compute-atlas

Compute Atlas is an open, source-cited dataset covering data centers, crypto-mining sites, and the dedicated power generation built to supply them across the United States, from proposed and permitted through under construction and operational. It also records civic footprint information such as energy, water, subsidies, jobs, emissions, and community reception, gathered from county planning portals, SEC filings, state air-quality permits, ISO/RTO interconnection queues, utility filings, and subsidy disclosures. The project states that no national registry of data centers exists and positions itself as a sourced alternative. Every facility record cites at least one public source with a URL, label, source kind, and retrieval date, and a sourceIndex links specific claims to specific documents. Records carry a confidence level (confirmed, reported, rumored) and a status (proposed, permitted, under construction, operational, cancelled). Access options include a filterable map and table, an explore view by state, operator, metro, power, status, crypto, and community opposition, plus a public CORS-open API with endpoints for facilities, individual facilities, search, stats, and schema. The complete dataset is also published as a single forkable JSON file, with a separate metadata file recording its asOf timestamp. Writes are moderated: contributions and URL leads are staged for human review, and submission management requires an admin token. Stated design principles include treating missing fields as unknown rather than zero, reporting what share of records disclose an aggregated field, keeping ranges and modeled projections in notes rather than numeric fields, forbidding derived figures, and requiring explicit human approval before anything publishes. Candidates from the discovery pipeline have cited URLs fetched and compared against the claims they support before staging. The data model is validated against a Zod schema as a discriminated union on facilityType, with fields for id, name, operator, status, confidence, AI classification, location, capacity in megawatts, status history, sources, and lastUpdated, plus type-specific and civic blocks. The test suite validates every record on each pull request. The stack is Next.js, React, TypeScript, Zod, Neon Postgres with Drizzle, MapLibre GL, Tailwind with shadcn/ui, Vitest, and Playwright. Running without a database URL renders from the JSON snapshot. The project targets WCAG 2.2 AA, with the data table as a first-class alternative to the map and automated accessibility audits in the end-to-end suite. Code is MIT licensed and the data is CC BY 4.0; the project describes itself as non-partisan and provides the data as is without warranty.