Sources, Updates & Methodology
How we compile the data, what can be compared and what the figures do not represent.
Compute Atlas: US infrastructure projects
Public Compute Atlas API records are imported server-side. We keep only facilityType=data_center; separate crypto-mining and power-generation categories are excluded.
Records may identify operating, construction-stage, permitted, proposed or cancelled projects. AI classifications are shown only if supplied explicitly. Source confidence and original document links remain visible on facility profiles.
Licence: CC BY 4.0. Data from Compute Atlas by Edward Kubiak, dataset DOI. This application adds its own interface and analysis; the underlying data remains attributed.
EIA: hourly regional demand
US EIA Form-930 observations are fetched from /v2/electricity/rto/region-data/data/ using respondent and type D filters at hourly UTC frequency.
Critical limitation: these series report total balancing-authority demand, not data-center demand or AI electricity use. They cannot be divided into AI and non-AI loads from this dataset alone.
We preserve the observation period and source timestamp instead of calling delayed grid observations “live”. Official API documentation ↗
Incomplete coverage and power numbers
Compute Atlas is an independently curated collection, not a comprehensive census. Absence of a project does not establish that it does not exist. Public filings and announcements can be corrected or delayed.
Planned and operating MW are displayed separately. Missing capacity is unknown rather than 0, and multi-site entries, status changes and different capacity definitions make direct summation unreliable. The interface deliberately does not display a nationwide total MW capacity.
Freshness and change detection
CLI cron imports Compute Atlas once per day and EIA hourly data several times daily. Records are upserted by stable source IDs, hours by respondent + UTC period, and changed facility records are stored in the change log. Individual source observation dates may predate the import time.
Successful update timestamps are shown at the bottom of the pages. After a failed sync, the last saved data remains displayed and the sync log retains the failure. The application does not automatically delete records missing from a subsequent source response.