Zhipu acquired hundreds of millions of yuan AIInfra Company Zhongke Jiahe fully made up for the shortcomings of the underlying heterogeneous computing power project

According to the "AI Technology Review" report, Intelligent Spectrum, the leading company of China's model, has spent hundreds of millions of yuan to complete the acquisition of Zhongke Jiahe, an AI heterogeneous computing power software infrastructure company. This move aims to completely fill in the shortcomings of Intelligent Spectrum in the underlying engineering and compiler capabilities of large models to cope with the structural shortage of computing power and high concurrency reasoning problems caused by the surge in user numbers.<br /> Zhongke Jiahe Technology originated from the Compilation Laboratory of the Institute of Computing, Chinese Academy of Sciences and was founded by Dr. Cui Huimin. Its core team has been deeply involved in the research and development of compilers for many domestic chips such as Loongson, Shenwei, Cambrian and Huawei Shengteng. Zhongke Jiahe's core advantage lies in its virtual instruction set technology, which can unify the chip ecosystem of different brands and models through mid-layer software, and assemble scattered domestic chips into a unified ultra-large-scale cluster, thereby greatly improving the overall computing power utilization; According to official claims, its SigInfer inference engine can reduce the inference latency of large models by up to 74 times.<br /> Recently, Zhipu's CodingAgent business has experienced explosive growth. In the first week of launch of the newly released GLM-5.2 model, the average daily Token invocation volume soared 27 times, causing its reasoning infrastructure to be exposed in high-concurrency and long-context scenarios. Systematic engineering bottlenecks are exposed. After being included in the list of U.S. entities, Intelligent Spectrum actively promoted domestic substitution and has completed reasoning adaptation to eight domestic computing power platforms including Huawei Shengteng, Pingtouge, and Moore Thread. Taking Zhongke Jiahe into the bag this time will not only directly improve the inferential cost and output quality per unit Token of Intelligent Spectrum, but will also provide core underlying compiler technical support for its previously announced self-developed and customized AI inference chip plan.