Overview
- The data center, which multiple outlets reported on Monday, draws about 1 gigawatt of power and Z.ai says it will use the site to train its GLM family with clusters that hold more than 10,000 domestically made accelerators each.
- The facility contains no Nvidia hardware and builds on Z.ai’s June release of GLM‑5.2, which company and industry reporting say was trained on Huawei Ascend accelerators.
- Analysts caution that Chinese accelerators lag current Nvidia Blackwell parts on performance per watt, so a 1 GW power bill does not equal the same usable training compute as an equivalent Nvidia site.
- Scaling usable capacity may be constrained by supply limits in China such as high SMIC wafer‑node utilization and scarce domestic high‑bandwidth memory (HBM), which slow how fast Ascend‑class cards can be produced.
- Beijing’s draft five‑year, roughly 2 trillion yuan plan to build a national AI data‑center grid and rules that favor domestic sourcing create a steady market for local chips and could reshape where future AI compute demand flows.