. Unnamed objects receive a descriptive placeholder with their OSM identifier.

How records are combined

Merge candidates are compared using proximity—approximately one kilometre—and a matching site name or operator. Nearby sites with different identities remain separate. Higher-priority details are preserved, and contributing sources are retained. Rounded locations and ambiguous campus boundaries still require review.

Records use operating, under construction, planned or unknown status, according to the evidence. Confidence is high for Epoch records or corroboration by at least two independent sources, medium for one named source, and low for OSM-only evidence. An OSM record copied into another dataset is not independent corroboration. Confidence describes the record's evidence, not a safety rating.

MW figures are displayed only when supplied by a source, retaining whether they describe an estimate, IT load, facility power or planned capacity. Missing capacity is not calculated from building size, server counts or other proxies.

Refresh dates, changes and reuse

The downloadable dataset includes a manifest with source timestamps and a changes log. Local import times are recorded separately from upstream record dates. The date shown on the map represents the latest recorded dataset retrieval or local import, not a new review of every facility. A page rebuild alone is not a data refresh.

A compact index supports search and clustered pins. Detailed records load in geographic chunks as readers explore at closer zoom or select a pin.

The combined database incorporating ODbL material is available under ODbL 1.0, with downloadable data and source credits. The published dataset uses U.S. filtering, field normalization and duplicate merging. Attribution: Epoch AI; Compute Atlas by Edward Kubiak; © DataCentersExposed contributors; © OpenStreetMap contributors. Source authors do not endorse Grid & Silicon’s analysis.

Send a correction with the site name and a public source.