Zhipu Joins the ASIC Race: GLM-5.2 Token Usage Explodes 27x in Week One
TL;DR
Zhipu is in early talks with Chinese chip designers to build a custom ASIC for GLM-5.2, whose token usage surged 27x in one week.
Zhipu AI, the Chinese large-model lab behind the GLM family, is in preliminary talks with several domestic chip-design houses about building a custom ASIC to replace part of its Nvidia GPU footprint. The Information broke the story on July 7, followed the next day by TrendForce and Investing.com. No partner has been picked; internal estimates put a viable ASIC — through design, tapeout, validation and software adaptation — at more than two years out.
The trigger is GLM-5.2. Since its launch last month, daily token usage on Vercel's AI platform has surged 27x in the first week, making it the fastest-growing model on that platform. Zhipu is simultaneously supporting API commercialization and inference for a Malaysian national AI project, and the compute gap is now obvious. Tightening US export controls on H100 and H200 supplies, plus mounting losses, push the company to squeeze inference cost per token.
The playbook is well-worn. Google's TPU, OpenAI's Broadcom-fabricated silicon, ByteDance, Alibaba, and last week's DeepSeek disclosure all follow the same path — pull inference off Nvidia's general-purpose GPUs and claw margin back with in-house ASICs. What is different: Google and OpenAI can lean on TSMC's leading-edge nodes, while Chinese labs at best reach SMIC N+2. Binding an ASIC to a domestic foundry chain means betting on the design team and yield at the same time.
If it works, GLM gets pricing power over its own compute and Nvidia disappears from Zhipu's inference bill. If it doesn't, in two years Zhipu holds a chip already outpaced by the next GLM, the compute shortage persists, and the cash is gone. A Zhipu insider told The Information the company needs to "build out semiconductor capability, navigate chip validation, and adapt the software ecosystem" all at once — three things no Chinese artificial intelligence company has yet completed end to end.
via The Information / TrendForce / Investing.com
The trigger is GLM-5.2. Since its launch last month, daily token usage on Vercel's AI platform has surged 27x in the first week, making it the fastest-growing model on that platform. Zhipu is simultaneously supporting API commercialization and inference for a Malaysian national AI project, and the compute gap is now obvious. Tightening US export controls on H100 and H200 supplies, plus mounting losses, push the company to squeeze inference cost per token.
The playbook is well-worn. Google's TPU, OpenAI's Broadcom-fabricated silicon, ByteDance, Alibaba, and last week's DeepSeek disclosure all follow the same path — pull inference off Nvidia's general-purpose GPUs and claw margin back with in-house ASICs. What is different: Google and OpenAI can lean on TSMC's leading-edge nodes, while Chinese labs at best reach SMIC N+2. Binding an ASIC to a domestic foundry chain means betting on the design team and yield at the same time.
If it works, GLM gets pricing power over its own compute and Nvidia disappears from Zhipu's inference bill. If it doesn't, in two years Zhipu holds a chip already outpaced by the next GLM, the compute shortage persists, and the cash is gone. A Zhipu insider told The Information the company needs to "build out semiconductor capability, navigate chip validation, and adapt the software ecosystem" all at once — three things no Chinese artificial intelligence company has yet completed end to end.
via The Information / TrendForce / Investing.com
