Compute map
Where America is putting its AI compute.
794 U.S. data center facilities, 1930–2026. Plotted on top of the projected labor-market tightness for each metro under the baseline scenario. Hyperscale campuses cluster where the workforce is already pressured. Colocation facilities follow the fiber.
Show
Hyperscale facilities
303
Built by Amazon, Google, Microsoft, Meta and similar — running the operator's own AI and cloud workloads.
Colocation facilities
220
Run by data-center landlords (Equinix, Digital Realty, CyrusOne) and leased to remote tenants — including the AI-compute customers who don't show up locally.
Peak buildout year
2024
98 new facilities opened — the highest of any year in the 96-year record.
Who's building
The five largest operators run more than half of every data center in the dataset.
- 01 Amazon Web Services 176
- 02 Unknown 143
- 03 Digital Realty 71
- 04 Meta 61
- 05 Google 59
- 06 Microsoft 57
- 07 Equinix 35
- 08 Quality Technology Services 23
- 09 Flexential 22
- 10 CyrusOne 19
- 11 NTT 13
- 12 CoreSite 10
- 13 Stack Infrastructure 7
- 14 Switch 7
- 15 Apple 6
- 16 Centersquare 6
About this dataset
What the dots mean — and what they don't.
Each dot is a single data center facility, plotted at its mapped lat/lon. Circle size is electrical capacity (MW) where reported; capacity is missing for 34% of facilities — those render at the minimum size. Color encodes type: hyperscale (operator-owned), colocation (leased), or unknown (cannot confidently classify from public data).
The background heatmap is the model's projected worker gap by metro at the baseline scenario × 5% wage ceiling, the same value you can change on the Atlas. Tighter metros sit in terracotta; metros at or near zero gap sit in cream. Out of respect for model uncertainty, the gap is shown as quantile bins, not exact numbers.
What this map is not. The visual co-location of dots and color is a correlation, not a causal claim. The Bahar & Wright working paper that assembled this dataset estimates causal effects using synthetic-control methods; that's a separate question, and the answers are nuanced. Here, the goal is descriptive: see where the compute is, see where the workforce is squeezed, ask the questions yourself.
The hyperscale-vs-colocation classification uses the underlying capacity provenance (whether MW data come from an operator-imputation model or from the public datacentermap registry). Some facilities are confidently one or the other; some genuinely sit in between.
Facility data: Bahar & Wright (2026, working paper). Not peer-reviewed; treat correlations as descriptive. Wage-pressure source: BW-LaborShortages GE model, see Methods.