Chinese AI developer Zhipu AI is engaging in initial conversations with domestic chip designers about creating a custom processor tailored to its GLM model family, sources told media. The discussions come as usage of Zhipu's latest release, GLM-5.2, surged dramatically - daily token consumption on one model aggregation platform rose as much as 27 times in the model's first week - and as U.S. export restrictions have limited access to the most advanced foreign GPUs.
The potential move toward a bespoke application-specific integrated circuit, or ASIC, represents a strategic pivot from reliance on general-purpose GPUs supplied by foreign vendors. ASICs are engineered to execute narrowly defined tasks tied to particular model architectures, and once a model's structure stabilizes they typically offer meaningfully better energy efficiency and lower per-token inference costs than off-the-shelf GPUs. That cost profile makes ASICs attractive for organizations primarily focused on high-volume inference workloads rather than exploratory training.
According to the reporting, Zhipu has not finalized a partner and has only initiated exploratory outreach to several Chinese design houses. Any program to produce a custom chip would be multi-year in scope: the lab would need to recruit or expand a semiconductor engineering team, shepherd the design through testing and validation cycles, and adapt its software stack to exploit the new hardware. Those steps, even in a best-case sequence, suggest the effort could extend beyond two years.
From the perspective of foreign GPU vendors, the prospect of Chinese AI developers migrating inference workloads to domestic ASICs is a structural concern. High-end GPUs from foreign manufacturers - in particular the cards widely used for large language model development and deployment - have been the de facto compute substrate for many Chinese AI labs. Each successful transition of inference tasks onto a custom ASIC reduces the addressable market for those GPUs in China and makes that demand more difficult to recapture should export-control regimes change in the future.
The immediate catalyst for Zhipu's inquiries appears to be a combination of rapidly growing demand for GLM-5.2 and a tighter supply environment for advanced foreign chips. The GLM-5.2 release showed exceptional adoption on at least one model-aggregator platform, with daily token usage spiking up to 27 times within the first week of availability. At the same time, U.S. export controls have constrained the ability of Chinese labs to obtain the latest and most capable chips, turning compute availability into an operational constraint rather than a simple cost input.
Industry observers have generally noted that inference-optimized ASICs can materially lower operating costs for mature, high-volume workloads relative to general-purpose GPUs, although savings vary by model architecture and utilization rates. That dynamic is a central part of Zhipu's calculus: if GLM-related inference continues to compound at rates suggested by the early usage numbers, the business case for investing in a bespoke processor strengthens.
The move would place Zhipu on a path similar to several large technology firms that have already developed custom silicon for model inference. Companies operating their own proprietary chips have cited reduced dependency on external GPU suppliers and lower costs for running large models as primary motivations. For Zhipu the rationale is sharpened by regulatory realities - unlike U.S.-based labs that can generally access the latest foreign silicon, Chinese developers face limits that make even nascent domestic alternatives strategically valuable.
On the domestic design side, established and emerging Chinese firms are active in the AI ASIC market. Names such as Cambricon Technologies and Biren Technology have been participants in the space, and the broader domestic ecosystem has expanded since the first rounds of export restrictions took effect. That growth gives Chinese AI developers a larger set of potential partners than was available previously, although none of these firms have been confirmed as Zhipu's chosen collaborator.
For investors in foreign GPU suppliers, the aggregation of these developments matters. Shares of the major GPU provider referenced in reporting continue to trade as a widely held AI infrastructure name, and China has historically represented a meaningful portion of its data-center revenue. Repeated reports that Chinese labs are evaluating or building custom silicon - even if projects are early-stage and carry multi-year timelines - add incremental evidence to a bear case that the vendor's exposure to the China data-center market is subject to structural pressures, not solely political noise.
The most immediate test of Zhipu's initiative will be execution. Designing and deploying a custom ASIC requires navigating complex technical challenges, securing foundry capacity, and reworking software to exploit a new instruction and compute paradigm. If GLM demand indeed continues to expand rapidly, the economic incentive to overcome those hurdles strengthens. A public selection of a design partner would be a clear sign that the effort has moved from exploratory outreach to a committed development program.
Summary
Zhipu AI is in early talks with Chinese chip designers about an ASIC optimized for its GLM models after GLM-5.2 saw daily token usage jump up to 27 times in its first week. The project is at a preliminary stage with no partner selected, and could take more than two years to complete. The push is driven by explosive demand for GLM-5.2 and constrained access to the most advanced foreign GPUs due to U.S. export rules.
Key points
- Zhipu has begun preliminary discussions with domestic chip design houses about a custom ASIC for GLM models - no partner chosen and conversations remain early.
- Daily token usage for GLM-5.2 surged up to 27 times on at least one model-aggregation platform during the model's first week, creating a strong inference demand signal.
- U.S. export controls on advanced semiconductors have tightened access to top-tier foreign GPUs for Chinese labs, elevating the strategic value of domestic silicon solutions.
Risks and uncertainties
- Execution risk - Building and deploying a custom ASIC involves complex design, testing, and software integration work that could extend beyond the projected multi-year timeframe, impacting implementation timelines for AI infrastructure.
- Foundry and supply constraints - Securing manufacturing capacity and navigating semiconductor supply chains are potential obstacles that could delay or complicate delivering a functional ASIC.
- Market and vendor exposure - Continued migration of inference workloads to domestic ASICs could reduce demand for foreign GPU suppliers in the China data-center market, creating strategic revenue pressure for those vendors.