What do AI energy and water disclosures actually tell you?
A company-wide power or water report describes infrastructure and practices within a stated boundary. It does not automatically measure the energy or water for a particular model, app, or request. Check the reporting year, metric, region, and workload before comparing numbers.
What kinds of water information do providers publish?
Water withdrawal, consumption, water-use effectiveness, and replenishment measure different things. WUE describes water intensity per unit of data-center IT energy, not water per prompt. Replenishment describes restoration projects, not a reduction in local withdrawals. AWS publishes data-center and regional WUE ↗; Google reports freshwater use and replenishment ↗; Microsoft reports fleet water intensity ↗; Meta reports stress-region restoration ↗.
Does a report measure one chatbot or AI request?
Product-level impact needs a known serving route, location, model, workload, and measured method. Company totals do not allocate use to a prompt. The directory’s energy practices score assesses documentation and controls; it is separate from water disclosure and does not estimate a request footprint. Check the scope for ChatGPT, Gemini, Microsoft Copilot, and Grok.
How can I compare providers fairly?
Compare the same reporting year and boundary: direct operations or full value chain, company or named facility, withdrawal or consumption, and measured or estimated data. Note whether high-stress regions are disclosed and whether the provider maps a service to a location. If the route is unknown, the product-level answer remains unknown. Browse compare providers and read each profile’s environmental sources.