Cooperative perception can extend the visible range of autonomous vehicles, but dense raw-data or feature exchange is difficult to sustain under vehicular bandwidth and latency constraints. This paper proposes a bandwidth-budget-constrained sparse semantic Token framework for collaborative 3D semantic occupancy completion. Multimodal observations are encoded as candidate Tokens carrying content, spatial indices, timestamps, and predicted code lengths. A semantic-utility and redundancy score identifies Tokens that are informative for occluded, uncertain, and long-range regions. During training, differentiable soft gates and rate-distortion-aware bit allocation are optimized jointly with the completion task, during inference, the gates are converted into hard selections by a utility-per-bit ranking and a budget-feasible greedy/local-swap procedure. The implemented communication cost includes entropy-coded Token values, delta-coded spatial indices, timestamps, source identifiers, quantizer identifiers, packet headers, CRC, and byte padding. At the receiver, an asynchronous Token memory performs pose compensation, temporal decay, and contextual retrieval before a dual-path 3D backbone and a progressive generative decoder complete occupancy and semantic categories. All quantitative results are reported on one V2X-Sim-derived semantic occupancy benchmark, and every compared method uses the same split, sensor inputs, voxel grid, training schedule, hardware, asynchronous link protocol, and serialized communication budget. Under a total budget of 0.36 Mbit/frame, the proposed method achieves a geometric occupancy IoU of 51.04% and a semantic mIoU of 31.42%, with an end-to-end computational latency of 44 ms and 14.9 FPS. Compared with Early Fusion, communication is reduced by 98.06% while semantic completion accuracy is improved.