Predictive GPU Power Modeling for Generative AI: Anticipating the Energy Consumption of Future Computational Tasks Before Execution

The increasing deployment of generative artificial intelligence (AI) systems has introduced substantial electricity consumption associated with large language model (LLM) inference, image generation, video generation, multimodal processing, and other computationally intensive workloads. As generative AI infrastructures become increasingly dynamic, their electricity consumption is determined not only by the number of incoming requests but also by the characteristics of individual computational tasks, GPU architectures, model configurations, execution strategies, and workload concurrency. Existing GPU resource management approaches primarily monitor utilization, memory allocation, execution time, and aggregate power consumption during or after computation. However, these measurements provide limited information about the electrical requirements of computational tasks that have not yet been executed. This paper introduces Predictive GPU Power Modeling for Generative AI (PGPM-GAI), a framework that investigates the prediction of GPU power consumption and task-level energy requirements before anticipated computational workloads are executed. The proposed framework extends predictive computing beyond the forecasting of computational demand, resource availability, and economic execution opportunities by introducing an energy-oriented prediction layer that estimates the expected electrical behavior of individual precomputation tasks. Its architecture integrates computational workload characterization, GPU power profiling, task-level power prediction, energy consumption estimation, uncertainty-aware energy forecasting, and power-aware predictive scheduling. Candidate workload features include model architecture, parameter scale, input sequence length, output token length, batch size, numerical precision, attention configuration, GPU memory requirements, hardware architecture, expected execution duration, and concurrent workload conditions. The framework distinguishes instantaneous GPU power, task-attributable power, incremental energy consumption, and total system energy, recognizing that these quantities are related but not interchangeable. A unified mathematical formulation is developed to predict task-level power profiles and cumulative energy consumption while accounting for idle power, overlapping workloads, prediction uncertainty, GPU operating conditions, and measurement overhead. The resulting predictions can be integrated with Economic-Aware Predictive Computing to determine not only when and where anticipated workloads should be executed, but also how much electricity their execution is expected to consume. An experimental methodology is proposed to evaluate prediction accuracy, energy estimation error, task-level attribution, scheduling effectiveness, and total electricity consumption under controlled and realistic generative AI workloads. The framework does not assume that predictive power modeling automatically reduces energy consumption; rather, it investigates whether advance knowledge of computational energy requirements can improve the coordination of precomputation tasks and support more energy-aware operation of generative AI infrastructure. Its practical effectiveness remains subject to empirical validation.