This article presents a spindle condition monitoring methodology using a low-power smart vibration sensor and a near real-time deep neural network (DNN) classifier. The most frequent spindle failures, such as imbalance, ingression, and evidence of a crash with the workpiece, are analyzed in this study. Experiments were designed to induce various failure events to monitor the spindle behavior using conventional vibration, current and temperature sensors, and an intelligent vibration sensor. The smart sensor is a device with internal signal processing identifying eight dominant frequencies and the amplitude/power distributions. It requires low power and generates narrow bandwidth messages that can be communicated wirelessly. A Fog device and a test plan are designed to monitor and store a dataset needed to train a DNN classifier. The Fog device generates temperature, current, and vibration signals from sensors connected to the spindle and sends them to data storage in the cloud. The signals were analyzed using both conventional vibration analysis and Artificial Intelligence-based classifiers. Metrics such as crest factor, skewness, kurtosis, and overall enveloping were used to assess their ability to identify the failure condition. The data from the smart sensor are used to train an optimized DNN, and the spindle defect classification performance is measured. With 960 data points per failure mode and training data taken over 960 min of operation, the optimized DNNs can classify the spindle states with an accuracy of 98%. The study shows real-time spindle condition classification feasibility over long periods using inexpensive and low-power smart vibration sensors.