Forecasts for abundant compute usually assume that suitable sites will be available when the hardware is ready. For a hyperscale project, the site may need hundreds of contiguous acres and a credible path to hundreds of megawatts. Finding both in the required schedule is not a simple property search.

Grid timelines illustrate the problem, although generation queues are not the same as data-center load connections. Berkeley Lab's 2024 queue study found that US generation projects built in 2023 took a median of five years from interconnection request to commercial operation. I use that as evidence that grid access requires early diligence, not as a site-specific forecast for a new data center.

Siting starts as search

Teams often evaluate a small set of parcels that somebody has already surfaced. The earlier question is which candidates deserve that expensive review. The inputs are scattered across utility and grid records, county GIS portals, satellite rasters, regulatory dockets, and local reporting.

I built datacenter-geo to search that wider field and retain the evidence behind each result.

What it measures

The current model covers 59 factors across eight domains: power, water, climate, land, connectivity, regulation, community, and supply chain. They include transmission proximity, water stress, terrain, flood exposure, connectivity, tax treatment, permitting, and local opposition.

The economic importance of a factor depends on the proposed facility. Tax treatment can materially affect a GPU-heavy project, while land price may dominate a smaller development. Community approval and water constraints can each block a site; the tool does not claim a universal ordering between them.

How the agents work

A prospector tiles a region and applies arithmetic filters such as slope, floodplain, protected land, and distance to mapped infrastructure. Survivors receive a written hypothesis. Domain analysts then look for evidence against that hypothesis, and a final red-team pass searches for reasons the site should be rejected.

Every datapoint records how it was obtained: a machine API, a government or utility file, cited web research, or an estimate. The score carries a confidence band so that precise measurements are not presented as equivalent to sparse reporting or modeled assumptions.

The public analyses currently include seven known sites. Two parcels in the same Virginia county had different approval outcomes, but their overall confidence bands still overlapped. The tool therefore declined to rank them.

That refusal is deliberate. The output is not a substitute for utility studies, land diligence, or local approval. It is a way to identify which parcels deserve that work and which claims still need a human source.

Sources

The repository contains the factor specifications and evidence ledger. Data adapters include ERA5-derived weather through Open-Meteo, WRI Aqueduct, Global Energy Monitor, OpenStreetMap via Overpass, PeeringDB, and Berkeley Lab's interconnection data.