The increasing scale and computational demands of large artificial intelligence models (LAIMs) pose significant challenges for efficient inference among distributed devices in resource-constrained environments. This paper presents a multi-cluster LAIM co-inference framework in which an edge server, equipped with multiple graphics processing units (GPUs), coordinates user clusters to fulfill inference requests. We assume that devices within each cluster capture data from different perspectives, and then use on-device lightweight LAIMs to generate local features. The edge server collects and fuses these features to produce a more accurate inference result. To capture the fundamental trade-off between model pruning and co-inference performance under multi-device scenarios, we theoretically characterize the impact of pruning ratio and device contribution based on the rate-distortion analysis and partial information decomposition. Guided by the analysis, we then jointly optimize the pruning ratio, task scheduling, and transmit power to minimize inference distortion subject to the latency, energy, and capacity constraints. Extensive simulations demonstrate that the proposed framework achieves superior performance in comparison to existing baselines, offering improved inference quality and resource efficiency in multi-cluster edge environments.