Global Compute Infrastructure Map

The publishable corpus holds 4,226 facility records. 586 of them have no coordinates, so there is nothing to place, though the country is usually known and the operator often named. 1 more has coordinates and is withheld from the plate by two drawing rules: a record sharing a campus with another collapses into that campus’s single pin, so one site does not read as several, and an OpenStreetMap record above 70° north carrying no operator, no power figure, no chip count and no description is dropped as noise. 587 records in total are held and not drawn.

A further 316 records never enter that corpus. Each comes from a licensed vendor feed whose terms permit analysis and forbid republication, so every public page here reads a filtered set that leaves those rows out. Plate II draws them as an outline: counted, with everything else withheld.

Three populations run through this page: 3,639 placed marks, 4,226 records in the publishable corpus, and a raw facility table larger than either, which carries the licence-withheld rows. A national compute share computed off the pin count divides by the first. Plate II puts all three in one frame, at one scale.

A sparse region on Plate II is sparse for one of two reasons: there is little compute there, or there is little record of it. An external proxy would settle it, since a country importing accelerators plausibly runs them, and this platform holds bilateral semiconductor trade. Those rows are HS 8542 lines, which is every integrated circuit including the ones in phones and cars, with no data-processing-machine line to isolate. Joining on them would convert “imports chips” into “runs data centres”, so no join was built.

A source’s coverage is the geography of whatever it was compiled for. 90 per cent of every placed mark comes from 4 general registries: participation-based feeds, where a site appears because someone listed it or mapped it. None of them was compiled to find compute. The compute-scoped catalogues, where a site is on the map because somebody went looking for compute and found it, place 80 of 3,639.

So a sparse region is a region this portfolio does not reach, and the claim is falsifiable: add a source with reach there and the region fills. Plate III draws the portfolio.