Energy-Aware Adaptive AI Inference in Edge–Cloud Computing

Abstract

Adaptive AI inference distributes the computation across devices, edge servers, and cloud infrastructure to optimize task quality, latency, and energy. This study provides a critical overview of the literature on offloading, split inference, early exits, model routing, hardware adaptation, and speculative decoding selected from 2018 to 2026. Targeted public-source searches and full-text inspection enable comparison of mechanisms, experimental settings, and energy boundaries; exhaustive systematic coverage and meta-analysis are not claimed. The synthesis differentiates between measurements of client-energy and wider node measurements, analytical estimates, and communication and monetary proxies. Qualified numerical examples demonstrate the importance of having baseline choices and differences to go with reported savings. The evaluation requirements go beyond traditional DNN inference, as the models are based on language and multimodal models introduce repeated verification, rejected computation, and response-length dependence. The primary recommendation is to consider policies at specified quality points and latency goals and to factor in communication, monitoring, transitions, and escalation. It is important to have a clear view of accounting and reproducible workloads to determine if adaptation ultimately improves overall energy efficiency or moves energy between tiers.