Replication data for: Time to power paces AI infrastructure: Industrial policy and scarcity rents in the United States
Description
This dataset contains the replication package for "Time to power paces AI infrastructure: Industrial policy and scarcity rents in the United States," submitted to Structural Change and Economic Dynamics (VSI: Strategic technologies and industrial policies). The package reproduces every table and figure in the paper. It contains: (1) Raw inputs as retrieved: - Epoch AI frontier data-center census and construction timelines (CC BY 4.0) - EIA Electric Power Annual, Tables 2.8 and 2.10 (state industrial prices and retail sales, 2025) - EIA natural gas marketed production by state, 2024–2025 - HIFLD electric power transmission lines (345 kV and above) - EIA interstate and intrastate natural gas pipelines - U.S. Census Bureau cartographic boundary files and 2024 county population estimates (2) Constructed data: - Geocodes for the 77 US frontier AI campuses - State-to-market crosswalk (ERCOT, other organized markets, non-RTO) - 5,000 uniform control points - 345 kV network nodes and distance variables - Time-to-power intervals (3) Code: Python scripts 00–08 (data build, symbolic theory checks, PPML siting regressions, Monte Carlo forecast, forecast verification, figures, spatial case–control and point-pattern analysis, maps), run in sequence by run_all.sh. Script 00 retrieves the GIS layers; scripts 01–08 run offline with fixed seeds. The forecast uses 20,000 draws with seed 20261104. (4) Outputs: CSV and JSON files behind Tables 1–10 and B1–B4, and Figures 1–7 at 600 dpi. (5) Reference verification: Crossref/doi.org checks for all journal references and URL checks for grey literature. See README.md for variable construction, data sources and run instructions.
Files
Steps to reproduce
1. Requirements: Python 3.10 or later with numpy, pandas, scipy, statsmodels, sympy, geopandas, shapely, pyproj, matplotlib and python-docx. Install with: pip install -r requirements.txt. If requirements.txt is not present, install the packages listed above. 2. Unzip Replication_package_SCED.zip and open a terminal in the replication folder. 3. Run the full pipeline offline: bash run_all.sh. This runs scripts 01–08 in sequence on the raw inputs included in the package: 01_build_data.py builds the state panel, campus file and time-to-power intervals. 02_theory_checks.py re-derives Propositions 1–5 symbolically and produces Table 2. 03_siting_ttp.py estimates the PPML siting regressions (Table 3) and time to power (Table 5). 04_forecast_mc.py runs the Monte Carlo forecast (20,000 draws, seed 20261104). 05_verification.py produces Tables 8–10 and the 2025 calibration. 06_figures.py produces Figures 1, 3, 5 and 7. 07_spatial.py runs the case–control and point-pattern analysis (Tables 4, B2, B3). 08_maps.py produces Figures 2, 4 and 6. 4. Optional: to re-download the GIS layers and geocodes from the original sources, run FETCH=1 bash run_all.sh. This adds script 00 and needs internet access. 5. Outputs are written to output/tables (CSV and JSON) and output/figures (PNG, 600 dpi). With the fixed seeds, results match the paper exactly. A full run takes a few minutes on a standard laptop. See README.md for variable construction and data sources.