Cloud computing has revolutionized modern IT architectures by enabling scalable and flexible resource allocation. However, accurately forecasting cloud workloads remains a significant challenge due to their high variability, non-stationarity, and complex temporal dependencies. In this paper, we analyze and compare traditional statistical models, such as ARIMA and VARIMA, with Recurrent Neural Network-based approaches, specifically Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks, for predicting CPU utilization workloads. Experiments were conducted on a real-world dataset comprising Azure Virtual Machine traces, containing fine-grained CPU usage metrics sampled every five minutes. Results show that neural models outperform VARIMA in capturing workload fluctuations, with GRU achieving the lowest normalized prediction errors across metrics. These findings provide insights for selecting forecasting models that support efficient and cost-effective resource management in dynamic cloud environments.