Predictive maintenance (PdM) in Industry 4.0 (I4.0) increasingly relies on machine learning (ML) techniques to minimize unplanned downtime and enhance operational efficiency. While cloud-based ML solutions offer scalability and strong predictive performance, their reliance on network connectivity introduces latency and reliability issues that hinder real-time industrial applications. This study investigates the deployment of lightweight autoencoder (AE)-based models optimized for edge computing environments, comparing their performance against traditional cloud-hosted alternatives. Multiple model architectures were evaluated, and inference latency was benchmarked across four deployment scenarios: cloud-hosted PyTorch, native PyTorch on Raspberry Pi 3B, TensorFlow Lite (Python runtime), and TensorFlow Lite (C++ runtime). Latency measurements, averaged 100 executions per model, reveal that edge deployment can reduce inference time by up to 7000× compared to cloud containers, with TensorFlow Lite C++ deployments achieving latencies as low as 60 microseconds. These results demonstrate that edge-based ML deployment is a viable strategy for enabling timely, autonomous fault detection in real-time PdM systems.