The United States data centre record
1,853 facilities at real coordinates, counted from public records. Three pens on one sheet: what was reported, what was inferred from it, and what is only projected — with the paper left blank wherever the instrument could not see.
- Station
- United States
- Chart speed
- 1 yr / div
- Last advance
- 2026-07-29
- Regions
- 276 · 47
Move across the chart, or focus it and use ← →
Three kinds of number
They are not interchangeable, and this is the only page that has to explain it before you read anything else.
A location is reported — a contributor mapped it and you can open the element and check. A date is observed: the month OpenStreetMap first recorded the facility, never the month it was built, because OpenStreetMap carries no construction dates. A megawatt figure is inferred, a share of a national total published by Lawrence Berkeley National Laboratory divided up by building floor area, and it is an upper bound rather than a meter reading. Anything past 2024 is predicted — a published scenario, drawn as a range because that is how it was published.
The count and the curve disagree, deliberately. The trace ends at 1,666 while the snapshot holds 1,853. The 187 in between are facilities whose creation falls outside the recorded edit history, or arrived after the trace last advanced. They are in the count and not in the curve, and that is stated rather than reconciled away.
Where it concentrates
The densest 11 counties, each drawn at the same gain. Scaling every row to its own maximum would make them all look equally busy; the finding is that they are not.
- Gain
- 239 full scale
- Net, 12 mo
- +345
347 of them carry no power figure
A campus mapped as land, a bare point with no geometry, and a site still under construction are all counted and all located. None of them has a floor plate, so none of them takes a share of the national total. Their power is unknown, which is not a small number.
The other 1,506 are buildings and do take a share. Every bar below ends at a stop mark, because each is an upper bound: the pieces re-sum to LBNL's published 21,918 MW for 2024 exactly, and a test enforces it. inferred
What moved
The last 500 edits: 452 arrivals and 48 removals.
A removal means the element stopped matching the data-centre filter in OpenStreetMap. It is not a demolition and this site never says it is.
- 2026-06-19Lumen Minneapolis 3 · Hennepin County, MN · appeared
- 2026-06-18DC Blox Data Center · Jefferson County, AL · appeared
- 2026-06-16DataBank SLC3 · Salt Lake County, UT · appeared
- 2026-06-16NTT · Dallas County, TX · appeared
- 2026-06-16DataBank SLC2 · Salt Lake County, UT · appeared
- 2026-06-16DataBank SLC4 · Salt Lake County, UT · appeared
- 2026-06-16Verizon · Salt Lake County, UT · appeared
The sheet itself
Every mapped facility, over the substations and power plants they have to connect to — 62,427 nationally. The coastline under them is drawn in the paper's own hairline rather than in a pen, because it is the one thing on this sheet that was not measured: it is dissolved from the same county boundaries that decide which county each facility belongs to. Everything in ink is a record.
The sheet covers the contiguous states, so it draws the contiguous share of that total rather than all of it. Alaska and Hawaii carry grid assets and no mapped data centres, so they are off it rather than empty on it; Puerto Rico's two facilities get an inset, and its grid, like theirs, stays off the paper. The figure the sheet itself is drawn from is in its description.
The rest of the roll
Each of these answers one question this sheet only summarises.
And a second, much deeper dataset: in one Arizona valley, Helios reads parcel transfers, permits and assessor records to argue where a data centre is being built — 13 candidate sites, each with its evidence chain and its confidence scored in the open. inferred