Zhipu, one of China’s leading large-model companies, has spent several hundred million yuan to acquire AI heterogeneous computing software infrastructure firm Zhongke Jiahe, according to a report from AI Technology Review cited by ChainCatcher.
The acquisition is intended to address weaknesses in Zhipu’s lower-layer model engineering and compiler capabilities as the company deals with structural compute shortages and the demands of high-concurrency inference brought on by rapid user growth.
Zhongke Jiahe’s technical roots and core advantage
Zhongke Jiahe’s technology originated from the compiler laboratory at the Institute of Computing Technology of the Chinese Academy of Sciences. The company was founded by Dr. Cui Huimin. Its core team previously participated deeply in compiler development for a range of Chinese chip projects, including Loongson, Sunway, Cambricon and Huawei Ascend.
According to the report, Zhongke Jiahe’s key advantage lies in its virtual instruction set technology. Through a software middle layer, the company can unify chip ecosystems across different brands and models, assembling scattered domestic chips into a single ultra-large-scale cluster and improving overall compute utilization.
The report added that, according to the company’s own claims, its SigInfer inference engine can reduce large-model inference latency by as much as 74x.
Pressure on Zhipu’s inference infrastructure
Zhipu’s Coding Agent business has recently seen explosive growth. Its newly released GLM-5.2 model posted a 27x surge in average daily token calls during its first week on an aggregation platform, the report said. That increase exposed systemic engineering bottlenecks in Zhipu’s inference infrastructure under high-concurrency and long-context workloads.
After being placed on the U.S. Entity List, Zhipu has been pushing domestic substitution efforts. The report said it has already completed inference adaptation for eight domestic computing platforms, including Huawei Ascend, T-Head and Moore Threads.
What the deal could change inside Zhipu
According to the report, bringing Zhongke Jiahe in-house could directly improve Zhipu’s per-token inference cost and output quality. It could also provide core lower-layer compiler support for Zhipu’s previously reported plan to develop a custom AI inference chip.