Zhipu AI Considers Developing Custom Chips as GLM-5.2 Usage Surges 27x
Taylor Wilson
Chinese AI lab Zhipu AI (智谱AI) is evaluating the feasibility of designing its own AI chip, as daily token usage of its GLM-5.2 model surged up to 27× in its launch week. This means → demand is outrunning compute, and with U.S. export controls choking off chip supply, Zhipu is being pushed toward the hardest path: building its own silicon.
Why is Zhipu suddenly looking at making its own chips?
Two pressures are converging: GLM-series demand is surging, while U.S. export controls have cut off access to Nvidia chips.
Zhipu is on the U.S. Commerce Department's Entity List, banned from purchasing any American technology. Its chip partners must come from inside China.
This means → Zhipu isn't choosing to go custom — under its current supply-chain reality, it has almost no alternative.
How strong is GLM-5.2 demand, exactly?
On Vercel's model aggregation platform, GLM-5.2 has been the fastest-growing model since its release last month, with peak daily token usage surging 27× in launch week.
Coinbase CEO Brian Armstrong publicly said Coinbase is trialing GLM-5.2 to cut AI costs; the model is also now available on Oracle Cloud Infrastructure.
In plain terms = major overseas customers are adopting it voluntarily, and major platforms are listing it unprompted — this is real commercial pull, not a lab-stage experiment.
What chips is Zhipu using now — and are they enough?
Zhipu currently runs a mix of Huawei chips, other domestic Chinese chips, and a small number of Nvidia chips. In January it released its first image-generation model trained entirely on Huawei silicon.
But Huawei chips face their own export-control constraints — Chinese fabs have limited access to advanced manufacturing equipment, and Zhipu needs extra software engineering to run models efficiently on Huawei hardware.
This reflects a deeper problem: even switching to domestic chips isn't a plug-and-play fix — the software adaptation alone is a high-cost hurdle.
What does the custom-chip plan look like?
Sources say Zhipu has had preliminary talks with several Chinese chip-design firms about a custom AI processor — an ASIC (a chip purpose-built for a specific task, offering higher efficiency and lower power draw than general-purpose GPUs).
The goal: a chip optimized specifically for running GLM-series models. Discussions are still at an early stage; no design partner has been selected.
If the project proceeds, fabrication will also happen at Chinese fabs — the entire chain stays domestic.
How long will this take?
Sources are explicit: custom silicon is a long-term solution, not a quick fix. Building a semiconductor team, choosing partners, testing the chip, and adapting software adds up to a timeline of more than two years.
In plain terms = even if everything goes smoothly, this chip won't be running production workloads until 2027 at the earliest.
This means → before the chip lands, Zhipu must sustain its demand growth on the existing Huawei-plus-domestic mix — that gap period is the real test.
What does this tell us about the broader industry?
Zhipu's move tracks the same path as every major global AI developer: Google, OpenAI, ByteDance, and Alibaba have all launched or are developing custom chips.
OpenAI's latest — the "Jalapeño" chip, co-developed with Broadcom — is slated to run GPT-series models before year-end.
This reflects an emerging industry consensus: once model scale and usage cross a threshold, general-purpose GPUs stop being the best answer — whoever controls the chip controls the cost floor.
Content is for reference only, not financial advice.