Methodology

Release: abf-layers, version 0.1.0 (DOI: 10.5281/zenodo.23295409) and abf-boundaries, version 0.1.0 (DOI: 10.5281/zenodo.23295311) · published 15 October 2026

What this site is

AI Buildout Frontier is an open observatory of data layers on the physical side of AI. This first release covers the U.S. power system at county level: what generates electricity, how much is consumed, where transmission lines run, and what is waiting to connect to the grid. The next releases add France, data centers themselves, and the restrictions placed on them around the world.

The project started from a research question: which of the indicators that AI governance relies on can actually be collected from open data and checked by someone else? The map is how we work on that question. Every layer on it can be downloaded, cited and rebuilt from its source.

How to read a layer

A layer is named after its source, not after what it might stand for. “Generation · EIA-860 · county” tells you what EIA-860 lists for a county. It does not tell you how much power that county can supply. We keep the two apart because a layer states a fact. An estimate would need its own assumptions, and those belong in a methods text, not in the data.

Blank is not zero. We separate three states: there is a value; we looked and found nothing; we have not looked. A county with 0 MW of generation is a statement that EIA-860 lists no operating generator there. A hatched county is not covered by the layer. Sometimes a covered county has an empty field because the value does not exist, for example the highest voltage in a county with no line. The pop-up then says why.

Each layer keeps its source's license. The project has no single data license. Our own code is under the MIT license, and our own compilations, such as the boundary set, are under CC BY 4.0. If you combine layers, join them by unit code (unit_id or fips). The combined file carries the obligations of every license in it.

Every layer shows its dates: the period the source covers and the day we built the layer. The map shows them too, because the sources update on different schedules, from yearly to never again.

Every layer states how it was checked. In this release all energy layers are published as the source gives them (source_as_is). Our scripts place, sum or measure the source's own records and run the automated checks listed below. No value was re-checked by a person or by an AI agent. Later layers that rest on judgment, such as deciding which registry entry is a data center, will state how much of that judgment a person or an agent re-checked.

Layers in this release

Every county layer lists all 3,144 counties and county-equivalents of the United States. Values cover 3,109 of them: the 48 contiguous states and the District of Columbia. Alaska and Hawaii are listed but marked not covered, because they need their own map projection. In the queue layer the nine Connecticut planning regions are not covered either (see below).

Generation · EIA-860 · county

• Edition. Form EIA-860, 2025 final · data through 2025-12-31. EIA publishes a new edition every year.
• What is counted. Only generators with status “OP”, operating. Generators on standby or temporarily out of service are left out. The value is net summer capacity as reported to EIA. We count both generators and plants, because one plant often has many generators.
• How plants are placed in counties. By their coordinates: first a plain point-in-polygon test. A few plants on the shore fall just outside every county, because the Census cartographic boundaries are trimmed to the coastline. For those we take the nearest county within 1 km. When we first built the layer, on the 2024 edition, we checked this rule against EIA's own county field for the 8 plants it affected, and all 8 matched. Three offshore wind plants, Block Island, South Fork Wind and the Coastal Virginia Offshore Wind pilot, lie farther out. Together they have 171.3 MW and are not assigned to any county.
• What it does not show. Form EIA-860 covers power plants with a combined nameplate capacity of 1 MW or more. Emergency and standby generators at data centers and other buildings are generally not reported to it. Generators that entered service after 2025-12-31 are not in this edition.

Demand · NREL OEDI 8562 · county

• Edition. OEDI submission 8562, submitted 2025-11-05, page last updated 2026-06-02; year 2023 of the 2016–2023 series · data through 2023-12-31. The dataset is static: a later year would need a new submission.
• What the value is. For each county, the highest hourly electricity demand in 2023, in MW. 2023 is the last year of the dataset's 2016–2023 series.
• How the dataset is modeled. Utilities do not meter or publish demand by county, so the dataset estimates it. Retail electricity sales reported by utilities are totaled by state and allocated to counties by each county's share of the state population. The hourly shape comes from the reported demand of the surrounding balancing-authority regions. Power plants' reported direct use and estimated self-consumption of rooftop solar are added separately.
• What it does not show. Measured demand at any place. Because sales are spread by population, a large load that does not follow population, such as a data center campus in a county with few residents, is spread across its state rather than placed in its county.
• Attribution. Kodi Obika, Wesley Cole and Marie Rivers, National Laboratory of the Rockies (NLR), DOI 10.25984/3366592. The laboratory was called the National Renewable Energy Laboratory (NREL) until December 2025; the layer keeps the name NREL, under which the dataset was announced.

