Carbon-Intensity Predictive Computing for Generative AI: Forecasting Spatiotemporal Electricity Emissions for Carbon-Aware Computational Orchestration

The increasing deployment of generative artificial intelligence (AI) systems has created substantial electricity demand associated with large language model (LLM) inference, model loading, prompt processing, autoregressive generation, and distributed computational services. Although predictive computing provides opportunities to anticipate future computational demand and prepare reusable resources before actual requests arrive, the environmental consequences of these operations depend not only on how much electricity is consumed but also on the carbon intensity of the electricity supplying the computational infrastructure. Electricity carbon intensity varies across geographical regions and changes over time according to electricity generation composition, renewable energy availability, grid conditions, weather patterns, and electricity demand. Consequently, computational workloads with identical energy requirements may produce substantially different operational emissions depending on where and when they are executed. This paper introduces Carbon-Intensity Predictive Computing for Generative AI (CIP-GAI), a conceptual framework that extends predictive computational orchestration through the forecasting of spatiotemporal electricity carbon intensity. Rather than relying exclusively on historical or current carbon intensity measurements, CIP-GAI investigates whether future regional electricity emissions can be predicted and incorporated into computational preparation, workload scheduling, GPU allocation, and geographically distributed inference decisions. The proposed architecture integrates six principal functions: (1) regional electricity carbon-intensity data acquisition, (2) spatiotemporal carbon-intensity forecasting, (3) uncertainty-aware carbon emission estimation, (4) carbon-aware predictive workload preparation, (5) geographically distributed computational orchestration, and (6) continuous carbon accounting and adaptive optimization. The framework distinguishes average grid carbon intensity from short-run marginal and long-run marginal emissions, recognizing that these measures represent different environmental accounting perspectives and may lead to different scheduling decisions. It further distinguishes electricity consumption from electricity carbon intensity, thereby separating computational energy efficiency from the environmental characteristics of the electricity supply. A unified optimization formulation is developed to minimize expected operational carbon emissions while accounting for GPU energy consumption, regional carbon-intensity forecasts, prediction uncertainty, workload deadlines, response latency, geographical transfer costs, cloud resource prices, and computational preparation expenditure. Particular attention is given to the interaction between predictive computation and carbon-aware execution. Anticipated inference workloads may be prepared in advance during periods and locations with comparatively favorable electricity carbon intensity, provided that their computational states remain reusable and their preparation does not compromise immediate service requirements. The framework also investigates the possibility of selecting geographically distributed GPU resources according to predicted regional carbon intensity rather than relying exclusively on computational availability or monetary cost. An experimental methodology is proposed to evaluate carbon-intensity forecasting accuracy, regional scheduling decisions, operational emissions, computational efficiency, economic expenditure, and service-level performance. Comparisons with reactive inference, energy-aware scheduling, price-aware predictive computing, and non-predictive carbon-aware scheduling are established. No experimental results or verified emission reductions are claimed in this conceptual study. The proposed framework provides a theoretical foundation for investigating whether forecasting where and when electricity will be relatively carbon-intensive can improve the environmental efficiency of predictive computing infrastructures for generative AI.