Sharding, as a crucial method for enhancing blockchain scalability and throughput, has become a focal point of current blockchain research. With the deepening of blockchain sharding studies, handling cross-shard transactions has emerged as a significant challenge. Most existing sharding systems split cross-sharing transactions into multiple subtransactions, which are then processed by the respective shards. This approach greatly increases the communication overhead and results in high latency for cross-shared transactions. In this paper, we propose a three-layer blockchain sharding system called Tower, which is designed to address the issue of low throughput in cross-sharing transactions. The top layer stores a snapshot of the entire blockchain system’s state and is used to handle complex cross-sharing transactions. The middle layer shards store complete records of the underlying shards they oversee, enabling efficient processing and verification of cross-shard transactions involving the bottom layer shards within the middle layer itself. This approach eliminates the high latency caused by the two-phase commit (2PC) protocol commonly used in traditional blockchain systems. The bottom layer shards are responsible only for processing transactions within their respective shards. To address the communication needs between different shard layers, we design two hierarchical consensus protocols, resolving the consensus issues between bottom and middle layer shards as well as between middle and top layer shards. Additionally, we theoretically analyze the system’s security, scalability, and transaction complexity. Finally, we developed and implemented a prototype of Tower and compared it with classic single-layer and dual-layer sharding systems such as Pyramid. The results demonstrate that in a system with 15 shards and 3,000 nodes, our transaction throughput increased by approximately 1.5 times compared with that of Pyramid.