Online Adaptive Predictive Model Updating for Generative AI: Continuously Refining Computational Demand Forecasts from Incoming Requests

Predictive computing for generative artificial intelligence (AI) relies on the ability to anticipate future computational demand before corresponding inference requests materialize. However, predictive demand models are inherently exposed to changing user behavior, emerging information trends, evolving workload distributions, and unexpected computational requirements. A forecasting model trained using historical request distributions may gradually become inaccurate as new requests arrive and the underlying demand patterns change. Static predictive models may therefore produce increasingly unreliable estimates of future inference workloads, potentially resulting in unnecessary precomputation, inefficient GPU allocation, inappropriate resource reservations, and inaccurate economic decisions. This paper introduces Online Adaptive Predictive Model Updating for Generative AI (OAPMU-GAI), a conceptual framework that continuously refines predictive computational demand models using newly observed requests, realized computational outcomes, and prediction errors. Rather than treating forecasting models as fixed components of predictive computing infrastructure, the proposed framework considers them as continuously evolving computational decision mechanisms. Its architecture integrates incoming request monitoring, prediction-outcome comparison, online data assimilation, concept drift detection, adaptive model updating, model evaluation, update scheduling, and controlled deployment. Newly observed requests provide information about changing demand frequencies, query characteristics, token distributions, inference costs, GPU utilization, and resource consumption. The framework uses these observations to determine whether existing predictive models remain sufficiently accurate or require incremental parameter updates, adaptive retraining, model recalibration, or structural modification. A unified mathematical formulation is developed to characterize the relationship between prediction error, model update expenditure, future forecasting performance, and downstream computational benefits. Particular attention is given to the distinction between updating demand prediction models and updating the underlying generative AI models themselves. The proposed framework focuses primarily on predictive infrastructure models responsible for forecasting future requests and computational requirements, while treating changes to foundation-model parameters as a separate and optional extension. An experimental methodology is established to evaluate forecasting accuracy, adaptation speed, computational update overhead, prediction stability, and the resulting effects on proactive inference preparation and economic-aware resource allocation. The framework further investigates how online model updating can interact with predictive GPU scheduling, off-peak precomputation, dynamic electricity pricing, and cloud resource selection. Its effectiveness remains subject to empirical validation, particularly under nonstationary workloads, delayed feedback, sudden demand changes, and limited model-update budgets. By introducing continuous predictive model refinement as an explicit component of predictive computing, this research explores a mechanism through which generative AI infrastructures can adapt their understanding of future computational demand as new operational evidence becomes available.