# US Data Centers - Geospatial Research Infrastructure

A comprehensive geospatial research project investigating the spatial concentration, infrastructure dependencies, and socioeconomic/environmental impacts of US data center locations.

## Documentation

- **[Database Tables](docs/database-tables.md)** - Complete database schema with table descriptions, column definitions, and SQL examples
- **[Research Ideas](docs/research-ideas.md)** - Future research directions, data improvements, and potential collaborations
- **[SQL Queries](docs/query_legiscan_bills.sql)** - Pre-built legislative analysis queries

## Project Overview

This repository implements a PostGIS-based analytical framework that integrates multiple data sources to examine:

- **Spatial concentration patterns**: Where are data centers located and why?
- **Infrastructure dependencies**: How do data centers relate to submarine cables, power grids, and watersheds?
- **Equity and impact**: Do data center host communities bear localized burdens while benefits are nationally dispersed?
- **Demographics**: Who lives in data center-hosting census tracts?
- **Environmental exposure**: What are the water, energy, and natural hazard exposures?

## Key Research Question

**Do data centers represent "concentrated costs / dispersed benefits"?** Host communities bear localized infrastructure burdens (power, water, land use) while cloud computing benefits are nationally dispersed.

## Major Findings

### Spatial Concentration
- **State level**: Top 5 states (VA, TX, CA, OR, OH) hold 51% of all US data centers
  - Virginia alone: 20.6% (378 of 1,833 facilities)
- **Tract level**: Top 1% of data center-hosting census tracts hold 14.6% of all facilities
  - Only 0.86% of data center-state residents live in a hosting tract
  - Per-capita burden is **115× higher** in host tracts vs. state average
- **Watershed level**: Half of all US data centers sit in just 15 of 2,139 HUC8 watersheds
  - Single watershed (Middle Potomac-Catoctin / Loudoun County): 12.8% of US facilities

### Demographics of Host Communities
Compared to the US average, data center host communities are:
- **Wealthier**: Median household income $103,623 (vs. $78,538, +32%)
- **More educated**: 49% bachelor's+ (vs. 35%, +14 pp)
- **More diverse**: 50% non-Hispanic white (vs. 58%), driven by high Asian share (13% vs. 6%)
- **Better connected**: 94.9% broadband (vs. 89%)

### Infrastructure Insights
- **89% of data centers are in metropolitan tracts** (vs. 80% of all US tracts) - only 1.11× over-index
- **Non-metro data centers (11%)** are dominated by hyperscalers:
  - AWS (67), Meta (22), Microsoft (10), Google (4) = 55% of non-metro facilities
  - 66% are in Oregon + Washington (Columbia River hydro corridor)
- **Grid saturation**: 4 states have >2/3 of generation within 50 km of a data center:
  - New Jersey: 83%, Nevada: 75%, Tennessee: 70%, Oregon: 68%
- **Hyperscaler energy strategies** (non-metro sites):
  - AWS: 114 GW wind + 66 GW hydro
  - Microsoft: 13 GW nuclear (Palo Verde co-location)
  - Meta: 16 GW solar

### Clustered vs. Isolated Facilities
Facilities in DBSCAN clusters differ significantly from isolated sites:
- **$35K income gap**: Clustered sites in tracts with median income $108K vs. $73K for isolated
- **+18 pp education**: 51% bachelor's+ vs. 33%
- **More diverse**: 25 pp less non-Hispanic white
- **2× energy infrastructure**: 89 vs. 40 generators within 50 km

### Submarine Cables
- **Data centers are NOT systematically closer to cables** than ordinary US cities
- Only 21.4% of data centers are within 100 km of a submarine cable landing point
- Largest clusters (Ashburn VA, Columbus OH, Iowa) are inland, driven by fiber/power/tax incentives, not cables

## Data Sources

### Primary Data Center Inventories
- **Curated Sample**: 1,489 facilities from web scraping + manual curation, geocoded via Census TIGER + Nominatim
- **OpenStreetMap**: 1,549 OSM features tagged as data centers (via Overpass API)
- **IM3 Model Data**: PNNL's Integrated Multisector Multiscale Modeling existing facilities
- **Master Table**: 1,833 deduplicated facilities merging all sources

