Methodology

Grid Feasibility Index v1.0

The Grid Feasibility Index quantifies structural pressure on regional electricity infrastructure from announced AI compute demand at sub-national resolution (L1-L2, counties, , districts, etc.). It is calculated for each region as:

Available Capacity = grid headroom (installed generation minus peak demand) + substation bonus (proximity to high-voltage critical-class substations) − transmission penalties (regional bottlenecks) − water penalties (in cooling-constrained regions). Headroom is sourced from EIA Form 861 spatially joined to HIFLD utility service territories for U.S. counties; from Statistics Norway (SSB) Tables 08308 and 08313 for Norwegian fylker; and from IEA, IRENA, and GASTAT national totals disaggregated by population for Gulf regions.

Realistic Queue Pressure = announced project capacity (MW) × (1 − historical completion rate). Queue data comes from Berkeley Lab's “Queued Up 2024” interconnection database for the United States, Statnett capacity reservations for Norway, and announced project pipelines from Global Energy Monitor for the Gulf. Historical completion rates differ by region: 19.7% for PJM (U.S.), 40% for Norway, 55% for the Gulf, with an additional 30% geopolitical discount applied to UAE.

The output is mapped to five categories: low pressure (ratio ≥ 2.0), moderate pressure (1.0–2.0), high pressure (0.5–1.0), severe pressure (< 0.5), plus a separate “existing concentration” overlay for regions where data center clusters are already established (Loudoun, Fairfax, Prince William, Arlington, and Oslo).

Important interpretive note. The index measures structural pressure, not project approval likelihood. High-pressure regions can and do see active construction. The score quantifies the cost of that construction — in rate increases for residential customers, multi-year power delivery delays, and risk of stranded transmission assets — not whether projects proceed.

Validation Experiment: Prince William County

To test interpretation, we cross-referenced our Phase A index against actual under-construction data center projects in Prince William County (PWC), Northern Virginia, using the county's official interactive build-out dashboard.

PWC has a Grid Feasibility Index of 0.149 — severe pressure category. The index reflects PWC's structural position: the county has negative net headroom (−481.8 MW), meaning it already imports electricity from neighboring regions through the PJM transmission network, while its interconnection queue contains 970 MW of new project capacity awaiting connection.

As of 2026, 31 data center buildings are listed as under construction in PWC.

This is not a contradiction. It is the central observation of the project.

Three independent data points tell one consistent story:

1. The Grid Feasibility Index identifies PWC as the highest-pressure region in Virginia.
2. Dominion Energy itself acknowledges, in its 2024 Integrated Resource Plan, that contracted data center demand exceeds available transmission capacity by years; the utility constrains delivered load below contracted ESA amounts for periods of up to seven years.
3. Virginia residential customers experienced electricity rate increases exceeding $11 per month in 2026, attributed by PJM's independent market monitor primarily to data center load growth.

The construction continues despite the structural deficit. The deficit is real. Both facts coexist because power delivery to a built data center is decoupled from the building's completion: developers race to break ground in anticipation of grid expansion that takes years to deliver. The result is exactly what the index describes — high structural pressure manifesting as externalities (rate increases, multi-year delays, transmission stress) rather than as project denials.

The PWC validation confirms that the index is identifying the right phenomenon. It also reveals a known limitation: at county-level resolution, the index cannot distinguish between data centers connecting to substations with reserved capacity and those entering an indeterminate queue. Phase B addresses this through substation-level integration of HIFLD substation data and PJM interconnection queue data per substation.

Roadmap to v2.0

Three improvements are scheduled for Phase B (post-funding):

The methodology and underlying data are CC BY 4.0 licensed. Source code and data pipelines are available at GitHub. Detailed component breakdowns for each region are accessible through the interactive map by clicking on individual features.

Datacenter sources

Three independent datasets show where AI-relevant compute infrastructure exists, is being built, or is announced. Each is displayed as a separate toggleable layer on the map.

Frontier AI Datacenters

Source: Epoch AI Frontier Data Hub (CC-BY-4.0).

Coverage: 38 datacenters globally, of which 33 have verified geographic coordinates. The remaining 5 are known to exist but lack precise location data and are excluded from the map.

Methodology: Epoch AI identifies installations housing the largest training clusters by H100-equivalent compute. Coordinates are verified via satellite imagery where possible.

Last updated: 2026-05-04.

GPU Clusters

Source: Epoch AI GPU Clusters (CC-BY-4.0).

Coverage: 786 clusters globally; 598 have coordinates and appear on the map. The remaining 188 are known to exist but lack precise location data.

Methodology: Compiled from corporate disclosures, news reports, and regulatory filings. Each cluster carries a Certainty rating (Confirmed / Likely / Unlikely) and a Status field indicating lifecycle stage.

Last updated: 2026-05-04.

OpenStreetMap Datacenters (US)

Source: OpenStreetMap (ODbL).

Coverage: 1,317 facilities tagged as datacenters in US OpenStreetMap. Currently US-only — global OSM coverage is uneven and would require additional curation.

Methodology: Community-maintained tags building=data_center or man_made=data_center. Includes all datacenter types, not only AI-relevant. Useful as broad infrastructure context against which specialized lists (Frontier, Clusters) can be compared.

Note on overlapping records

A single physical datacenter may appear in multiple datasets — for example, a large hyperscaler facility may be listed in the Epoch Frontier set, the GPU Clusters set, and OSM. We display each dataset independently rather than deduplicating, because each source applies different criteria for what counts as a "datacenter" and at what threshold. Researchers can compare and triangulate across layers.

Note on geocoding precision

Some sources report coordinates at varying precision: street-level (precise), city-level (centroid of municipality), regional (state/province centroid), or country-level (national centroid). We encode this visually as point opacity on the GPU Clusters layer once Epoch begins publishing precision metadata. As of the current data version, all coordinates render at default opacity (treated as city-level equivalent).