Predictive Carbon-Aware Computing for Generative AI: Forecasting Low-Carbon Computational Windows Under Latency Constraints

The increasing deployment of generative artificial intelligence (AI) systems has introduced substantial electricity consumption and associated carbon emissions across GPU-intensive inference infrastructures. Although predictive computing enables generative AI systems to anticipate future computational demand and prepare reusable computational resources before actual requests arrive, the environmental impact of such preparation depends not only on how much computation is performed, but also on when and where it is executed. Performing anticipated computations during periods of high electricity carbon intensity may undermine the environmental benefits of predictive inference preparation, even when those computations reduce subsequent online inference costs. This paper introduces Predictive Carbon-Aware Computing for Generative AI (PCAC-GAI), a conceptual framework that integrates future computational demand prediction, electricity carbon-intensity forecasting, low-carbon execution-window selection, deadline-aware scheduling, and predictive computational preparation. The framework investigates how generative AI infrastructures can anticipate future inference workloads and strategically schedule flexible computational tasks during periods when electricity generation is expected to produce lower carbon emissions, while preserving the latency requirements of subsequent user requests. The proposed architecture integrates six principal functions: (1) predictive computational workload identification, (2) temporal and geographical carbon-intensity forecasting, (3) low-carbon computational opportunity-window identification, (4) latency-constrained task scheduling, (5) carbon-aware predictive resource preparation, and (6) adaptive emissions measurement and control. Unlike conventional carbon-aware scheduling approaches that primarily shift existing workloads toward cleaner execution periods, PCAC-GAI connects carbon-aware scheduling with the predictive preparation of generative AI resources, including anticipated responses, compatible KV cache states, contextual representations, and other reusable inference components. A unified optimization formulation is developed to minimize expected operational carbon emissions while accounting for workload energy consumption, carbon-intensity uncertainty, task execution duration, predicted request-arrival times, response freshness, computational capacity, and service-level latency constraints. The framework distinguishes between delay-tolerant predictive preparation tasks and time-critical online inference requests, ensuring that environmental optimization does not take precedence over mandatory response deadlines. The research further distinguishes electricity carbon intensity from electricity prices, GPU utilization, and total computational expenditure. It emphasizes that the economically cheapest execution window is not necessarily the environmentally preferable one, and that workload shifting should be evaluated using appropriately defined operational emissions rather than assumed environmental benefits. An experimental methodology is proposed to compare immediate execution, utilization-aware scheduling, price-aware scheduling, conventional carbon-aware scheduling, and the complete predictive carbon-aware architecture. Evaluation metrics include operational carbon emissions, energy consumption, deadline satisfaction, P95 and P99 response latency, prediction accuracy, preparation reuse rate, and total computational expenditure. The framework does not assume that low-carbon scheduling is universally beneficial; its effectiveness depends on carbon-intensity variability, workload flexibility, forecasting accuracy, execution deadlines, resource availability, and successful reuse of prepared computational resources. Its practical benefits remain subject to empirical validation.