### Geospatial Context Layers
- **US Census**: 2024 TIGER tract boundaries, ACS 2024 5-year demographics (85k+ tracts)
- **USDA RUCA 2020**: Rural-Urban Commuting Area codes for metro/micropolitan/rural classification
- **USGS HUC8 Watersheds**: 2,139 subbasin polygons for water-stress analysis
- **FEMA NRI**: National Risk Index with 18 natural hazard risk scores by census tract

### Infrastructure Layers
- **Submarine Cables**: 693 cables, 3,361 landing points (TeleGeography-style)
- **EIA Energy Data**: Operating generator capacity (4.7M monthly records, 2008-2026), facility fuel, state energy data
- **FCC Broadband Data**: Provider availability by location/block

### Additional Data
- **RDH Precinct Vote Data**: Election results for political-economy analysis
- **NOAA HMS Smoke Data**: Wildfire smoke exposure (2005-2025)
- **USDM Drought Data**: Drought severity
- **Utility Rate Tracker**: State-level electricity rate increases
- **LegiScan Legislative Data**: All US state + federal bills 2016–2026 (1.3M bills, 646 sessions), tagged for data center, ratepayer, grid, tax, and siting topics

## Repository Structure

### Core Python Scripts

**Data Ingestion** (`scripts/`)
- `scripts/load_postgis_data_centers.py` - Load curated data center CSV into PostGIS
- `scripts/load_postgis_osm_data_centers.py` - Fetch OSM data centers via Overpass API
- `scripts/build_master_data_centers.py` - Deduplicate & merge curated + OSM sources
- `scripts/load_postgis_internet_cables.py` - Load submarine cables and landing points
- `scripts/ingest_eia_energy_layers.py` - Ingest EIA energy data via API
- `scripts/build_watershed_huc8_tables.py` - Load USGS HUC8 watersheds
- `scripts/ingest_legiscan.py` - Download all US state/federal bills 2016–2026 via LegiScan API, tag for data center research topics

**Enrichment**
- `scripts/create_data_center_census_tract_table.py` - Join data centers to Census tracts with ACS demographics
- `scripts/build_fcc_bdc_broadband_connection_table.py` - Build per-facility broadband provider table
- `scripts/build_fcc_bdc_location_provider_aggregates.py` - Aggregate FCC BDC data by county/tract

**Analysis**
- `scripts/analyze_dc_tract_concentration.py` - Tract-level cost concentration analysis (Gini, HHI, demographic deltas)
- `scripts/analyze_cables_concentration.py` - Test if data centers cluster near submarine cables
- `scripts/make_data_center_map.py` - Generate Leaflet map of data centers
- `scripts/make_internet_cables_map.py` - Generate Leaflet map of data centers + cables

### Key Jupyter Notebooks
- `spatial_clustering_master_data_centers.ipynb` - DBSCAN clustering of data centers
- `cluster_analysis.ipynb` - Main demographic/energy/watershed analysis
- `fema_nri_data_centers.ipynb` - Join data centers to FEMA hazard scores
- `rdh_precinct_vote_data_centers.ipynb` - Join data centers to election data
- `usdm_drought_data_centers.ipynb` - Drought exposure analysis
- `hms_smoke_data_centers.ipynb` - Wildfire smoke exposure
- `enhanced_data_center_cluster_map.ipynb` - Generate enhanced cluster visualization

### Output Files
- `output/data_center_demographic_ruca_energy_summary.md` - Flagship analysis report
- `cables_concentration_report.md` - Cable proximity + cost/benefit concentration analysis
- `data_center_map.html` - Basic data center locations (Leaflet)
- `data_centers_cables_map.html` - Data centers + submarine cables (Leaflet)
- `output/enhanced_master_data_center_spatial_clusters_map.html` - Enhanced cluster visualization

## Technical Architecture

### Database
- **PostgreSQL 13+** with **PostGIS 3.x**
- Database name: `data_centers`
- See [database-tables.md](database-tables.md) for complete schema documentation

### Python Environment
- **Python 3.10+**
- Key libraries: `psycopg2`, `geopandas`, `shapely`, `scikit-learn`, `pandas`, `numpy`, `requests`, `folium`

