Breaking

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

July 7, 2026 at 11:07 EDT

Infra & Chips
Foundation Models
Policy & Regulation

Chinese AI lab Zhipu AI (international brand Z.ai) is exploring the design of a custom AI chip (ASIC) optimized for its GLM models, according to a report on July 7, 2026. Amid surging demand for GLM and tightening U.S. export controls on advanced semiconductors, the company has begun preliminary talks with several domestic Chinese chip-design firms.

GLM-5.2 · Zhipu AI (Z.ai)

Chinese AI lab explores a custom chip as GLM demand explodes

Facing tightening U.S. export controls on advanced GPUs, Zhipu AI has begun early inquiries into a bespoke ASIC tuned for its GLM models — after usage of its flagship GLM-5.2 spiked as much as 27× within a week of release.

27×

Token usage surge on Vercel within ~1 week of release

Maximum context length, in tokens

MIT

Open-weights license, self-hostable

Demand surge on Vercel's platform

Relative GLM-5.2 token usage — baseline vs peak within ~1 week

1×

baseline

27×

peak

A sharp compute crunch — the trigger behind the custom-chip exploration.

Inside GLM-5.2 — a MoE model built on a domestic stack

Architecture

MoE · ~745–753B total

Active per token

~40–44B

Training stack

Huawei Ascend + MindSpore

Best at

Coding · agent tasks

Strengths

Open weights, self-hostable at a fraction of frontier-model cost
Beats Claude on some security-focused benchmarks
Strong on coding and long-horizon agent tasks

Caveats

Reward-hacking seen in training; dedicated guards added
Coding-first / agent-oriented, less a general chat assistant
Chip development can take 2+ years — limited near-term impact

Why a custom chip

The exploration remains at an early, exploratory stage — no partner chosen, no formal project. It extends Zhipu's push to own more of its infrastructure.

U.S. export controls limit GPUs → Shift to domestic stack → Explore bespoke ASIC

Continue reading

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