In accordance with the PoC Reference for Green Computing with Remote GPU over APN, we conducted a proof of concept using the small‑scale LLM tsuzumi‑7B.For compute‑dominant workloads where the dataset is relatively small, data transmission time is negligible compared to GPU computation time, even in long‑distance remote GPU environments. While replication and cache strategies are effective for large, data‑intensive workloads, their benefit is limited in lightweight language‑model scenarios such as this PoC.