Zhipu AI, one of China's leading large language model companies, is exploring the development of custom inference chips and has already reached out to domestic chip design firms about potential collaboration. The company is currently on the US entity list, blocking it from purchasing advanced NVIDIA GPUs.
As Zhipu's open-source GLM-5.2 model saw daily token consumption on the developer platform Vercel explode by 27 times in a single week, the strain on computing resources has become increasingly acute.
While Zhipu has already deployed domestic alternatives like Huawei's chips and invested heavily in software adaptation, the company is now looking to follow the paths of Google's TPU and OpenAI's custom silicon to break free from supply chain constraints and reduce long-term cloud inference costs.
The self-designed chip project is still in early-stage discussions. If it moves forward, production would be handled by domestic fabs, and the timeline from design to tape-out is expected to take at least two years.