Storage-Aware Predictive Computing for Generative AI: Anticipating Future Storage Demand for Answer Caches, KV Caches, and Model Artifacts

The increasing deployment of generative artificial intelligence (AI) systems has introduced substantial storage requirements associated with response caching, transformer key-value (KV) states, model weights, adapters, checkpoints, and other computational artifacts. As generative AI services expand across heterogeneous hardware infrastructures and distributed inference environments, storage management is becoming an increasingly important component of overall computational efficiency. Existing storage optimization approaches primarily respond to observed storage utilization, incoming requests, and resource shortages. Although predictive computing has begun to investigate the anticipation of future inference demand, GPU resource requirements, and computational preparation opportunities, the prediction of future storage demand remains an important complementary research direction. This paper introduces Storage-Aware Predictive Computing for Generative AI (SAPC-GAI), a conceptual framework that investigates whether future storage requirements can be anticipated and selectively prepared before actual storage demand materializes. The framework focuses on three major categories of generative AI storage resources: (1) answer cache storage, including generated responses and reusable informational content; (2) KV cache storage, including temporary and persistent transformer key-value representations; and (3) model artifact storage, including model weights, quantized model variants, LoRA adapters, checkpoints, and related inference artifacts. Its architecture integrates predictive storage demand modeling, workload-to-storage estimation, storage capacity forecasting, predictive data placement, adaptive cache management, storage tier selection, and coordinated resource allocation. A unified optimization formulation is developed to determine how much storage capacity should be reserved, which computational artifacts should be retained, where they should be placed, and when storage resources should be expanded, migrated, or released. The framework further distinguishes between transient memory requirements and persistent storage requirements, recognizing that GPU memory, system memory, local SSDs, distributed storage, and object storage exhibit different capacity, latency, and economic characteristics. An experimental methodology is proposed to evaluate storage prediction accuracy, capacity utilization, cache hit rates, data access latency, storage-related costs, migration expenditure, and overall inference efficiency. Particular attention is given to the interaction between predicted computational demand and predicted storage demand, since advance computation may generate reusable artifacts whose future storage requirements must also be anticipated. The proposed framework does not assume that predictive storage management is universally beneficial; its effectiveness depends on workload predictability, artifact reuse probability, storage costs, data lifetime, access patterns, and resource constraints. By extending predictive computing toward storage infrastructure, this research investigates a broader approach in which generative AI systems anticipate not only future computation but also the storage resources required to support future inference operations.