Zhipu AI Eyes ASIC, GLM Usage 27×
Zhipu AI, a Beijing‑based artificial‑intelligence laboratory, has entered early discussions with several domestic chip‑design houses about creating a bespoke application‑specific integrated circuit (ASIC) optimized for its GLM family of large language models. The company has made preliminary inquiries but has not yet selected a partner, and the conversations remain at an exploratory stage. According to The Information, the ASIC development effort could extend beyond two years, requiring Zhipu to assemble or expand a semiconductor engineering team, navigate full chip design and testing cycles, secure foundry capacity, and rework its software stack to exploit the new hardware.
The catalyst for this initiative is the explosive growth of Zhipu’s latest model, GLM‑5.2, which has become the fastest‑growing model on Vercel’s model‑aggregator platform since its launch last month. Daily token usage for GLM‑5.2 surged as much as 27‑times during its first week, creating a strong demand for more efficient inference compute.
At the same time, tightening U.S. export controls on advanced semiconductors have made it increasingly difficult for Chinese AI labs to obtain Nvidia’s high‑end GPUs, turning compute availability into a binding constraint rather than a purely cost consideration. ASICs, unlike general‑purpose GPUs, are engineered to perform specific model‑related tasks, delivering superior energy efficiency and lower per‑token inference costs once a model’s architecture stabilises. Industry analysts note that inference‑optimised ASICs can substantially cut operating costs for mature workloads, although exact savings vary by model and utilisation.
Zhipu would be following a path already taken by larger peers such as Google, OpenAI, ByteDance and Alibaba, all of which have developed proprietary chips to reduce reliance on external GPU suppliers and lower inference costs. Reuters recently reported that DeepSeek is also pursuing custom chips to lessen dependence on Huawei and Nvidia. The urgency for Zhipu is heightened by the export‑control backdrop: unlike U.S. counterparts that can still access Nvidia’s latest silicon, Chinese labs face a regulatory ceiling that makes any domestic alternative strategically valuable.
On the domestic side, Chinese firms including Cambricon Technologies Corp Ltd (SS:688256) and Biren Technology have been active in the AI‑ASIC space, though The Information did not name either as a prospective partner for Zhipu. The broader ecosystem of domestic designers has expanded meaningfully since the initial rounds of U.S. export restrictions took effect, providing Chinese AI labs with more options than existed two years ago.
For Nvidia investors, each report of a Chinese AI lab exploring custom silicon adds incremental structural pressure on Nvidia’s China data‑center revenue, which historically represented a meaningful portion of its overall sales. The cumulative effect of multiple labs pursuing domestic ASICs reinforces a bear case that Nvidia’s exposure to the Chinese market faces not only political but also enduring technological headwinds.
The immediate execution risk for Zhipu lies in the multi‑year horizon: the lab must manage chip‑design complexity, secure foundry access, and adapt its software stack simultaneously. Should GLM demand continue at a pace comparable to the launch‑week surge, the business case for overcoming these hurdles strengthens. The next visible milestone will likely be the public selection of a chip‑design partner, which would signal a transition from exploratory talks to a committed development programme.