Zhipu AI explores an in-house chip route
According to monitoring by Dongcha Beating, Chinese frontier model company Zhipu AI is considering the development of a custom inference chip and has already approached domestic chip design companies to discuss possible cooperation. The effort is still in an early-stage contact phase, with no indication that it has moved into full execution or production planning.
The reported push toward an internal chip strategy is being shaped by two pressures at once. First, Zhipu AI has been placed on the U.S. blacklist, which prevents it from purchasing Nvidia’s advanced chips. Second, demand for its open-source GLM-5.2 model has risen quickly, making existing compute constraints more visible.
GLM-5.2 usage spikes on Vercel
The report said daily token usage for GLM-5.2 on the developer platform Vercel surged 27 times in one week. That increase has intensified the company’s compute shortage. While Zhipu has already deployed domestic computing resources, including Huawei hardware, and completed extensive software adaptation, the company is still evaluating a deeper hardware move to reduce long-term dependence on constrained supply chains.
Another stated goal is to lower long-term cloud inference costs. In that sense, the strategy is described as following a path similar to Google’s TPU program and OpenAI’s own chip efforts. Even so, any such project would be a long-cycle undertaking rather than a near-term fix.
Manufacturing timeline likely to stretch beyond two years
If the project moves forward smoothly, manufacturing would reportedly be handled by Chinese foundries. However, the timeline from chip design to actual tape-out is expected to take more than two years. That means the initiative, if realized, would be aimed more at medium- to long-term supply resilience and cost control than at solving immediate infrastructure bottlenecks.