Between 2000 and 2025, 38 U.S. states adopted tax incentives to attract data center investment. We provide causal evidence that these incentives increased state-level data center construction, with effects concentrated among large-scale facilities for cloud computing and, more recently, large language models. We develop a quantitative model endogenizing development location decisions and state tax incentives. The model implies strong fiscal competition among states: a coordinated repeal of incentives would improve total state-government welfare. It also implies large expected economic gains from data centers. We explore job market spillovers and find null results for data centers’ impact on local tech-related employment.

States and counties are offering large incentives to attract data centers, yet it remains unclear whether host communities benefit from the buildout of AI infrastructure. We find no detectable average effect of data-center load on residential or commercial electricity prices, and no broad gains in local revenue, employment, or IT spending. These results suggest that host communities help finance the buildout of AI infrastructure without receiving broad local gains. We argue that the historical null in electricity prices is consistent with supply-side absorption: during our sample, load growth appears to have landed largely on slack segments of local supply curves. Private procurement and behind-the-meter contracts matter mainly for forward-looking scenarios, not for explaining the historical null.

We construct a novel dataset that links mortgage records with detailed information on datacenter characteristics and tenant contracts. We show that banks with high securitization propensity, those more likely to sell, syndicate, or securitize loans, exhibit significantly weaker underwriting standards than banks that retain loans. Properties financed by high securitization propensity lenders carry substantially higher leverage, measured by debt per square foot, debt-to-rent ratios, and loan-to-value ratios. These patterns persist in matched samples and when restricting attention to single-property loans, indicating that they reflect systematic differences in screening rather than deal structure. Our findings extend evidence on originate-to-distribute moral hazard from residential mortgages to digital infrastructure finance, a rapidly growing market characterized by technological obsolescence risk and by collateral values that we show are meaningfully lower than total invested capital.