What this is

The atlas maps 341 facilities in the physical supply chain of AI — mines, refineries, wafer plants, toolmakers, fabs, packaging and assembly lines, datacenters, exchanges, and power stations. Each site carries what it does, who operates it, why it matters, its status, and source links.

Where the data comes from

The dataset was assembled with AI assistance from public reporting, company disclosures, and trade press, then run through automated checks (no missing coordinates, no out-of-country coordinates). It is honest work, but it is not a census: capacities and dollar figures are best public estimates, coordinates are city-level rather than parcel-level, and the industry moves fast enough that some statuses will drift between snapshots.

Hero claims in the story — the numbers attached to named sites — get a hand-verification pass against their listed sources before this piece is declared finished. Until then, treat figures as drafts with sources attached.

Confidence flags

Every site carries a confidence flag. High (286 sites) means multiple independent sources agree on the facility's existence, operator, and role. Medium (55 sites) means the site is well-attested but some details — capacity, timing, exact role — rest on fewer or softer sources. The atlas carries 662 source links in total, one to three per site.

Chokepoint severity

Sites are tagged by how concentrated their step of the chain is: monopoly (48 sites — one indispensable supplier), duopoly, diversified, or not applicable. The tags drive the atlas filter and the red ticks on pins. They are judgments about the leading edge — many steps have trailing-edge alternatives that don't change the strategic picture.

Corrections

If you operate one of these facilities, or simply know better, I want the correction. Open an issue on GitHub or reach me through the site. Corrections ship in the next data snapshot with credit if you want it.

See also: the Deep Dive primer — the full technical reference behind the story.