Infrastructure Observatory Methodology
How we identify, classify, and characterize AI-related infrastructure sites.
Inclusion criteria
We track infrastructure that is materially connected to AI development, training, inference, or deployment at scale. Inclusion requires at least one of the following:
- –The facility is publicly identified by its owner or operator as AI-purpose infrastructure.
- –The facility hosts compute clusters used primarily for AI training or large-scale inference.
- –The facility supplies power, water, or land that is contractually or operationally dedicated to AI infrastructure.
- –The facility is a hyperscale data center operated by a company whose primary business is AI or cloud AI services.
We do not include general-purpose enterprise or colocation data centers unless there is documented evidence of substantial AI workloads. Absence from this database does not indicate environmental compliance.
Classification definitions
Explicit AI Infrastructure
explicit_ai
Facilities that are purpose-built or publicly identified as AI training or inference infrastructure. Examples: GPU supercomputer clusters, AI training data centers contracted to a specific AI lab, inference farms with documented AI workloads.
AI-Capable Hyperscale
ai_capable
Large hyperscale data centers operated by cloud providers or AI labs that host substantial AI workloads alongside general cloud services. These facilities could accommodate AI training or inference at scale, and may host it, but AI workloads are not the sole or documented primary use.
Associated Infrastructure
associated
Substations, power generation projects, transmission lines, water infrastructure, or land sites that are contractually or operationally tied to AI infrastructure. Includes dedicated interconnects, renewable energy projects constructed specifically to serve AI data centers, and utility-scale substations sited for AI campus load.
Environmental data sources
All environmental context values (water stress, grid carbon intensity, drought risk) are researcher-assigned at time of data entry. We never fetch these values from live APIs. The following sources inform our environmental characterizations:
Environmental values are reviewed when a site record is updated. The last_verified date reflects when all values for a site were last reviewed against current sources.
Confidence framework
The infrastructure observatory applies the same A–E/ND confidence framework as our entity-level data, adapted for site-level characteristics:
Measured & Verified
Directly measured by an accredited third party, or from a government regulatory record with verified data.
Disclosed (Unverified)
Reported by the facility owner or operator in a public document, not independently verified.
Estimated (Documented Method)
Derived from proxy data (e.g. grid region average, basin-level risk index) using a documented methodology.
Estimated (Uncertain)
Derived from indirect or analogous data with significant uncertainty.
Unknown / Modeled
No direct or proxy data; value derived from sector-level models or analogous cases. Treat with substantial caution.
Not Disclosed
The facility owner has not published this information. Non-disclosure is recorded.
The overall_confidence field on each site reflects the weakest-link confidence across all recorded values: the overall rating can be no higher than the least-certain individual field.
What last_verified means
Every site record carries a last_verified date. This is the date on which a researcher reviewed all fields in the record against current sources and confirmed they remained accurate or updated them as needed.
It is not the date on which the underlying event occurred (e.g. a data center opened, or a planning objection was filed). The underlying event dates are captured in opened_year and source publication dates.
Community concern standards
A community concern flag is set when there is a documented, verifiable public record of the concern. This means:
- –A regulatory body received and recorded a formal objection or complaint.
- –A planning authority published a record of objection during a permitting process.
- –A local government body adopted a resolution or issued a public statement.
- –Journalism with attributed sources documented community opposition or formal complaint.
We do not set community concern flags based on social media, informal commentary, or unverified reports. Each flag is linked to a source in the Evidence section of the site record.
What OpenButterfly does not claim
We do not make causal claims about environmental harm attributable to specific facilities. Reporting that a site is located in a high water stress basin does not mean that the site is causing water stress. It means the site is located in a context where water scarcity is a documented concern.
We do not attribute workloads, energy consumption, or emissions to specific sites. A site's capacity (in MW) is a physical characteristic of the facility, not a claim about actual utilization or emissions.
Community concern flags document events that are on the public record. They do not imply that concerns are valid, ongoing, or unresolved, only that they were documented in a verifiable source.
How to contribute or correct
If you have a source that contradicts or updates information in this database, or if you are aware of a site that should be included, we welcome contributions.
To suggest a correction or new site, please open an issue on our GitHub repository with the site name, a link to your source, and what it establishes. All contributions will be reviewed before being added to the database.