AI Buildout Frontier

Project Details

Problem

Decisions about AI, from national compute strategies to grid planning and verification agreements, lean on numbers like how much compute sits where and how much power a site draws. Few of these numbers are available in public records and can be verified by a third party, and nobody has measured how few. Where they exist, they sit in national registers in different languages and formats. Some are getting harder to see. In 2025–2026 a U.S. federal portal of infrastructure data closed without a stated reason, France's grid operator removed line coordinates citing public security, and the EU's data-center reporting publishes only national aggregates.

Current status

Live at aibuildoutfrontier.org since October 2026: a globe, five U.S. energy layers by county, and France by département and region, including 43 data centers confirmed in the national permit register. Three datasets are on Zenodo with DOIs.

A pilot checked 35 U.S. data centers against four open channels. EPA's national facility registry found 23 of 25 sites drawn at random from Virginia's permit list, but only 6 of the 10 largest AI data centers elsewhere, mostly under names that say nothing about who runs them. EIA's inventory of power plants contains none of the 35. Thirty-five sites are enough to see a pattern and too few to measure it.

Work so far

The first version, in spring 2026, was a grid index for Virginia's 133 counties and cities. Tested against its own raw inputs, the index turned out to be mostly local demand under another name. I retired it and rebuilt the project as data layers named after their sources, where every cell is a value: "checked and not found" or "not checked".

I'm not a programmer. I write the specs and the rules; AI agents in separate roles write the code, and an agent that didn't write the rules re-checks samples against the original documents.

Expected outputs (next 12 months)

• A world layer of data centers showing which open sources see each site: Epoch AI's catalog, the Atlas of Data Center Politics, and national registers where they exist.
• The U.S. pilot published as a public layer.
• An observability layer by country. For each of the nine channels, it records found, not found, or not checked, with a link, a date, and a license verdict. There is no score.
• Datasets for more countries after France.

Every layer can be downloaded (GeoParquet, GeoJSON, CSV) and cited by DOI.

Links

• Site: https://aibuildoutfrontier.org
• Code: https://github.com/Geofuturist/ai-buildout-frontier
• Data (France, latest): https://doi.org/10.5281/zenodo.23336103