Breaking

Zhipu Weighs Its Own AI Chip Tuned for GLM Models

July 8, 2026 at 08:30 EDT

Infra & Chips
Foundation Models
Open Source

Chinese AI lab Zhipu AI (Z.ai) is exploring the development of a custom AI chip optimized for its GLM family of large language models. The move comes as U.S. export controls make it harder to source advanced GPUs while demand for GLM models surges, turning compute into a growing bottleneck.

Zhipu AI · Z.ai

Zhipu Explores a Custom Chip to Feed Its GLM Models

As U.S. export controls squeeze access to high-end GPUs and demand for GLM-5.2 surges, the Beijing lab is weighing an in-house ASIC tuned for inference — pushing model-to-hardware integration entirely inside China.

27×

Daily token usage jump on the aggregator in GLM-5.2's first week

Context length supported by GLM-5.2 (MoE architecture)

2+ yrs

Estimated time for a custom chip to reach practical use

The Demand Trigger

GLM-5.2 daily token usage — first week on aggregator

1×

Launch
day

27×

One week
later

Surging inference demand meets GPU scarcity — turning compute into the bottleneck driving the chip plan.

The Path to Domestic Silicon

NVIDIA GPUs

limited by export controls

→

Domestic chips

Ascend · Cambricon · Moore Threads

→

Custom GLM ASIC

tuned for inference (exploratory)

Developers praise

"A viable substitute for Opus-class models in real work"
"Top-tier open weights with strong value for money"
Open weights under the MIT license

Developers caution

"Benchmarks look good but it underdelivers in real agent loops"
Trails leading commercial models on some tasks; verbose reasoning
Variability across providers and data-privacy concerns

Continue reading

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