Where the Machines Land: Mapping the Physical Limits of the AI Buildout

This summer a set of before-and-after satellite images went viral, appearing to show data centers rising over cleared forest. The reaction split immediately: outrage on one side, and on the other the objection that the images compared different seasons and measured the wrong thing anyway. It was a legitimate question asked with no data behind it - which is exactly the kind of problem geospatial data science exists to settle. So I went and measured it. In this talk I walk through what I found, across three questions. First: what do these facilities actually replace? I built season-controlled Sentinel-2 NDVI composites for the largest U.S. data centers, each matched to its own peak-greenness month and differenced against its own surroundings, so weather and crop cycles cancel out. The answer turns out to be mostly farmland and already-developed ground — forest is a small minority of the converted area. Forest clearing is real, and where it happens it produces the sharpest signal in the whole sample, but among the biggest facilities it's the exception rather than the rule. Second: how much room is left? Using an H3 hexagonal grid over the contiguous U.S., I combined grid adjacency, generation capacity, climate normals, flood and wetland exposure, terrain and protected land, and inferred feasibility from where hyperscale facilities have actually been built rather than from demand forecasts. Two independent unsupervised models converge on the same answer: the national feasible envelope is in the tens of gigawatts, not the hundreds. That's a physical ceiling, and no amount of capital moves it without changing the energy system underneath. Third, the same NDVI break-detection logic works in near-real time, which means you can watch construction happen before it's announced. And by laying IUCN species habitat data over the feasibility surface, I'm asking what I think will be the next siting fight — how much of the technically buildable land sits on top of habitat that matters. Everything runs on free, open data: Sentinel-2 through STAC, cloud-optimized GeoTIFFs read in place, H3, kernel density and similarity models. I'll show the maps live rather than on slides, and the methods are reproducible end to end — you could point the same pipeline at your own region on the way home.
Format: TECHNICAL TALK, RESEARCH TALK
Location: ROOM B
Speakers: Milan Janosov
Level: Intermediate
Audience: Registered participants