China Telecom Deploys First Domestic TPU Cluster in Hangzhou: Zhonghao Xinying Chana TPU With 400 PFLOPS, National-Grade Computing Chain From Chip to Platform

TL;DR AI
2 min readKey summary
China Telecom, ZTE, and Zhonghao Xinying launched a domestic TPU cluster in Hangzhou with 1,024 chips delivering 400 PFLOPS of compute.
The first phase runs on the Chana TPU, while the second-generation Xuyu TPU is planned for WAIC 2026 with higher performance and lower power use.
Native prefill-decode separation boosts utilization and reportedly lifts token output efficiency by more than 10x.
Inference cost per million tokens has dropped from about 100 yuan to under 10 yuan, highlighting a major cost breakthrough for local AI workloads.