Transmission lines · HIFLD · county

• Edition. HIFLD archive of 2026-02-26; newest record 2024-09-26; publisher metadata edition 2024-09-30 · data through 2024-09-30. The source is no longer updated.
• What the values are. Kilometers of line inside the county by voltage class, their total, the highest voltage present, whether the county has any line at all, and what share of its line length has no recorded voltage. The classes are 69, 115, 138, 230, 345 and 500 kV, 735 kV and above, and one class for lines below 69 kV or without a recorded voltage.
• Inside, not across. We measure line length inside each county. We do not count lines crossing a county boundary, because a boundary shifted by a kilometer can flip that count while the length inside barely changes. The kilometer columns are rounded to whole kilometers, so their total can differ from the true length by up to 4 km.
• Empty values. Twelve county-equivalents have no HIFLD line inside them: four independent cities in Virginia, the islands of Nantucket and Martha's Vineyard, and six rural counties. In two more, Hamilton County, Iowa, and Clare County, Michigan, the line inside is shorter than 0.5 km, so the total rounds to 0 km while “has any line” says yes. Where a county has lines but none with a recorded voltage, the highest voltage is left blank.
• Where the data comes from. HIFLD Open, the public portal that distributed this file, was discontinued by the U.S. Department of Homeland Security on 26 August 2025; the announcement gave no reason. We use an archived copy of the national file: the copy is dated 26 February 2026, and its newest record is from 26 September 2024. Other public archives of the file exist, for example on DataLumos (DOI 10.3886/E240591V1). Which archive our copy came from was not recorded, and we have not matched it to those archives record by record. So we publish our copy itself next to the layer, and anyone can check our processing against it.
• License. Public domain (publisher statement in HIFLD metadata; produced by ORNL for DOE). Our archive carries no dataset metadata of its own, so the statement is taken from a copy of the publisher's metadata for the same edition, kept by the Penn State Data Commons. The lines were produced by Oak Ridge National Laboratory, which a contractor runs for the Department of Energy. That is why we rely on the publisher's statement and not on the rule for works of federal employees.
• Was the geometry withdrawn? Before we publish the location of infrastructure, we check whether its publisher withdrew it for security reasons (see “What we do not publish”). HIFLD did not: no reason was given for closing the portal, and the U.S. Energy Atlas of EIA still offers the transmission-line file as an archived layer.
• What it does not show. How much power can flow into a county. An earlier version of this project estimated that from line ratings, and we dropped the estimate (see “Why there is no index”).

Transmission lines · HIFLD · segments

• What it is. The 93,880 line segments of the same archived file inside the contiguous United States. Each keeps its HIFLD ID, voltage, voltage class, owner and status (in service or not available).
• Voltage band. Each segment also gets a band shared by all transmission layers of the project: below 100 kV, 100–199, 200–299, 300–399, 400–599, 600 kV and above, direct current, unknown. The band is computed from the recorded voltage; the class is used only when no number is recorded. A segment without a usable voltage is in the band “unknown”. The direct-current band is empty in this release, because no segment in the file is marked as direct current.
• Edition, license and provenance are the same as for the county layer above.

