Spatial Non-Stationarity and Wealth Accumulation: Multiscale Regression of Locational Endowments in a Heterogeneous Metropolitan County

<title>Abstract</title> <p>The rapid expansion of hyperscale data centers in Texas has intensified conflicts between computational infrastructure growth and regional water security, culminating in San Marcos' 2026 zoning ban, the first such prohibition by a Texas city. This study develops a hybrid GeoAI and Multi-Criteria Decision Analysis (MCDA) pipeline to identify legally viable data center sites within the Guadalupe River Basin, which overlies the sensitive Edwards Aquifer recharge zone. The six-stage framework integrates Boolean pre-filtering for legal and geotechnical exclusions; unsupervised Isolation Forest suitability scoring with permutation-based feature weighting; spatially constrained regionalization via an Automatic Zoning Procedure; Spatial Durbin Model validation of spatial spillovers; and zonal Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) ranking. Applied to 16,041 census blocks, Boolean pre-filtering excluded 33.1\% of the basin on legal and geotechnical grounds, leaving 10,736 eligible blocks. Isolation Forest scoring, necessitated by the absence of existing data centers within the basin boundary, produced a right-skewed suitability distribution (mean = 0.218, median = 0.173) and objective, algorithmically derived criterion weights. Partitioning eligible blocks into five ecologically coherent zones improved suitability discrimination, while the Spatial Durbin Model confirmed substantial spatial spillovers (R-squared = 0.703) and residual spatial autocorrelation justifying zonal ranking. Final closeness coefficients, ranging from 0.154 to 0.977 across 9,034 blocks, identify site shortlists that direct development away from the Edwards Aquifer recharge zone and flood-prone lowlands. Replacing subjective expert weighting with empirically derived machine-learned criteria, this framework offers regulators and grid planners a highly reproducible, spatially explicit template for resolving water-energy conflicts in hydrologically constrained basins. The pipeline demonstrates that unsupervised learning can effectively guide infrastructure siting in regions lacking positive training examples, providing a scalable approach applicable to water-stressed regions across the United States. JEL Classification: C21 , C38 , D31 , R14 , R23 , R31</p>