Breaking

Zhipu Weighs Designing Its Own AI Chip for GLM

July 8, 2026 at 19:00 EDT

Infra & Chips
Foundation Models
Open Source

Chinese AI lab Zhipu AI (international brand Z.ai) has entered early talks with several domestic chip design firms about the possibility of designing a custom AI processor (ASIC) for its GLM models. Aiming to cut inference costs and secure computing capacity, it would join OpenAI, Google, Alibaba and ByteDance in pursuing in-house silicon. Shares of the related Hong Kong-listed Knowledge Atlas Technology (2513.HK) reportedly rose 9.9% to 13.4% right after the report.

Zhipu AI · Z.ai

China's Z.ai Weighs Its Own Custom AI Chip for GLM

After a 27-fold surge in daily token usage on GLM-5.2, the Tsinghua-born lab has opened early talks with domestic chip designers to build a custom inference ASIC — cutting costs and reducing NVIDIA dependence under U.S. export controls.

The demand shock that triggered the chip plan

1×

Before

27×

After GLM-5.2

Daily token usage jumped 27-fold, straining compute and pushing the lab toward in-house silicon.

~750B

Total MoE parameters (~40B active)

Token context window

1/5–1/10

Of Claude's cost at similar performance

GLM-5.2 vs leading overseas models

Terminal-Bench 2.1

GLM-5.2

81.0

Claude Opus 4.8

85.0

SWE-bench Pro: 62.1 — beats GPT-5.5. Optimized for coding and agent tasks at repository scale.

Toward a full-stack, self-reliant AI

GLM Models

Open-weight series

→

Custom ASIC

Tuned for GLM inference

→

Compute Secured

Lower cost, less NVIDIA

Already optimized for domestic hardware — Huawei Ascend, Cambricon and Moore Threads — the lab joins OpenAI, Google, Alibaba, ByteDance and DeepSeek in pursuing in-house silicon.

Momentum

Claude-class coding at a fraction of the cost
Reliable on long, repository-scale tasks
Stock in related listing rose 9.9%–13.4% on the report

Caveats

Only "under consideration" — no partners, specs or timeline
Project could take more than two years
Some note stronger censorship in places

Continue reading

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