Interconnection queue · LBNL Queued Up · county

• Edition. Queued Up 2026 data file · data through 2025-12-31. A new edition appears every year.
• What the value is. Megawatts and number of projects with an active request to connect to the grid, by the county where the project would be built. All grid operator regions are included. When a request has more than one capacity value in the file, we take the largest, not the sum.
• What is left out. Requests with a capacity of zero or less; requests outside the contiguous states or with no state given; requests without a county code; and requests in Connecticut. The source codes Connecticut projects by the state's former counties, which do not map onto the planning regions the Census has used since 2022. So the nine regions are marked not covered rather than shown as zero. The layer's known gaps list each excluded group with its count and megawatts. Together, the groups account exactly for the difference between the layer and the source file.
• What it does not show. These are power plants and batteries asking to connect, not electricity users. Most requests are never built. There is no public county-level queue of new large loads such as data centers: utilities and grid operators collect that information, but in the cases we checked they keep it confidential.
• Attribution. “Lawrence Berkeley National Laboratory and GridTracker”, as the license asks. This is the jointly released data file under CC BY 4.0. It is not the GridTracker website, whose terms do not allow reuse.

Administrative boundaries · ADM0, ADM1, ADM2

• One frame for the whole world. Natural Earth at 1:10 million, version 5.1.2. Where borders are disputed, we use the U.S. government's point of view, because the main sources in this project are U.S. federal. The boundaries do not express a position of the project. Five disputed areas belong to no country in this view: Abyei, the Ilemi Triangle, the Siachen Glacier, the Southern Patagonian Ice Field and Rincón de Artigas. They are drawn with a neutral fill, not left blank like the sea.

• Levels follow each country's own structure, not whatever a source calls “admin-1”. This release:

  Country	First level	Second level	Source
  World	242 countries and territories	—	Natural Earth
  United States	51 (states and the District of Columbia)	3,144 counties and county-equivalents	U.S. Census Bureau cartographic boundaries 1:500,000, 2023
  France	18 regions	101 departments	Natural Earth, merged
  United Kingdom	4 countries	—	Natural Earth, merged
  Norway	15 counties and Svalbard	357 municipalities	Kartverket, CC BY 4.0, reference date 1 January 2026

  Other countries show national borders only, until we check their subnational divisions against an official list. An outdated division on a site about checkable data would be worse than none.

• Fitting national data to the frame. National boundaries of the United States and Norway are fitted to the frame along land borders with neighbors only. The coastline stays national.

• Two geometries. Calculations use the official boundaries at full resolution: for the United States, the Census 1:500,000 files. The map shows a simplified version. Both carry the same unit codes, so what you see is what we counted on. The subdivisions of France and the United Kingdom come from Natural Earth and are generalized: they are not the official national boundaries.

• Unit codes. Countries: ISO 3166 three-letter codes (USA). Subdivisions: {country}-{level}-{national code}, for example USA-1-51 (Virginia) or USA-2-51059 (Fairfax County).

• Checks run at every build. No overlaps between units. No gaps except a reviewed list of seas, disputed areas and slivers. Every unit inside its parent. Unit counts as expected. Subdivisions are published only for countries whose entry in the boundary registry has been verified. The build ran twice with byte-identical results and was reproduced on a second machine.

• License. CC BY 4.0 for the boundary set. Each unit carries its own source and license.

Counties, county-equivalents and cities

“County” on this site means a county or county-equivalent as the Census defines it. In Virginia that is 133 county-equivalents (95 counties + 38 independent cities). Independent cities are separate units: St. Louis city and St. Louis County in Missouri are two rows, not one. Connecticut appears as 9 planning regions, as in the 2023 Census files, not as its 8 former counties. Areas are computed in equal-area projections.

Downloading and citing

• Formats. Every layer comes as GeoParquet (the main format, readable by GeoPandas and DuckDB), gzipped GeoJSON and CSV without geometry. Map tiles are for display only.
• Joining. Join on unit_id or on the five-digit fips code. Read codes as text, or leading zeros disappear.
• Stable links. Files sit at fixed, versioned addresses. A new version gets a new address, and old ones keep working. The current version is listed in manifest/latest.json with file sizes and checksums.
• Citing. Cite the layer, the version and the DOI: 10.5281/zenodo.23295409 for the energy layers, 10.5281/zenodo.23295311 for the boundaries. Each layer card gives a ready citation, and each pop-up gives one for the single record. Please also credit the original source as its license requires. The attribution text of each layer is in its metadata and on the Sources and licenses page.

