Breaking

Zhipu AI in Early Talks With Domestic Firms Over Custom GLM Chip

July 7, 2026 at 10:30 EDT

Infra & Chips
Foundation Models
Policy & Regulation

Chinese AI lab Zhipu AI, known internationally as Z.ai, is in early talks with several domestic chip design firms to develop a custom AI chip (ASIC) optimized for its GLM models. The move, driven by surging model usage and U.S. export controls, aims to reduce reliance on Nvidia and Huawei. No partner has been selected and the discussions remain at an early stage, according to reporting.

Z.ai (Zhipu AI) · GLM Series

China's Zhipu Weighs Its Own AI Chip as GLM Demand Explodes

Cut off from Nvidia's top GPUs by US export controls, the Tsinghua-born lab is in early talks to design custom silicon — while its open-source GLM-5.2 rivals frontier models at a fraction of the cost.

27×

Peak surge in developer token usage during GLM-5.2's first week

~1/5

The cost of matching a leading US flagship model's performance

Maximum context length, tokens

First-Week Token Usage Surge

Demand for GLM-5.2 climbed as much as 27-fold — the pressure now driving custom-chip talks.

1×

Baseline

27×

Peak week 1

Benchmark Edge

GLM-5.2 62.1%

GPT-5.5 58.6%

SWE-bench Pro (coding) · MIT open-source license

Domestic Chip Shift

Huawei Ascend
Cambricon
Moore Threads
T-Head

Ascend 910C ≈ 60% of H100 inference performance

Why Build Its Own Chip?

US Entity List

No high-end Nvidia GPUs

→

Soaring GLM Demand

Compute pressure builds

→

Custom Silicon Talks

2+ years to deliver

Early discussions with several Chinese chip-design firms; no partner, specs, or timeline confirmed.

Developer praise

Strong agentic performance and cost efficiency; top-tier Chinese-character generation and text rendering among open-source models.

Open questions

H100-class hardware may have been used in training; domestic chips still face supply, ecosystem maturity and CUDA lock-in hurdles.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.