TL;DR
- Facility Status: Chinese model developer Z.AI’s reported domestic-chip data center has started partial operation.
- Designed Scale: Its claimed 1GW design uses over 10,000 domestic chips to develop Z.AI’s General Language Model (GLM) family.
- Software Mechanism: Zhongke Jiahe may help Z.AI run models across chips from Chinese suppliers including Huawei and Cambricon.
- Performance Proof: Customers still need verified chip identities, commissioned power, reliability, utilization, and sustained GLM throughput.
Z.AI’s reported completion of a domestic-chip data center includes the start of partial regional operations. Separately, Z.AI recently completed its acquisition of infrastructure and foundational software startup Zhongke Jiahe on July 21.
Z.AI’s facility has reportedly 1 gigawatt of designed capacity and is intended to train its General Language Model (GLM) family.
No public Z.AI confirmation establishes the facility so far and site and investment cost remain undisclosed, along with the exact accelerator mix and achieved operating capacity. Partial operation of a 1GW design would not establish that the entire power envelope has been commissioned for model training.
What a Gigawatt-Scale Domestic Stack Must Solve
The site’s reported inventory of more than 10,000 domestic AI chips says little by itself about accelerator performance, interconnect speed, memory, power efficiency, or reliability. Cooling, networking, and storage also draw from the 1GW electrical envelope, reducing the power available to working accelerators. Sustained throughput and completed training runs would provide more useful measures than chip count or rated capacity alone.
Nvidia still remains the dominant U.S. supplier of advanced AI accelerators, specialized chips used to train and run models, in spite of increased efforts by Huawei and other companies to catch up with domestic AI chips.
Replacing its hardware at scale requires software that can coordinate different Chinese processors as one usable system. Chinese AI infrastructure and foundational software startup Zhongke Jiahe, acquired by Z.AI may develop software that spans Chinese chip designs, including hardware from Chinese AI-chip suppliers Huawei and Cambricon, but neither company has identified the hardware installed at the site.
Cross-chip software could reduce the engineering needed to adapt one model to several accelerator designs. It could also let Z.AI combine available processors instead of relying on one supplier, although that flexibility would not guarantee equal performance across the chips.
Z.AI’s GLM Models
The recently released GLM-5.2 model achieved a 62.1 SWE-bench Pro score, compared with 58.4 for GLM-5.1 on the software-engineering benchmark.
Z.AI positions GLM-5.2’s 1M-token window for long-horizon work, which adds memory and computing demand across accelerators, networking, storage, and software. Neither model result verifies how the data center distributes workloads, handles failures, or converts its electrical envelope into completed training work.
Lower running costs could broaden adoption of Chinese models. Lian Jye Su, chief analyst at Omdia, said: “They can be run at a fraction of the cost that OpenAI charges its clients”.
The Domestic Baseline and the Next Proof Point
Z.AI tied domestic-chip model training to Huawei hardware in February, although technical specifics may have changed.
Industry estimates assign Chinese manufacturers about 41% of accelerator shipments in China in 2025, or 1.65 million of roughly 4 million GPUs.
Alibaba’s own reported plan for a 10,000-chip data center in Guangdong dates to April, and the project remained incomplete when Z.AI completed the Zhongke Jiahe acquisition.
For Louis Gave, founding partner and co-chief executive of Gavekal Research, limited access to cheap capital may continue to push Chinese AI companies toward open-source software, cheaper local chips, and electricity-grid investment.
