The burgeoning field of Multimodal Artificial Intelligence (AI) aims to develop systems capable of processing and understanding information from diverse sensory inputs, such as vision, language, and audio. A significant bottleneck in training these sophisticated models is the immense cost and effort associated with annotating vast quantities of multimodal data. Unsupervised representation learning offers a promising solution by enabling models to learn meaningful feature representations directly from unlabeled data. Among the myriad unsupervised techniques, contrastive learning has emerged as a particularly powerful paradigm, demonstrating remarkable success in both unimodal and, more recently, multimodal contexts. This article provides a comprehensive review of unsupervised representation learning with contrastive learning in multimodal AI systems. We elucidate the core principles of contrastive learning, its evolution from unimodal applications to cross-modal alignment, and its capacity to learn robust, transferable representations across heterogeneous data sources. By synthesizing key architectural designs, empirical successes, and applications, we highlight how contrastive learning facilitates better understanding, alignment, and fusion of information from different modalities. Furthermore, we discuss the inherent challenges, such as handling unaligned or sparse multimodal data, and outline critical future research directions towards building more versatile and data-efficient multimodal AI.