Skip to content

Sources and method

204 builds so far (+193 in 30 days) · 93 GW stated · 240 history rows · 186 companies

What counts as a build

An announced AI data center campus, or a named phase of one, with a place and a company that runs it: the Stargate sites, Meta's Prometheus and Hyperion, Microsoft's Fairwater sites, AWS Project Rainier, xAI's Colossus, and campuses that Crusoe, Vantage, Nscale, Nebius, TeraWulf and others build for AI labs. Chip deals and cloud contracts without a named site are left out; they live on Partnerships.

Sources

Every build was read from its operator's own announcement where one exists (company newsrooms and blogs, SEC filings) and from reputable reporting for what happened after (DCD, CNBC, Reuters, Business Insider, local public radio). Each build page lists its sources, and each change links the article that reports it.

The numbers

  • Capacity is the campus capacity the company announced, in megawatts, usually IT load. A note says when it is IT power, a first phase or a target. When a company never gave a figure (Microsoft's Fairwater, AWS Rainier) the field is empty rather than an estimate.
  • Investment is the amount the company stated, in US dollars; "reported" marks a figure that comes from reporting rather than the company.
  • Status: Online, Partly online, Building (construction started), Announced, Paused, Cancelled.
  • Operator runs the site; tenant is the main customer when one was announced. Company names link to Companies and use its slugs.

History

Nothing is overwritten. Each change to capacity, status, target date, tenant, investment or operator, and each dated event that is not one field (a permit, a fine, a force majeure notice) is a row with its date and source, kept forever on the build page, on Changes and at /api/changes.

Scope and coverage

The aim is every announced AI or hyperscale data center campus of 100 MW or more, or USD 1 billion or more, anywhere, from 2023 on. Companies in scope are the companies.fru.dev registry's cloud companies and every company named in a build: 14 of 28 (50%) run or rent at least one build on file; the rest are the gaps being closed. 16 builds are marked Unverified until a person re-reads their source.

Updates and sources

Every day at 14:30 UTC. First, free discovery: the run reads the RSS feeds and newsrooms on file (57 so far, +7 in the last 30 days), adds RSS feeds it finds on newsroom pages and on the sites of registry companies, and keeps every item that names a data center with megawatts or dollars as a lead (14 waiting to be read). Then it reads SEC full-text search for 8-Ks that mention a data center. On Tuesdays it also runs at most three news searches. It reads at most six articles, leads first, and asks a model once to list only the changes and new builds those articles state. Its finds are marked Unverified, and its changes are held back, until checked by hand. Without a model or search key the run still finds sources and leads and refreshes the company data.

About this data

Builds from company newsrooms, SEC filings, RSS feeds and news, found by a daily agent and checked by hand; unverified ones are marked. Figures are as announced. The list grows as new builds are announced. Not investment advice. Logos via logo.dev; trademarks belong to their owners. Anonymous usage stats via Google Analytics. Company facts come from companies.fru.dev.

Data

JSON at /api/builds and /api/changes, plain text at /llms-full.txt. For AI agents: /agents and /openapi.json.

AI data center changes by email

Tuesdays, only in weeks when a build is announced, delayed, expanded, paused or cancelled.

Double opt-in. Unsubscribe any time.