Reservoir computing (RC), originating in the early 2000s, utilizes dynamical systems as reservoirs to adaptively learn spatiotemporal features and hidden patterns in complex time series. It has shown potential for high-precision prediction in chaotic systems and has attracted significant interest in nonlinear dynamics and complex systems. However, further research is needed to unlock its full capabilities for fast, lightweight, and interpretable learning in temporal dynamical systems. This Perspective discusses the parallel progress in mathematical theory, algorithm design, and experimental realizations of RC, identifying opportunities and challenges for large-scale industrial adoption, along with ideas for resolving these challenges through collaboration between academic and industrial researchers. RC is an alternative to deep learning, offering a compact design and fast training, making it suitable for many industry-level signal processing and learning tasks. It has been shown to be effective in image recognition, climate modeling, and weather forecasting. RC systems typically consist of an input layer, a middle processing layer with nonlinear recurrent dynamics, and an output layer that recombines signals to produce the final output. The core idea of RC is to design and use a dynamical system as a reservoir that adaptively generates signal basis and combines them in an optimal way to mimic the dynamic behavior of a desired process. RC has been shown to be effective in various applications, including signal classification, time series prediction, control of system dynamics, and PDE computations. However, it still faces challenges in terms of performance on certain tasks and scalability. Despite these challenges, RC has the potential to be a key technology for future applications, including 6G, next-generation optical networks, green data centers, intelligent robots, and AI for science. RC's compact and lightweight network structure, rich functional interfaces, and low-complexity training and computing nature make it a promising candidate for edge-side information processing. The development of RC is expected to play a critical role in several important application domains, as it offers a unique combination of performance, efficiency, and adaptability.Reservoir computing (RC), originating in the early 2000s, utilizes dynamical systems as reservoirs to adaptively learn spatiotemporal features and hidden patterns in complex time series. It has shown potential for high-precision prediction in chaotic systems and has attracted significant interest in nonlinear dynamics and complex systems. However, further research is needed to unlock its full capabilities for fast, lightweight, and interpretable learning in temporal dynamical systems. This Perspective discusses the parallel progress in mathematical theory, algorithm design, and experimental realizations of RC, identifying opportunities and challenges for large-scale industrial adoption, along with ideas for resolving these challenges through collaboration between academic and industrial researchers. RC is an alternative to deep learning, offering a compact design and fast training, making it suitable for many industry-level signal processing and learning tasks. It has been shown to be effective in image recognition, climate modeling, and weather forecasting. RC systems typically consist of an input layer, a middle processing layer with nonlinear recurrent dynamics, and an output layer that recombines signals to produce the final output. The core idea of RC is to design and use a dynamical system as a reservoir that adaptively generates signal basis and combines them in an optimal way to mimic the dynamic behavior of a desired process. RC has been shown to be effective in various applications, including signal classification, time series prediction, control of system dynamics, and PDE computations. However, it still faces challenges in terms of performance on certain tasks and scalability. Despite these challenges, RC has the potential to be a key technology for future applications, including 6G, next-generation optical networks, green data centers, intelligent robots, and AI for science. RC's compact and lightweight network structure, rich functional interfaces, and low-complexity training and computing nature make it a promising candidate for edge-side information processing. The development of RC is expected to play a critical role in several important application domains, as it offers a unique combination of performance, efficiency, and adaptability.