1. Introduction
Purpose, audience, and context of the methodology.
This document describes Greenpixie's methodology for estimating the energy consumption, carbon emissions, and water consumption of cloud-based AI inference workloads. It is intended for customers, partners, procurement teams, auditors, regulators, and technically literate readers.
Organisations that consume hosted AI models through cloud provider APIs generate energy demand and associated carbon emissions with every request. As AI adoption grows and enterprise usage shifts toward higher-volume inference, this demand increases. Published analyses by Patterson et al. and MIT Technology Review attribute approximately 80% to 90% of total AI compute to inference, with model training accounting for the remainder.
Greenpixie provides per-token energy, carbon, and water estimates for cloud AI inference, supporting decisions about model selection, deployment configuration, and sustainability reporting. Each estimate is expressed as a median with an interquartile range, on the basis described in Section 7.6. This methodology is a companion to the Greenpixie Methodology for Cloud Emission Measurement, which details the broader cloud emissions framework and provides the carbon intensity, water, and embodied emissions factors applied here.