This work proposes a bandwidth-scalable channel estimation algorithm based on convolutional neural networks (CNNs) in line with fifth-generation (5G) standards. Support for different bandwidth configurations is essential to meet the heterogeneous and challenging demands of the 5G and next-generation wireless communication systems. Typically, each bandwidth configuration requires a dedicated CNN architecture, which also implies an individual training process. Our approach introduces a single module adaptable to all bandwidth parameters. Promising results for a subcarrier spacing of 15 kHz are achieved considering the mean-square error (MSE) and bit error rate (BER) performance metrics. Furthermore, an online inference on a field-programmable gate array (FPGA) is conducted to evaluate a comprehensive trade-off analysis between performance, computational efficiency, scalability, and generalizability.