China's Zhipu AI begins trading in HK after $558m IPO

Zhipu, an AI startup based in Beijing, began trading in Hong Kong after raising US$558 million in its IPO.

Also known as Knowledge Atlas Technology JSC Ltd., it is the first major Chinese generative AI firm to list publicly.

The company offered 37.4 million shares at HK$116.20 (US$14.9) each, with retail investor allocation oversubscribed by more than 1,159 times.

Ahead of its debut, shares rose as much as 35% in gray market trading.

Founded in 2019 by researchers from Tsinghua University, Zhipu reported revenue of 312.4 million yuan (US$44.6 million) in 2024.

The company is backed by Alibaba, Tencent, and several local government funds.

Its listing comes amid a wave of Chinese semiconductor and AI companies raising new funding, as US export controls continue to restrict access to advanced chips.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Zhipu’s IPO pricing signals investor appetite, yet profits remain undisclosed

Retail oversubscription hit 1,159x while the prospectus lists research and development (R&D) spend of RMB 2.1954 billion against RMB 312.4 million in 2024 revenue 1.
R&D outlay reached RMB 1.59 billion in the first half of 2025 1; cash burn likely continues as model work takes priority over profit.
Revenue more than doubled in 2023 and again in 2024 1; with R&D staff at 74% of headcount 1 unit economics stay opaque without gross margins or customer acquisition costs.
The company is shifting to standardized application programming interface (API) services targeting 50% of revenue 2; the US$558 million IPO roughly matches a year of R&D spend.

Enterprise SaaS teams can plug in Zhipu’s lower-cost APIs for bilingual builds

The GLM-4.5 API is priced at RMB 0.8 per million input tokens and RMB 2 per million output tokens (tokens are billing units for text) 3; these rates undercut many proprietary models, boosting developer margins.
Zhipu serves 12,000+ enterprise customers and 45 million+ developers 1; third parties cite strong domestic standing 4; bilingual support helps SaaS teams ship local AI features without retraining.
The MIT-licensed GLM-4.6 model uses a 355 billion-parameter Mixture of Experts (MoE) design 4; enterprises with data privacy needs can self-host to remove recurring API fees at high volume.
A 200K token context window 4, the text span the model reads at once, suits document-heavy fields like legal tech, finance, or research tools where many rivals charge premiums.