Every forecast for abundant intelligence assumes the compute arrives. Fabs are the story people tell. The constraint I keep hitting is duller: a few hundred contiguous acres, flat, unflooded, with a credible path to 300 MW before 2030.

That path is the hard part. In most US interconnection queues it runs four to seven years. In India it is a DISCOM sanction and open-access approval. In China it is whether you sit inside a designated East Data West Computing hub. Land is cheap next to power that arrives on time.

Siting is prospecting, not evaluation

The industry treats siting as evaluation: hold three candidates, rank them. That is backwards. The real question is which twenty of two thousand deserve a look, and no database answers it. The inputs sit in grid queues, county GIS portals, satellite rasters, regulatory dockets and local newspapers.

So I built datacenter-geo, a harness that goes hunting instead.

What it measures

Fifty-nine parameters across eight domains: power, water, climate, land, connectivity, regulation, community, supply chain. Transmission distance and substation headroom. Basin stress and withdrawal rights. Wet-bulb temperature, which is what actually sets cooling cost.

The decisive ones are rarely the obvious ones. A sales tax exemption on equipment outweighs land price by an order of magnitude when the equipment is ten billion dollars of GPUs. Community opposition kills more American projects than water scarcity does.

How the agents work

A prospector tiles a region and eliminates most of it arithmetically — slope, floodplain, protected land, distance to transmission. Survivors get a written, falsifiable hypothesis. Eight domain analysts try to disprove it in parallel. A red team runs last with one instruction: kill this site. Without it the analysts drift optimistic. They find reasons a site works, because that is the shape of the task you handed them.

The part that matters

Every number carries how it was obtained — live API, government file, web research, or estimate. A score of 82 built on satellite measurements is a different asset from 82 built on news articles, and most tools present them identically.

I ran it against seven known sites. Two parcels in one Virginia county: approved unanimously, rejected four to three. They diverge sharply on community, and the harness still refuses to rank them, because the confidence bands overlap.

A model that cannot say "I don't know" is not measuring anything.

That refusal is the product. Investors do not need another confident number. They need to know which parcels deserve a human, and how thin the evidence is when one arrives.

References

Source, factor specifications and the full evidence ledger for every run: github.com/not-vibhu/datacenter-geo. The seven analyses are browsable at geo.vibhutripathi.com.

Measurements come from open data: ERA5 reanalysis via Open-Meteo for wet-bulb and free-cooling hours, WRI Aqueduct 4.0 for basin stress, Global Energy Monitor for generation and retirements, OpenStreetMap power infrastructure via Overpass, PeeringDB for exchanges, and LBNL Queued Up for interconnection timelines. Forty-eight of the sixty-nine registered sources need no API key.