What we checked

On every build, for every layer:

• every input file is fingerprinted (SHA-256), and the fingerprint is published in the layer's metadata;
• if an input is missing or incomplete, the build stops, so a failed download cannot turn into zeros;
• field names and types are checked against the layer's declared schema;
• every county layer lists each unit of the frame exactly once, and every unit code exists in abf-boundaries;
• the totals in the published files equal the totals of the build.

Against the source:

• Generation. The capacity of operating generators in the EIA-860 workbook equals the layer total plus the three offshore wind plants outside every county.
• Interconnection queue. The active requests in the Queued Up workbook equal the requests in the layer plus the excluded groups listed in its known gaps, both in number and in megawatts.
• Transmission lines, county. The kilometer totals are checked against the “has any line” field. The two counties where they disagree because of rounding are listed in the layer's known gaps.
• Transmission lines, county and segments. A script checks the license statement against the copy of the publisher's metadata.
• Transmission lines, segments. Every segment ID is unique.

Against our earlier Virginia pipeline. When we first built these layers in September 2026, on the 2024 edition of EIA-860 and the 2025 edition of Queued Up, we compared them with the pipeline behind version 1.0 for Virginia's 133 county-equivalents:

• demand matched in all 133;
• the queue differed in 1, explained by including all grid regions instead of one;
• generation differed in 10. These are 5 pairs of neighboring units, together 0.7% of Virginia's capacity: plants near a boundary landed on the other side, because the old and new boundary files are drawn slightly differently.

Published files do not change. Once a version is published, its files are fixed, and every new publication checks them by checksum. Corrections come as a new version.

Known limitations

• Generation covers generators operating at the end of the edition year. Generators that entered service later are missing until the next edition.
• Demand is modeled and spread by population, so it may misplace large single loads.
• Transmission comes from an archive whose newest record is from September 2024, and the source is no longer updated. Some lines have no recorded voltage.
• The queue lists requests, not plants that will be built, and generation, not demand.
• Alaska and Hawaii are not covered, and in the queue layer neither is Connecticut.
• Boundaries outside the United States, France, the United Kingdom and Norway are national only and generalized at 1:10 million.

What we do not publish

• Vulnerability of specific facilities: how they could be reached or disabled, single points of failure, security.
• How long a specific facility could run without the grid: fuel, permitted generator hours, backup configuration. Such figures appear only as totals for regions or classes, never for fewer than three facilities.
• Locations withdrawn for security. If the publisher of infrastructure locations withdraws them and gives security as the reason, we do not republish older copies. We publish totals by area instead.
• Substations from OpenStreetMap. We built this layer, but it is not in this release: coverage is uneven, and its license (ODbL) would carry over to any file it is merged into.
• A composite score. See below.

Why there is no index

Version 1.0 of this site, in April 2026, published a “grid feasibility index” for Virginia's 133 county-equivalents. While building a national version, we checked the index against its raw inputs and dropped it.

• The transmission term was demand under another name. In 128 of the 133 units it hit a cap set at half of adjusted peak demand. In those rows it was exactly 0.5 × demand, so its 0.9995 correlation with demand was an identity, not a finding.
• The cap threw away most of the transmission signal. Across those 128 rows the median import estimate fell from 1,281 MW to 41 MW (from 1,230 to 38 MW across all 133), and 94.5% of the total by sum.
• The “severe” band held only artifacts. Five units sat at −1.0: places with no generation and no transmission lines at all. Forty more sat at exactly −0.5, the floor the formula could not go below.

An index asks readers to trust its author. Layers ask them to trust EIA, NLR, HIFLD and LBNL, and then let them check. The version 1.0 methodology is archived at https://github.com/Geofuturist/ai-buildout-frontier/blob/gfi-v1.0/app/methodology/page.tsx.

How this was built

One researcher, working with AI coding assistants. The code is public on GitHub (https://github.com/Geofuturist/ai-buildout-frontier) under the MIT license. AI tools helped write the code and these texts, but they are not a source: every value comes from a file listed in the layer's metadata, and the checks above are scripts, not model output.

Found an error? Open an issue on GitHub. Corrections are listed with the version that fixes them.