### Data Formats
- CSV (raw data exports)
- GeoJSON (watershed/cluster geometries)
- Shapefiles (Census, USGS, FEMA inputs)
- HTML (interactive Leaflet maps)

### Configuration
Credentials stored in `~/.zsh_secrets`, loaded via environment variables:
- `PGWEB_*`: PostgreSQL connection
- `EIA_API_KEY`: EIA energy data
- `FCC_USERNAME`, `FCC_API_KEY`: FCC broadband data
- `RDH_USERNAME`, `RDH_PASSWORD`: Redistricting Data Hub
- `CENSUS_API_KEY`: Census ACS API
- `LEGISCAN_API_KEY`: LegiScan legislative data

## Quick Start

### Basic Rebuild Sequence

```bash
# 1. Load base data center data
python3 scripts/load_postgis_data_centers.py
python3 scripts/load_postgis_osm_data_centers.py
python3 scripts/build_master_data_centers.py

# 2. Enrich with context layers
python3 scripts/create_data_center_census_tract_table.py --replace-final
python3 scripts/load_postgis_internet_cables.py
python3 scripts/ingest_eia_energy_layers.py --category power
python3 scripts/build_watershed_huc8_tables.py

# 3. Run analyses
python3 scripts/analyze_dc_tract_concentration.py > output/tract_analysis.txt
python3 scripts/analyze_cables_concentration.py > output/cables_analysis.txt

# 4. Execute notebooks
jupyter notebook cluster_analysis.ipynb

# 5. Load legislation (all states, 2016-2026)
python3 scripts/ingest_legiscan.py --all
# Weekly refresh (skips unchanged sessions):
python3 scripts/ingest_legiscan.py --fetch --load
```

### Generate Maps

```bash
python3 scripts/make_data_center_map.py
python3 scripts/make_internet_cables_map.py
```

## Key Outputs

### Research Reports
- **Demographic, Energy & Watershed Analysis**: `output/data_center_demographic_ruca_energy_summary.md`
- **Submarine Cable Proximity**: `cables_concentration_report.md`

### Interactive Maps
- Data center locations with cluster assignments
- Submarine cable routes and landing points
- Energy infrastructure proximity
- Watershed boundaries with data center counts

### Data Exports
- `master_data_center_spatial_cluster_points.csv` - Data center points with cluster IDs
- `master_data_center_spatial_cluster_summary.csv` - Cluster-level statistics
- `output/master_data_center_huc8_watersheds.geojson` - Watershed polygons
- `output/master_data_center_map_context.csv` - Full context for mapping
- `output/master_data_center_state_energy_context.csv` - State-level energy statistics

## Data Quality Notes

1. **Incomplete power ratings**: Only 5.9% of data centers have power ratings (108/1,833)
2. **Operator fragmentation**: String variations ("Meta" vs. "Meta, Inc.", AWS variants) inflate distinct-operator counts
3. **45 facilities** use city-precision fallback coordinates (approximate tract assignment)
4. **7 facilities** failed RUCA join (Puerto Rico / non-US)
5. **Broadband subscribers** are a coarse benefit proxy (actual cloud users are global)
6. **EIA longitude correction**: 2008-2010 generator coordinates had sign errors, corrected in flat-table build

## Known Limitations

- **Power capacity**: Only 5.9% populated - nearby EIA generator capacity used as proxy
- **Operator strings**: Need deduplication (50 of 190 non-metro facilities have null operator)
- **Benefit measurement**: Broadband subscribers are an imperfect proxy for cloud computing benefits
- **Universe**: Limited to 46 DC-host states (excludes DC-free states from ACS comparison)

## Research Ideas & Future Work

See [research-ideas.md](research-ideas.md) for detailed next steps and potential research directions.

## Project Status

This is a **mature, publication-ready geospatial analysis infrastructure** combining authoritative government datasets (Census, EIA, USGS, FEMA) with novel data center location data to test political economy and environmental justice hypotheses.

The "concentrated costs / dispersed benefits" hypothesis is operationalized and tested with rigorous spatial statistics (Gini coefficients, HHI indices, Mann-Whitney tests).

## License

Research data compiled from public sources. Please cite appropriately if used in publications.

## Contact

For questions about this research project, please contact the repository owner.