Zhipu AI is considering developing its own in-house AI chips to meet the rapidly growing demand for its GLM models.

According to monitoring by Beating, China’s leading large language model (LLM) firm Zhipu AI is considering developing its own customized inference chips and has begun inquiring about cooperation with local chip design companies. Zhipu AI is currently on the U.S. Entity List, which bars it from purchasing NVIDIA’s advanced chips. As the daily token usage of its open-source model GLM-5.2 on developer platform Vercel surged 27-fold in a week, its computing resource crunch has grown increasingly severe. While Zhipu AI has deployed domestic computing resources from firms like Huawei and completed extensive software adaptation work, to escape supply chain constraints and lower long-term cloud inference costs, the company has chosen to follow the path of developing its own chips, similar to Google’s TPU and OpenAI’s in-house chip initiatives. This in-house chip project remains in the early discussion stage; if it advances smoothly, manufacturing will be handled by a domestic wafer foundry, with the entire process from design to actual tape-out expected to take over two years.