AI is reshaping the physical world. Most of its footprint goes unreported.

What the data shows

Across the latest reported year for each organization, the 11 we track consumed roughly 257 TWh of electricity, about the annual use of 24.5 million US homes. They withdrew around 348 GL of water, some 139,000 Olympic swimming pools, and their reported and estimated emissions total roughly 182 Mt CO₂e, comparable to 39.7 million passenger cars.

And not one of them discloses how much of that footprint is attributable to AI workloads specifically. These totals mix company-disclosed figures with our documented estimates; every underlying number carries its confidence tier and source on the entity profiles.

The global AI ecosystem consumes energy, water, land, and raw materials at a scale that is poorly understood and almost entirely undisclosed. Training large models requires data centers drawing tens to hundreds of megawatts. Inference, run continuously at scale, may dwarf training energy over a model's lifetime. The semiconductor supply chain stretches across continents, each node carrying its own environmental burden.

OpenButterfly exists to map this terrain with the rigor it deserves. We draw on government databases, corporate sustainability disclosures, academic research, regulatory filings, satellite imagery, and investigative journalism to build a picture of what is actually happening, not what is being claimed. Where data is absent, we say so. Absence of disclosure is itself a data point.

This is a slow, difficult, ongoing research project. We are not a rating agency, a ranking tool, or an advocacy campaign. We are an observatory, committed to methodological transparency, honest uncertainty, and public access.

The transparency problem

Most AI companies do not publish meaningful environmental data. Sustainability reports, when they exist, aggregate figures across entire corporate operations, making it impossible to attribute environmental costs to specific AI workloads. Some companies publish carbon figures that exclude Scope 3 emissions entirely. Others report water use only for facilities they own, not those they rent from cloud providers.

The gap between what is disclosed and what actually happens is not merely a reporting problem. It reflects a structural asymmetry: companies bear no regulatory obligation to disclose AI-specific environmental data in most jurisdictions, and investors have only recently begun to demand it. The result is a market where environmental externalities remain invisible by default.

OpenButterfly treats non-disclosure as a data signal, not a gap. When a company does not publish its water consumption, that fact is recorded and weighted in our methodology. The silence is as informative as the number would be.