Accuracy- and Utility-Driven Adaptive Predictive Computing for Generative AI: Automatically Optimizing Computational Strategies Through Prediction Performance and Realized Operational Benefits

The increasing deployment of generative artificial intelligence (AI) systems has created computational environments characterized by fluctuating inference demand, heterogeneous GPU availability, changing electricity prices, variable cloud resource costs, and uncertain future information requirements. Predictive computing provides a potential mechanism for anticipating future computational workloads and preparing reusable computational resources before corresponding requests arrive. However, the effectiveness of predictive computation depends not only on the initial quality of demand forecasting or resource allocation, but also on the system's ability to continuously evaluate and improve its operational strategies according to actual outcomes. A predictive strategy that performs effectively under one workload distribution may become inefficient when demand patterns, resource availability, computational costs, or information characteristics change. This paper introduces Accuracy- and Utility-Driven Adaptive Predictive Computing (AUAPC-GAI), a conceptual framework that enables generative AI infrastructures to automatically adjust their predictive computational strategies according to observed prediction accuracy and realized operational benefits. Unlike conventional predictive computing approaches that primarily determine which workloads should be prepared in advance, AUAPC-GAI introduces a continuous feedback mechanism connecting prediction evaluation, computational outcome measurement, strategy adjustment, and subsequent execution. Its architecture integrates prediction-performance assessment, realized computational benefit estimation, adaptive strategy selection, resource-allocation adjustment, feedback-driven policy optimization, and operational stability control. The framework distinguishes between prediction accuracy and computational utility, recognizing that accurate predictions do not necessarily generate economic or computational benefits, while moderately accurate predictions may still be valuable when their associated preparation costs are sufficiently low and their successful reuse provides substantial savings. A unified optimization formulation is developed to jointly characterize prediction quality, net computational benefits, resource constraints, service-level requirements, and adaptation costs. The framework further investigates how systems can automatically adjust precomputation intensity, prediction thresholds, task priorities, resource allocation, execution timing, and strategy selection based on accumulated operational evidence. An experimental methodology is proposed to compare static predictive strategies, accuracy-only adaptation, utility-only adaptation, and integrated accuracy-and-utility-driven adaptation under changing workloads. The proposed research does not assume that continuous adaptation necessarily improves system performance; rather, it establishes the theoretical and experimental foundations for determining when feedback-driven strategic adjustment can improve the long-term effectiveness of predictive computation. By introducing an outcome-driven adaptation mechanism, this research extends predictive computing from anticipating future computational demand toward continuously improving how that demand is anticipated, prepared, and managed.