Skip to main content
Compute Atlas

Manifesto & method · Edition 2026

Public, but scattered.

Compute Atlas is a community-driven, open, source-verified survey of the AI-datacenter buildout across the United States — where it is being built, by whom, and at what cost to energy, water, and the communities nearby.

Manifesto

§ The problem

The information is public. Assembling it is the hard part.

The facts about any given datacenter are, in principle, on the record — scattered across county permit filings, tax-abatement agreements, water-authority applications, interconnection queues, and local news. No one keeps them in one place. Gathering them, cross-checking them, and turning them into something comparable is genuinely difficult work. That difficulty is the problem this atlas exists to solve.

§ The method

A source for every record.

Every field traces back to a public source, cited on the facility’s page. We record what the documents say and no more: ranges and projections are marked as estimates, not presented as fact; where a figure genuinely isn’t known, the record says so rather than guessing. Nothing here is fabricated or inferred beyond what a reader can check for themselves.

§ The invitation

Built in the open, correctable by anyone.

The code and the data are public, and the atlas is meant to be stewarded by the people who use it — journalists, researchers, local officials, and residents. If a record is wrong, incomplete, or missing, you can fix it. Every contribution needs one thing: a public source anyone can verify.

§ The stance

Non-partisan, and not affiliated with anyone.

Compute Atlas takes no editorial position on whether any facility should be built. It is not affiliated with any company, advocacy group, or government agency. The aim is a factual, honest starting point — what to make of it is up to the reader. The source code and data are public.

§ 01 · Definition

What counts as an “AI datacenter”

Working definition: facilities primarily purpose-built or repurposed for large-scale AI and machine-learning training or inference, typically characterized by hyperscale GPU or accelerator clusters rather than general-purpose cloud or enterprise compute.

The category is inherently imprecise — hyperscale cloud campuses increasingly colocate AI workloads alongside other compute, and company announcements often do not distinguish between the two. Every record therefore carries an AI classification that reflects how confident we are that a facility is primarily AI compute:

Confirmed
The operator or a credible primary source explicitly describes the facility as an AI or GPU cluster (e.g., xAI Colossus).
Likely
The facility exhibits strong indicators (e.g., hyperscale GPU procurement, AI-specific power agreements) but has not been explicitly confirmed as AI-primary.
Mixed use
A multi-purpose campus where AI workloads are a known component but not necessarily the primary or exclusive use.

§ 02 · Status

Status definitions

Each facility carries one of five lifecycle statuses. Tracking status transitions over time is a core feature of Compute Atlas — the full history of known status changes is recorded for every facility.

Operational
Built and running.
Under construction
Actively being built.
Permitted
Approved/permitted, not yet under construction.
Proposed
Announced or proposed; not yet approved.
Cancelled
Cancelled or withdrawn.

§ 03 · Confidence

Confidence levels

In addition to AI classification, each record carries a confidence level that reflects the quality and independence of the underlying sources:

Confirmed
Verified by multiple independent sources or by an official operator announcement with supporting documentation (e.g., permit filings, utility agreements).
Reported
Covered by at least one credible news outlet or official filing, but not yet corroborated by multiple independent sources.
Rumored
Based on a single, unverified source or on indirect indicators (e.g., job postings, land acquisition records). Treat with caution and check the linked sources directly.

§ 04 · Sources

How the data is compiled

The current dataset is a curated seed drawn from publicly available sources. Every record links to the specific sources used to create or update it. Source types include:

  • Company announcements and press releases
  • Permit and zoning filings
  • Utility and ISO interconnection queue entries
  • Federal and state subsidy disclosures
  • OpenStreetMap data (for coordinates and facility boundaries)

The roadmap includes automated ingestion from interconnection queues and public permit databases, as well as a structured community submission process. If you have a correction or addition, see the Contribute section below.

§ 05 · Limitations

Limitations

Compute Atlas aims to be accurate and honest, but there are real constraints on what the data can reliably represent:

  • There is no national registry of AI datacenters. “AI datacenter” is not a legal or regulatory category.
  • Announcements are often aspirational. Projects are frequently delayed, scaled back, or cancelled after public announcements — tracking cancellations is a feature, not an edge case.
  • Coverage is partial and skews toward large, well-reported facilities. Smaller or less-publicized projects are likely underrepresented.
  • Capacity figures (in megawatts) are often estimates from third-party sources and may mix planned with operational capacity. They should be treated as approximate.
  • Coordinates are best-effort from public sources (OpenStreetMap, permit filings). Some coordinates reflect a nearby town center rather than the facility itself.

§ 06 · License

Attribution & licenses

§ 07 · Contribute

Contribute & corrections

Community contribution is how this atlas stays accurate and grows. If you have a correction, a missing facility, or an updated source, open an issue on GitHub — every submitted change needs one thing: a public source URL anyone can verify.

Open an issue on GitHub →

Prefer to support the work directly? Sponsor this project on GitHub.

← Back to the map