llms.blog
AI Hardware

Zhipu AI explores custom silicon as GLM-5.2 demand surges

Chinese AI lab Zhipu AI is in early discussions with domestic chip design houses about building a bespoke processor optimized for its GLM model family, according to a report by The Information. The move comes as daily token usage for GLM-5.2 surged 27-fold during its first week of release, straining compute capacity already squeezed by U.S. export controls on advanced semiconductors. The Beijing-based company, which trades on the Hong Kong Stock Exchange as Z.ai, has made preliminary inquiries

axc
axc
2 min

Chinese AI lab Zhipu AI is in early discussions with domestic chip design houses about building a bespoke processor optimized for its GLM model family, according to a report by The Information. The move comes as daily token usage for GLM-5.2 surged 27-fold during its first week of release, straining compute capacity already squeezed by U.S. export controls on advanced semiconductors.

The Beijing-based company, which trades on the Hong Kong Stock Exchange as Z.ai, has made preliminary inquiries with several Chinese ASIC design firms but has not yet selected a partner. The conversations remain exploratory, and any resulting chip would take more than two years to design, test, and bring to production.

The catalyst is straightforward. GLM-5.2, released in June 2026, became the fastest-growing model on Vercel's model aggregator platform, with daily token usage jumping as much as 27 times during launch week. At the same time, U.S. export restrictions have made it increasingly difficult for Chinese AI labs to acquire Nvidia's most capable GPUs, turning compute availability into a structural constraint rather than a cost issue.

ASICs, or application-specific integrated circuits, are processors engineered for particular model architectures rather than the general-purpose computation that GPUs provide. They typically deliver better energy efficiency and lower per-token inference costs once a model's architecture stabilizes, making them economically attractive for labs running high-volume inference workloads.

Zhipu would be following a well-established path. Google, OpenAI, ByteDance, and Alibaba have all developed proprietary chips to reduce dependence on outside GPU suppliers. Hours before The Information's report, Reuters reported that DeepSeek is also pursuing custom silicon to reduce its reliance on both Huawei and Nvidia.

The broader Chinese ASIC ecosystem has expanded since initial U.S. export restrictions took effect. Cambricon Technologies and Biren Technology are among the domestic firms active in the AI chip space, though neither has been named as a prospective Zhipu partner.

For Nvidia, each Chinese lab that transitions inference to domestic alternatives represents a slice of its China data-center revenue that becomes structurally harder to recover, regardless of how export-control policy evolves. The immediate question for Zhipu is execution: chip design, foundry access, and software adaptation must happen simultaneously, and the lab will need to build or expand a semiconductor team to see the project through.

Sources

Zhipu AI explores custom ASIC chip as GLM-5.2 usage surges 27x - Yahoo Finance / Investing.com

China's AI Lab Zhipu Weighs Custom Chip As Demand for its GLM Model Soars - The Information

AI HardwareZhipuChina AI
ShareX (Twitter)LinkedIn

Written by

axc

More to read

  • LLM Text Watermarking in Production: Statistical Logit Biasing, Cryptographic Signatures, and Evasion Vectors

    As regulatory frameworks such as Article 50 of the EU AI Act enforce machine-generated content provenance, text watermarking has transitioned from academic theory to a core component of production LLM serving stacks. Unlike post-hoc classifiers that evaluate perplexity or burstiness and suffer from high false-positive rates on formal or non-native writing, generation-time watermarks embed imperceptible statistical or cryptographic signals directly into the token sampling process. When engineere

    1 min

  • Grokking in Large Language Models: How Weight Decay and Circuit Efficiency Drive Delayed Generalization

    Grokking in Large Language Models: How Weight Decay and Circuit Efficiency Drive Delayed Generalization In standard machine learning paradigms, model generalization closely tracks training loss: as an optimizer minimizes loss on training data, performance on held-out validation data improves in tandem until the model begins to overfit. In 2022, researchers at OpenAI observed a phenomenon that inverted this assumption: small neural networks trained on algorithmic tasks achieved near-zero trainin

    1 min

  • Serverless GPU Inference in Production: Cold Starts, GPU Memory Snapshotting, and Weight Paging Architectures

    Serverless GPU Inference in Production: Cold Starts, GPU Memory Snapshotting, and Weight Paging Architectures Deploying large language models on dedicated cloud GPUs creates an uncomfortable financial trade-off: keeping enterprise accelerators such as NVIDIA H100s or A100s warm 24/7 costs thousands of dollars per instance each month, yet scaling instances to zero introduces severe latency penalties. When traffic arrives at a dormant node, a standard inference server cold start can take anywhere

    1 min