A company born inside a Chinese university lab just became a public-market experiment in how much the world will pay for a domestic large model (大模型). On 8 January 2026, Zhipu AI (智谱) opened trading on the Hong Kong Stock Exchange and closed 13% above its offer price. For a sector that has burned cash faster than it has earned it, the debut was less a victory lap than a stress test of a new funding model.
The numbers that actually printed
The pricing and first-day tape told a clear story:
- Offer price: HK$116.20 per share, raising about HK$4.35 billion (≈ US$558 million) before the over-allotment option.
- Listing-day close: HK$131.50, a 13% first-day gain and a market capitalisation of about HK$57.89 billion (≈ US$7.4 billion).
- Demand: the Hong Kong public offer was oversubscribed 1,159 times; the international tranche 15.3 times.
- Cornerstone cushion: 11 investors — Beijing state capital, insurers, funds and strategic backers — took HK$2.98 billion (≈ US$382 million) of stock, roughly 70% of the deal.
Zhipu (legal name: Knowledge Atlas Technology JSC Ltd) trades as 02513.HK. It is, by its own prospectus wording, the first publicly listed company built around an AGI foundation model.
Before the listing: eight rounds and a wall of money
Zhipu was founded in 2019 as a spinout from Tsinghua University's Knowledge Engineering Lab (KEG), the group behind the GLM pre-training architecture. Its funding history reads like a map of China's AI-investor universe:
- Eight private rounds, raising more than RMB 8.3 billion (≈ US$1.15 billion) in total.
- Investors span Alibaba, Tencent, Meituan, Ant Group and Xiaomi, layered on top of local state-owned funds from Hangzhou, Zhuhai and Chengdu.
- The final pre-IPO round (May 2025) valued the company at about RMB 24.4 billion (≈ US$3.4 billion) post-money — roughly 2.1× the IPO issue valuation, meaning late entrants more than doubled on paper within eight months.
That last point is the quiet headline. A company losing money hand over fist still convinced institutions to pay a premium eight months before the public priced it lower.
Why Hong Kong, and why now
Hong Kong gave Zhipu something the mainland A-share market could not easily offer in 2026: a deep, US-dollar-adjacent pool of international investors and a listing path tolerant of pre-profit technology names. The exchange had just come off its strongest IPO year since 2021, and a cluster of AI and chip names — Biren, Zhipu, MiniMax — listed within a single week in January 2026.
Zhipu also structured the offer to signal discipline. About 70% of net IPO proceeds go to R&D of general-purpose AI large models — pre-training, reasoning and AI agents (智能体). A further ~10% funds the MaaS (Model-as-a-Service) platform, and another ~10% goes to the partner network and strategic investment.
What the money is buying
The R&D budget is enormous relative to revenue, and that is the point. Figures from the prospectus:
- 2024 revenue: about RMB 312 million; H1 2025 revenue: RMB 190.9 million, up 325% year on year.
- H1 2025 adjusted net loss: about RMB 2.36 billion, with R&D spend of RMB 1.595 billion — roughly eight times revenue for the period.
- Total liabilities at mid-2025: about RMB 11.25 billion, after an IPO that lifted cash by roughly HK$4 billion.
In plain terms: Zhipu is selling compute and models cheaply — its coding tool costs as little as ¥20 (under US$3) a month, about one-seventh of a US rival — to buy market share and sovereign-AI contracts across Southeast Asia and Belt and Road countries. The bet is that volume and government anchoring beat near-term profit.
The honest caveats
Three facts deserve to be said out loud.
First, the valuation is rich. At the offer price, Zhipu traded at roughly 147× its 2024 revenue (per New Fortune analysis) — a multiple that assumes the large-model market compounds for years without a price war destroying margins.
Second, the cloud-deployment business is already compressing. As token volume exploded — from 500 million to trillions of tokens a day — gross margin on hosted services fell toward zero in 2025 as Zhipu cut prices to win users.
Third, losses are structural, not a one-off. Management has said outright that profitability is not the near-term priority. The IPO is explicitly a way to keep the R&D tap open.
What readers can do now
- If you follow AI infrastructure investing, treat Zhipu's 70%-to-R&D structure as the template: public Chinese model companies are funding compute through equity, not profit. Watch recurring R&D intensity, not headline revenue.
- If you build on Chinese models, test Zhipu's MaaS API latency and pricing against Qwen and DeepSeek today; the "one-seventh of US rivals" claim is worth benchmarking yourself.
- If you report on the sector, read the IPO as a sovereignty signal: a Tsinghua lab reaching public markets means foundation models are now treated as strategic infrastructure, not just software.
Honest limitations
Core facts (HKEX listing 8 Jan 2026; offer HK$116.20; ~HK$4.35B raised; close HK$131.50; HK$57.89B market cap; 1,159× HK / 15.28× intl oversubscription; 11 cornerstone taking HK$2.98B; 02513.HK; Tsinghua KEG / GLM origin; 8 private rounds >RMB 8.3B; May 2025 post-money ~RMB 24.4B; 70% of proceeds to R&D; 2024 revenue RMB 312M; H1 2025 revenue RMB 190.9M and adjusted net loss RMB 2.36B; mid-2025 liabilities ~RMB 11.25B; GLM-4.7 Dec 2025) come from China Daily / Reuters-sourced reporting, Global Times, CGTN, China News Service, Securities Times, Financial News and the company's prospectus. Currency conversions use ≈ 7.8 HKD/USD and ≈ 7.2 RMB/USD. The "147× 2024 revenue" multiple is a third-party (New Fortune) calculation; the low-price strategy and margin outlook are company statements and analyst interpretation, not independent audits. This article is current to 29 September 2026 and is not investment advice.