AI Meets Natural Hazard Risk: A Nationwide Vulnerability Assessment of Data Centers to Natural Hazards and Power Outages

Miguel Esparza, Bo Li, Junwei Ma, Ali MostafaviarXiv 2025 · DOI · arXiv

As the demand for Artificial Intelligence (AI) skyrockets, the physical infrastructure supporting it—data centers (DCs)—is increasingly coming into conflict with a changing climate and an aging power grid. These "factories of the 21st century" house the critical servers and cooling systems required for cloud computing and training Large Language Models. However, they are remarkably fragile; unplanned downtime can cost an average of $8,851 per minute.

While individual operators perform site-specific risk assessments, there has been a lack of a comprehensive, nationwide study overlaying existing DC locations with empirical data on natural hazards and grid stability. This research presents the first nationwide spatial assessment of U.S. data center vulnerability, identifying high-risk "hot-spots" and low-risk "cold-spots" to guide future infrastructure development.

Data Sources and Spatial Distribution

The study integrated three primary data sources to map the intersection of infrastructure and risk:

  1. Datacenter Map: Provided coordinates for 2,660 data centers across the contiguous United States.
  2. FEMA National Risk Index (NRI): Provided census-tract level data on six natural hazards: earthquakes, hurricanes, tornadoes, coastal flooding, riverine flooding, and wildfires.
  3. EAGLE-I (Department of Energy): A dataset of power service outages from 2014–2022, covering 3,044 U.S. counties.
Spatial distribution of 2,660 data centers across the contiguous United States.

Methodology: Identifying Spatial Clusters

The researchers employed a multi-stage spatial analysis approach to move beyond simple mapping:

  • Spatial Autocorrelation (Global Moran’s I): Used to determine if hazard risks and DC locations are statistically clustered rather than randomly distributed.
  • LISA (Local Indicator of Spatial Association): This technique identifies "Hot-Spots" (areas of high risk surrounded by high risk) and "Cold-Spots" (low risk surrounded by low risk).
  • Bivariate LISA (BI-LISA): Used to find correlations between two different variables across space—for example, identifying areas with both low earthquake risk and low power outage frequency.
  • Composite Vulnerability Score: A weighted formula integrating hazard data with physical and social factors like building value and community resilience.

Spatial Autocorrelation of Natural Hazards

The analysis revealed that natural hazard risks are not randomly distributed but exhibit significant spatial clustering. For instance, hurricanes, tornadoes, and earthquakes currently threaten the largest number of existing data centers. A significant portion of U.S. data centers are located in high-risk zones for hurricanes (34%) and tornadoes (29.8%).

Global Moran’s I and LISA results for six natural hazards.

Key Insights and Regional Winners

The study highlights a "Power-Hazard Paradox": some areas are naturally safe from disasters but suffer from unstable power grids. Grid stability is a major concern, with 18% of U.S. data centers located in "Hot-Spots" for total power outage events. Furthermore, the authors argue that "Community Resilience" is as vital as physical safety, as it dictates how quickly a region can recover after a disaster.

After calculating composite vulnerability scores (where 0 is the lowest vulnerability), the research identified the top three optimal locations for future DC development:

  1. Koochiching, Minnesota (Score: 0.091)
  2. Schoolcraft, Michigan (Score: 0.095)
  3. Houghton, Michigan (Score: 0.096)

Michigan and Minnesota consistently emerged as the states with the most low-risk, high-resilience counties, making them top contenders for the next generation of AI infrastructure.

Limitations and Future Outlook

While groundbreaking, the study acknowledges constraints such as the difference in data resolution (census-tract for hazards vs. county-level for power outages). Additionally, the model does not yet account for "network redundancy"—the ability of providers to shift workloads between states during a localized failure. Future research could integrate these operational fail-safes and more granular factors like proximity to fiber-optic backbones to further refine the vulnerability model.

Summary updated

BibTeX
@misc{esparza2025ai,
  author = {Miguel Esparza and Bo Li and Junwei Ma and Ali Mostafavi},
  title = {{AI Meets Natural Hazard Risk: A Nationwide Vulnerability Assessment of Data Centers to Natural Hazards and Power Outages}},
  year = {2025},
  eprint = {2501.14760},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.14760},
}
Figure