Beijing-based Zhipu AI finds itself at a familiar crossroads. Demand for its latest open-source models has outstripped available computing power. U.S. export rules keep tightening. So the company once known as one of China’s “AI tigers” has begun early talks with domestic chip design firms about building its own specialized processor.
The move, reported today by The Information, comes as daily token usage for GLM-5.2 surged 27 times in a single week. That kind of growth strains even the best-provisioned data centers. And it forces hard choices on any lab trying to serve both Chinese enterprises and a growing roster of international developers.
Zhipu, which rebranded internationally as Z.ai, listed on the Hong Kong stock exchange in January 2026. The IPO raised roughly $558 million and valued the firm at several billion dollars at debut, according to CNBC. Shares later climbed on enthusiasm for its rapid model releases. Yet the listing also brought greater scrutiny of its supply-chain vulnerabilities.
Those vulnerabilities trace directly to American sanctions. Washington has blacklisted Zhipu and restricted shipments of advanced Nvidia chips to Chinese AI developers. The company responded by shifting training and inference onto domestic hardware. In January it released GLM-Image, an open-source multimodal model that Bloomberg said represented “the country’s first to be fully trained using domestic chips.” The entire pipeline ran on Huawei’s Ascend Atlas 800T A2 servers with Ascend 910 processors. Bloomberg noted the milestone signaled progress toward Beijing’s goal of technological self-reliance.
By February, Zhipu unveiled GLM-5. Reuters reported the model featured stronger coding abilities and support for long-running agent tasks. It approached Anthropic’s Claude Opus 4.5 on certain coding benchmarks and beat Google’s Gemini 3 Pro on others. CEO Zhang Peng told the wire service that overseas revenue had started to pick up, though the firm still earned most of its money inside China. Reuters.
The pace has not slowed. In June the lab dropped GLM-5.2. Its own blog post described the model as delivering a “substantial leap in long-horizon task capability” with a full one-million-token context window that remains stable. The architecture includes IndexShare, which cuts indexer computation by 2.9 times at maximum context length, and multi-token prediction that boosts acceptance rates by up to 20 percent. On benchmarks it trails Claude Opus 4.8 by just one percent on FrontierSWE while edging out GPT-5.5 on another test. It also claims the top spot among openly available models on SWE-Marathon. Z.ai official blog.
That performance has translated into real adoption. The 27-fold jump in token usage cited by The Information reflects both developer enthusiasm and the model’s permissive MIT license. Anyone can download the weights from Hugging Face or ModelScope and run them locally with vLLM, SGLang or other frameworks. Enterprises pay for hosted access through Z.ai’s coding plan, where GLM-5.2 consumes quota at three times the normal rate during peak Beijing hours. A temporary promotion has lowered off-peak billing to one times through September.
Such popularity creates a classic inference-scale problem. General-purpose GPUs from Nvidia or even Huawei’s Ascend chips deliver flexibility. They also carry higher per-token costs and power demands when running the same model architecture millions of times per day. A custom ASIC tuned to GLM’s specific matrix shapes, attention patterns and activation functions could slash energy consumption and lower serving expenses. The trade-off is obvious. The chip would be useless for other models.
Zhipu has not picked a partner yet. Discussions remain preliminary. People familiar with the plans told The Information that the project could take more than two years to reach production. Still, the direction mirrors a broader trend. DeepSeek, another Chinese lab, is also exploring its own inference chip, Reuters reported on the same day. Both efforts aim to loosen dependence on foreign silicon while gaining tighter control over the full stack.
The timing carries extra weight. Zhipu’s shares fell 23 percent in late February after compute shortages sparked user complaints and forced the company to limit new sign-ups, Wikipedia’s entry on the firm notes. That episode highlighted how quickly software success can turn into hardware pain. A purpose-built processor offers one path to relief. So do continued optimizations in software. GLM-5.2 already ships with improved memory management and kernel tweaks designed to handle the massive key-value caches required by million-token contexts.
Analysts watching the sector see the custom-chip talk as more than cost control. It represents an attempt to harden the business against future export crackdowns. If U.S. rules expand to cover more categories of AI accelerators, a domestically fabricated ASIC designed in China and produced at a local foundry becomes harder to block. Of course, advanced manufacturing and high-bandwidth memory still present chokepoints. Those constraints explain why the effort remains early-stage.
Zhang Peng and his co-founders, Tsinghua University professors Tang Jie and Li Juanzi, built Zhipu on the premise that open models could compete with closed Western offerings. The strategy has paid off in downloads and developer mindshare. GLM-5.2 ranks as one of the strongest openly available models on agentic and coding tasks. Yet serving those downloads at competitive prices requires infrastructure the company does not fully own.
So the lab weighs its options. It could deepen partnerships with Huawei, Cambricon, or Moore Threads. It could invest further in software efficiencies that stretch existing silicon. Or it could follow the path of hyperscalers elsewhere and design silicon that matches its models exactly. The preliminary inquiries with Chinese design houses suggest the third option now sits on the table.
Success is far from guaranteed. Fabricating a competitive AI chip demands expertise in everything from architecture to packaging to yield optimization. Even companies with deep pockets have stumbled. The payoff, however, could be substantial. Lower inference costs would let Zhipu price its API more aggressively. Better power efficiency would ease data-center constraints inside China. And full stack control would accelerate iteration between model updates and hardware features.
For now the company continues to ship software updates at a blistering clip. GLM-5.2’s coding plan subscribers already have access. Local enthusiasts can run the weights on consumer hardware thanks to the open release. The model’s long-context strengths suit complex developer workflows, project-level reasoning, and multi-step debugging. Those use cases consume tokens quickly. Hence the 27x spike.
Industry watchers expect more Chinese AI labs to pursue similar hardware strategies. The combination of exploding model adoption and constrained chip supply leaves few alternatives. Zhipu’s exploration of custom silicon may prove an early data point in a larger shift toward vertical integration across China’s AI sector.
Whether the lab ultimately builds the chip or simply uses the threat to extract better terms from existing suppliers remains to be seen. What is clear is that the days of relying primarily on imported accelerators are ending. Demand for GLM models has made that math unavoidable.
WebProNews is an iEntry Publication