What data centers are actually permitted in your county.
There is a great deal of noise about data centers right now and very little that anyone can check. Groundwork reads the permits. For every county it tracks, it publishes what operators have actually filed with the regulator — how many sites, how many backup generators, how much they are permitted to emit — each figure sourced to the filing it came from. No press releases, and no panic.
Start with where you live.
Read the map as a floor, not a census. Coverage is uneven by design of the source, not by design of ours: air permitting is a state function, so there is no national permit list. The national layer is EPA's registry, where the industry code is self-reported — which is why Texas shows only 61 facilities here despite being one of the largest data center markets in the country. An empty county means nobody has filed under a code we can see, never that nothing is there.
Of the 412 sites located precisely enough to check, 16 sit inside a FEMA Special Flood Hazard Area and 176 draw from basins WRI rates high or extremely high for water stress. Virginia — the densest cluster on the map — has zero in a flood zone; its exposure is water.
91 of 503 tracked sites have no verified coordinate and are deliberately absent from the map — placing them on a county centroid would imply a precision the permit does not support. They are all reachable through the search box above and through the county tables below.
The aggregate nobody else discloses.
Permit-by-permit, a data center build-out looks routine. County-by-county, it looks like this. These totals exist only because someone added up hundreds of separately-published filings. Generator counts appear only for Virginia, where Groundwork reads the permit documents themselves.
| Locality | Sites | Permits | Generators permitted | In flood zone |
|---|---|---|---|---|
| Loudoun County | 85 | 92 | 5,882 | — |
| Fairfax County | 31 | 35 | 1,757 | — |
| Prince William County | 28 | 29 | 2,058 | — |
| Cook County | 25 | — | — | — |
| Douglas County | 20 | — | — | — |
| Bexar County | 15 | 42 | — | — |
| Fulton County | 15 | — | — | — |
| Dallas County | 11 | 19 | — | 2 |
| Maricopa County | 11 | — | — | — |
| Mecklenburg County | 10 | 11 | 16 | — |
| Henrico County | 8 | 8 | 248 | — |
| DuPage County | 7 | — | — | — |
| Tarrant County | 6 | 8 | — | — |
| Middlesex County | 6 | — | — | 2 |
| Montgomery County | 6 | — | — | — |
| Sarpy County | 6 | — | — | 1 |
| Collin County | 5 | 19 | — | — |
| Harris County | 5 | 11 | — | 1 |
| Arapahoe County | 5 | — | — | — |
| Johnson County | 5 | — | — | — |
The 20 densest of 131 counties. A dash in the permits or generators column means Groundwork has the facility from EPA's registry but has not yet read that state's permit documents — not that the facility is unpermitted.
Every county tracked
Permits stop just short of the line.
Virginia air permits tabulate how much nitrogen oxide a facility may emit each year. At 100 tons a year a new source becomes a major source, which brings a stricter review and a public participation process. Of the 155 data center permits where that figure is readable, here is where they land.
Tons of NOx per year permitted, per facility. Read from the emissions table of each permit and labelled probable — a figure parsed from a PDF, and a permitted ceiling rather than measured emissions.
It is not evidence of wrongdoing. Designing a facility to stay under a regulatory threshold is lawful, ordinary engineering, and every industry does it. These are also backup generators — permitted to run during outages and for testing, typically tens of hours a year.
Groundwork takes no view on any individual application, and does not claim any single permit was sized to avoid scrutiny.
The distribution is not what engineering need alone would produce. 70 permits land in the last ten tons below the line; 2 cross it. A threshold that was incidental would not bend the curve this hard.
Whatever else is true, the review process that 100 tons is meant to trigger is being triggered 2 times, in the densest data center market on earth.
Which facts you can lean on.
There is no blended score. A single number would hide which inputs came from a filing and which came from a fuzzy match, and that difference is the product.
| Tier | Means | Source |
|---|---|---|
| confirmed | Read straight from a filing, or an exact point-in-polygon lookup | State air permits · EPA registry · FEMA NFHL · WRI Aqueduct |
| probable | Matched on a non-unique key. Very likely right, not certain | Addresses and emissions read from permit PDFs |
| reported | Established by journalism, no filing yet. Upgrades automatically | Named publication, per entry |
| pending | Not collected yet. Never a claim that nothing is there | — |
How this is built, and where it is weak.
Sources
- VA DEQ data center permits — 201 permits, each PDF read for address, equipment and permitted NOx.
- EPA ECHO — 253 facilities in other states, by industry code and operator name.
- FEMA NFHL and WRI Aqueduct 4.0 — flood zone and water stress, by exact location.
Limits
- A floor, not a census. Permitting is a state function. Depth varies by what each state publishes — every county page says which.
- Virginia is deepest by accident of disclosure, not importance. It is the only state with a data-center-specific permit list.
- Permits are ceilings, not measurements, and approval is not construction.
- 91 sites have no verified coordinate and are off the map by choice. A county centroid would imply precision the filing does not support.
- 610 facilities were excluded — they carry the data-processing industry code but nothing identifies them as data centers.
Spotted an error?
Fuzzy-matched fields are labelled as such and some will be wrong. Open an issue — corrections are tracked in public and noted on the page.
Groundwork compiles US public disclosures. Not investment advice, not a flood model, and not a substitute for a site-specific engineering or insurance assessment. Confidence tiers describe how a fact was established, not how severe a risk is. Download the dataset.