The archive · AI & Models · Strategic decision · 2019–2026
Zhipu's open-source GLM bet pays off: HK$57.9B HKEX IPO as China's first LLM stock
Tsinghua-spawned Zhipu bet open-source, agent-native GLM models could fund an AGI quest; it listed on HKEX in Jan 2026 at a HK$57.9B market cap.
Z.ai (Zhipu AI, 智谱)
What the business is
Zhipu builds the open-source GLM family of large language models, sells enterprise access through its MaaS (model-as-a-service) platform, and ships consumer agent products such as AutoGLM.
How it started
Founded in 2019 from Tsinghua University technology transfer, Zhipu built the GLM architecture in 2021 and launched its MaaS development and commercialization platform the same year. The founding bet, per chairman Liu Debing: 'let machines think like humans,' with the Z in the name standing for the ultimate state of intelligence.
What happened
Revenue grew from ¥57.4M (2022) to ¥125M (2023) to ¥312M (2024), a 130% CAGR, then hit ¥191M in H1 2025 (+325% YoY) against widening R&D spending (¥2.2B in 2024; ¥1.59B in H1 2025, about 8.4x revenue). In July 2025 Zhipu released GLM-4.5, its first agent-native model (355B params flagship, 106B Air), open-sourced under MIT; in August 2025 it shipped AutoGLM 2.0, billed as the world's first mobile agent; in December 2025 GLM-4.7 topped open-source leaderboards. Local deployment accounted for about 85% of H1 2025 revenue.
How it ended up
Listed on HKEX on 2026-01-08 as China's first LLM IPO: priced at HK$116.2, opened at HK$120, closed at HK$131.5 for a HK$57.89B market cap. MiniMax listed the next day. CEO Zhang Peng had told CNBC the model could run on just eight Nvidia H20 chips.
Background
Zhipu (now Z.ai) was founded in 2019 from Tsinghua University technology transfer and launched the GLM architecture in 2021 alongside a MaaS development and commercialization platform. Its bet was that China would reward an open-source, agent-first foundation-model strategy — releasing models freely under MIT while charging enterprises for managed access.
The strategy produced fast iteration: revenue grew from ¥57.4M (2022) to ¥312M (2024), a 130% CAGR, then ¥191M in H1 2025 (+325% YoY), while R&D spending reached ¥2.2B in 2024 and ¥1.59B in H1 2025 — roughly 8.4 yuan of R&D per yuan of revenue. GLM-4.5 (July 2025) was Zhipu's first agent-native model, open-sourced at 355B parameters, and AutoGLM 2.0 (August 2025) was billed as the world's first mobile agent.
In December 2025 GLM-4.7 topped the Artificial Analysis and Code Arena leaderboards for open-source and domestic models. On January 8, 2026, Zhipu listed on HKEX as China's first large-model IPO, closing at HK$131.5 for a HK$57.89B market cap; MiniMax followed the next day.
What has to be true
- Open-source MIT releases gave GLM distribution and developer mindshare that closed rivals could not match, building a pipeline of enterprise MaaS customers.
- Local deployment — about 85% of H1 2025 revenue — showed large Chinese enterprises would pay for managed, private model access even while cloud pricing was loss-making.
- Being first to list turned China's LLM capital race into a public-market story, giving Zhipu liquidity and attention ahead of MiniMax and Moonshot.
- The 2–3 month base-model iteration cadence kept GLM at the top of open-source benchmarks, which became the headline traction that supported the IPO.
What can be applied
Open-sourcing the core model can be distribution, not giveaway: MIT releases built mindshare and MaaS revenue, and being first to list turned an R&D-heavy loss-maker into a public story.
Aftermath
As of September 2026, Zhipu trades on HKEX and continues GLM iteration; chief scientist Tang Jie's IPO-day letter promised GLM-5 and a new frontier-research X-Lab. The company remains loss-making, with cloud pricing and a price war among Chinese model vendors still pressing on gross margin, per its prospectus reporting.
Sources
- 智谱上市：一场关于AGI商业化的资本测试
- Zhipu launches GLM-4.5 model as China ramps up open-source AI race
- China's Zhipu Unveils World's First Mobile Agent, Turning Phones Into AI Assistants
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card