This paper presents a novel end-to-end (E2E) learning architecture for massive MIMO communication systems using complex-valued neural networks (CVNNs). Our approach leverages CVNNs to process complex signals directly, eliminating the need to split real and imaginary components, thereby preserving the natural structure of wireless signals. The proposed architecture integrates both the encoding and decoding stages, optimized for flat-fading Rayleigh channel conditions, focusing on improving transmission efficiency. A key contribution is the extension of the approach to multiuser MIMO scenarios, where the system is designed to orthogonalize data streams for several user equipment, improving spectral efficiency with federated learning. We show that it is possible to effectively transmit a number of data streams that exceed the channel matrix rank. Additionally, a power control mechanism based on regularization is introduced to ensure stable transmission power. The effectiveness of the proposed approach is rigorously validated through simulations across a range of scenarios, demonstrating significant improvements in the mutual information. The results are compared with theoretical limits and classical approaches, highlighting the potential of CVNN-based architectures for advancing future wireless communication systems in both single and multiuser contexts.