Project · Interactive tool
Where Should the Next Data Center Go?
Why I built this
Following the news about the data center boom, I kept wondering how companies decide where to build. Power and connectivity come first, but a site also has to work for the people around it: labor markets, land and housing costs, natural hazards, water, local politics, and whether a community welcomes a new facility or pushes back with a moratorium. Commercial siting tools weigh some of this, but I could not find an open, transparent way to look at all of it together.
So I set out to see how far a light project could get with machine learning and openly available data alone. I assembled about 80 county-level conditions from public sources (Census, EIA, FEMA, USGS, NOAA, LBNL and others), matched them to the facilities that were actually built, and trained models to learn where new data centers go.
This is a lightweight public version.
A fuller version is available privately. It adds more detail than this page shows, including the power tariff cost calculator, side-by-side county comparison, facility locations on the map, and additional data sources. If you are interested, please reach out at kyumin.kim.econ@gmail.com.
Methods
- Count models of new builds per county (Poisson objective), LightGBM tuned with Optuna, chosen after comparing ten model families: penalized regressions, random forest, XGBoost, CatBoost and a neural net.
- Validated with spatial block cross-validation and a forward test: trained on 2016–2020 builds, scored on 2021–2026.
- SHAP values explain each county's score; a survival model of past grid-connection requests estimates today's wait for power.
What it suggests
- Open data goes a long way: trained only on 2016–2020, the model placed 45–50% of the facilities built in 2021–2026 in its top 10% of counties, four to five times chance.
- Power dominates: new generation waiting to connect and nearby plant capacity are the strongest signals, ahead of land or labor costs.
- "White space" exists: counties with strong fundamentals but few facilities, which the market has not reached yet.
- Local acceptance is the blind spot: moratoria and community opposition shape real projects, but public data on them is thin.
Below, every county in the contiguous US gets a 0 to 100 score. Choose what matters, switch deal-breakers on or off, and click a county to see the reasons behind its score. Everything runs in your browser on pre-computed results.