The Case for Eco-cost Labelling of AI inference

AI's environmental costs fall disproportionately on the Global South, yet organisations typically choose AI models by their intelligence, cost, latency and suitability for sovereign deployment. These environmental costs are becoming a political issue worldwide, but consumers, governments and enterprises lack sufficient, trusted information about the ecological costs of their own AI usage. This paper proposes a common service and open-source code repository for estimating the operational eco-costs of inference served by hyperscalers and data centres worldwide. We take inspiration from environmental labelling initiatives, including LEED certification for buildings, appliance energy ratings and sustainable harvesting certifications, as market-based mechanisms for making externalities visible and supporting informed choices. Given a model, input and output token counts, date and time, and available data-centre or hyperscaler information, the proposed API would return estimates of greenhouse gas emissions, water consumption and energy use, together with energy-source information, uncertainty and provenance. We also present UX concepts for a proposed reference integration into an open-source AI orchestration platform for public-sector organisations, illustrating how buyers could consider trade-offs among cost, speed, intelligence, sovereignty and environmental impact as they develop AI solutions to advance the Sustainable Development Goals and their own objectives. The repository and integration are proposed work; the figures are illustrative, and their effects on buyer behaviour and emissions remain to be evaluated.