4. Methodology Overview
How the methodology works at a high level.
Greenpixie's AI token methodology estimates the energy consumption, carbon emissions, and water consumption of cloud AI inference on a per-token basis. A representative set of open-weights models is benchmarked on real GPU hardware, and regression models trained on those benchmarks generalise the findings across the full range of models in Greenpixie's database, including proprietary models that cannot be measured directly. Predicted energy values are then mapped to each cloud provider's billing configurations and combined with regional carbon intensity, hardware embodied emissions, and water intensity data to produce per-token carbon and water estimates.
The methodology operates through five components, being physical benchmarking, predictive modelling, billing configuration mapping, carbon and embodied emissions calculation, and water consumption calculation. Section 7 describes each in detail. Together they produce the per-token energy, carbon, and water values applied to customer usage data within Greenpixie's data product.
The architecture and deployment configuration of a proprietary model are not disclosed by its vendor. Where these characteristics cannot be observed, the methodology establishes the range of values each could plausibly take and samples across those ranges, producing a median and an interquartile range for every value it publishes. Section 7.6 describes this process.