AI Sustainability Research
Energy and carbon disclosures for AI are now common, but the freshwater cost of a single query is reported inconsistently and with incompatible accounting boundaries, which makes it effectively impossible to compare across models, prompt sizes, or regions. This research asks how to define a transparent, comparable per-query freshwater metric for AI inference. It is the freshwater companion to the lab's published Carbon Cost of Intelligence (CCI) framework.
AI sustainability reporting has converged on energy and carbon, but freshwater remains the least consistently measured cost of inference, and what providers do disclose uses incompatible accounting boundaries. This thread investigates how to define a transparent, comparable per-query freshwater metric that accounts for water consumed both directly, in cooling a data center, and indirectly, in generating the electricity that powers it, and that can reflect how much a given volume of water matters in the region where a query is served. It is scoped to operational water; embodied water in hardware manufacturing and construction is out of boundary. The work is the freshwater extension of the lab's published Carbon Cost of Intelligence (CCI) framework, and draws on sustainability accounting, the energy-water nexus, and transparent, comparable measurement. Faculty-advised.
