According to reports, Zhipu has invested hundreds of millions of RMB to acquire AI Infra company XCore Sigma, further enhancing its foundational model engineering capabilities and strengthening domestic computing power adaptation and inference optimization capabilities.

XCore Sigma is an AI heterogeneous computing software infrastructure company focusing on compilation technology. Its technology originates from the Compilation Lab at the Institute of Computing, Chinese Academy of Sciences. Its core team has participated in the development of compilers for domestic chips such as Loongson, Sunway, Cambricon, and Huawei Ascend. It possesses full-chain capabilities from virtual instruction set design, operator generation to optimization deployment, aiming to provide a unified software foundation for domestic AI large models.

Previously, Zhipu had invested in AI Infra companies such as Jiliu Tech, Wuwen Xinqiong, and Silicon Flow, covering areas such as computing clusters and inference optimization. However, with the rapid growth of model call volumes, infrastructure pressure gradually emerged. After the release of GLM-5 in March, the daily call volume of the Coding Agent reached hundreds of millions, and some users reported output anomalies in complex tasks. Zhipu's review found that there were engineering challenges in the inference architecture and KV Cache management under high concurrency scenarios.

This acquisition is expected to improve Zhipu's adaptation efficiency in multi-chip environments. XCore Sigma's compiler technology can connect different domestic chip ecosystems through an intermediate layer, improving computing power utilization. In addition, its SigInfer inference engine previously claimed to reduce inference latency and improve throughput efficiency; if implemented in production environments, it will further reduce the cost of model services.

As AI competition expands from model capabilities to computing power, software infrastructure, and ecosystem building, underlying technologies such as compilers are becoming a key direction for large model companies to build long-term competitiveness.