Knowledge Atlas Tech Joint Stock (02513.HK), the listed entity behind Chinese AI startup Zhipu AI, is exploring development of a custom ASIC chip to power its GLM family of large language models, according to a report from The Information.
The company has held preliminary discussions with several Chinese chip design companies about creating a specialized AI processor tailored to its models, the report said, citing people familiar with the matter. No development partner has been selected, and the project remains in an early phase that could take more than two years to complete.
Shares surged 13.6% to HKD 1,829 in Hong Kong trading on July 7, with turnover reaching HKD 11.2 billion on 6.7 million shares traded. The Hang Seng Index rose about 2% the same day. Short selling accounted for HKD 251.7 million, or 5.2% of total volume.
The chip initiative comes as demand for Zhipu AI's recently launched GLM-5.2 model strains computing capacity. Daily token usage jumped 27-fold during the first week after the model's release, the report said, intensifying the need for more efficient inference hardware.
Why Custom Silicon Matters
Unlike general-purpose graphics processing units, application-specific integrated circuits are built for dedicated workloads, allowing companies to improve energy efficiency and lower long-term operating costs for AI inference. An in-house ASIC could also reduce Zhipu AI's reliance on Nvidia's advanced chips, which have become harder to procure as U.S. export restrictions tighten access to products such as the H100 and B200.
The strategy mirrors moves by Google, which developed its Tensor Processing Unit; OpenAI, which is working with Broadcom on custom silicon; and Chinese peers Alibaba and ByteDance, which have both built proprietary AI processors. For Chinese AI companies, the push toward domestic chip alternatives has accelerated as Washington expanded semiconductor export controls in each of the past two years.
Investment Implications
If Knowledge Atlas moves forward with production, the chip would likely be manufactured at a Chinese foundry, the report said, though no fabrication partner has been named. The next milestone will be the selection of a chip design partner, which would signal the project has progressed beyond the exploratory stage.
Knowledge Atlas trades at a market capitalization of roughly HKD 110 billion based on the July 7 close. The stock has more than doubled this year as investor enthusiasm for Chinese AI names surged following the release of DeepSeek's R1 model in early 2026, which demonstrated competitive performance at a fraction of the training cost of U.S. rivals.
This article is for informational purposes only and does not constitute investment advice.