One of the main objectives of WP2 is to develop energy-efficient hardware platforms for AI-native transceivers, with a focus on enabling low-power, high-performance signal processing at the wireless edge. This deliverable presents the outcomes of hardware architecture investigations conducted in Tasks T2.1.2, T2.1.3, T2.2.3, and T2.2.4, centered on a hybrid analog-digital neuromorphic computing paradigm.To address the energy bottlenecks in neural inference, particularly for sequence modeling tasks such as adaptive symbol detection, we propose a mixed-signal architecture that combines analog in-memory computing for feed-forward layers with digital stochastic circuits for attention mechanisms. This architecture supports real-time processing of temporally encoded spike signals and reduces memory access overhead by co-locating computation and storage. In addition, hardware-aware training and drift compensation techniques are implemented to mitigate the impact of device nonidealities.Our design is evaluated on a spiking neural receiver use case. Experiments demonstrate significant improvements in computational and energy efficiency compared to conventional digital implementations, achieving up to 14.5× energy reduction and over 7× speed-up, while maintaining comparable detection accuracy. These results validate the feasibility of neuromorphic mixed-signal hardware for future 6G transceiver platforms and highlight the importance of hardware-software co-design in energy-constrained AI systems.