Big generative models have significantly impacted several domains like natural language processing, computer vision, and drug discovery. These developments, large-scale generative foundation models exemplified by architectures such as GPT-4 have emerged to address a broad spectrum of generalized tasks. These models are trained to capture the underlying general correlations from a large training dataset. Although big generative AI techniques have been explored in various wireless communication applications, such as channel generation, the integration of generalized foundation models into this domain remains limited. A core component of these models is the masked autoencoder. This work investigates the suitability of masked autoencoders for massive multiple-input multiple-output (MIMO) systems, focusing on their capacity to capture spatial and temporal correlations in massive MIMO channels. To this end, a massive MIMO scenario with user mobility is considered, where the channel state information (CSI) varies with both spatial and temporal correlation. A masked autoencoder is trained in a self-supervised manner using channel state information (CSI) from multiple users. The trained model is then tested for its performance in tasks of feedback compression, channel interpolation, and channel prediction. Experimental results demonstrate that masked autoencoders effectively capture inherent correlations within massive MIMO channels, underscoring their potential to advance foundational model-based approaches in wireless communications.