Inspiration
Data centers are being built at record speed, and each one can use as much power and water as a small city. Communities usually find out after the site is already picked. We wanted a tool that weighs the environment and nearby residents before a location is chosen.
What it does
Atlas maps all 3,109 counties in the lower 48 and answers two questions: where the industry is likely to build next (a model trained on 3,245 real data centers), and where it should build, based on grid carbon, water stress, endangered-species habitat, community vulnerability, power cost and infrastructure. Crossing the two gives a Recommended list, a Watchlist of likely-but-harmful places, and Opportunity counties the industry is overlooking.
How we built it
We combined public data from the EPA, EIA, WRI Aqueduct, FEMA, USFWS, the Census, OpenStreetMap, PeeringDB and Epoch AI into a county-level dataset with Python. We trained LightGBM and logistic regression models with SHAP explanations, and served it all through a Flask app with an interactive Leaflet map.
Challenges we ran into
Real labels were messy: build dates were sparse, some addresses wouldn’t geocode, and several government services had broken links or rate limits. Our first model was no better than “bigger counties get more data centers.” That pushed us to focus on hyperscale and AI campuses and to test the model fairly against counties of similar size.
Accomplishments that we’re proud of
Using only data available in 2021, the model ranked the counties that later got Google, Microsoft, Meta, xAI and CoreWeave campuses in its top 3–7%. We also tested it honestly against simple baselines instead of overclaiming.
What we learned
A model has to beat a simple rule like “pick the biggest county” before it means anything, and fair validation matters more than an impressive headline number. We also learned how much public environmental data is freely available, and how hard it is to stitch together.
What’s next for Atlas
We want to add parcel-level data, transmission lines, water-use estimates by cooling type, and local permit and zoning records. We also want to turn Atlas into a tool for local governments and community groups, so they can evaluate a data center proposal before it arrives.
Log in or sign up for Devpost to join the conversation.