Zhipu AI (2513.HK) ☆

Zhipu (Knowledge Atlas Technology, marketed as Z.ai) is a Tsinghua University spinoff that develops the GLM family of large models; it listed in Hong Kong in January 2026 under Chapter 18C of the Listing Rules, the regime for specialist technology companies. Its business turned over within a year. In the six months to 2026-06-30, per its interim results announcement, open platform and API revenue was RMB825.2 million, 86.5% of the total against 15.2% a year earlier, while on-premise deployment revenue fell 20.5% to RMB128.7 million; the company says it "proactively scaled back its on-premise model software licensing business". The demand behind that shift rests on model capability, which independent trackers rate highly: Let's Data Science (2026-06-27) reports that on Artificial Analysis's Intelligence Index GLM-5.2 "is now the top-ranked open-weights model overall and sits fourth on the full leaderboard", and OpenRouter said GLM 5.2 "has had a very fast uptake relative to most open-weight launches" (Gigazine, 2026-06-23). The same reports tie part of that uptake to timing: GLM-5.2 "landed days after U.S. export controls forced Anthropic to pull Claude Fable 5 and Mythos 5 worldwide in mid-June". What the record does not show is protection. The weights are MIT-licensed and can be self-hosted; China's MaaS market was led by Volcano Engine at 49.5% in 2025 per IDC (KrASIA); Reuters reports rivals cutting prices and racing to launch models; and Zhipu lost RMB2,071,992 thousand in the half while spending RMB2,131,171 thousand on R&D, with group gross margin down to 26.4% from 50.0%. It was also added to the US Entity List on January 16, 2025, and now runs inference "across a cluster of over 100,000 domestic chips". Zhipu is a leading open-weight lab whose position has to be won again with each release, at a loss, so it has no moat yet.

評級: 無

The record shows a frontier-grade model developer with nothing yet protecting its profits. Zhipu's interim results announcement for the six months to 2026-06-30 reports revenue of RMB953.9 million against a loss for the period of RMB2,071,992 thousand and R&D expenses of RMB2,131,171 thousand, and says gross profit margin fell from 50.0% to 26.4% as lower-margin cloud revenue grew. Its prospectus (https://www1.hkexnews.hk/listedco/listconews/sehk/2025/1230/2025123000017.pdf) warns that "even if our models are more advanced, effective and cost-efficient than those of our competitors, current or potential customers may accept competitors' models in lieu of ours", and that competition "may also result in continued pricing pressures". Reuters, via The Standard (https://www.thestandard.com.hk/finance/article/341440/Chinas-Zhipu-AI-revenue-quintuples-in-first-half-loss-narrows, 2026-08-31), reports rivals including Alibaba, ByteDance and Moonshot "cutting prices and racing to launch new models". KrASIA's report on IDC data (https://kr-asia.com/how-bytedances-volcano-engine-holds-nearly-half-of-chinas-maas-market) finds that the share which held in China's model-as-a-service market in 2025 was ByteDance's Volcano Engine's, credits that to scale, and adds that "For followers that lack economies of scale, matching low prices often means greater cost pressure and may even lead to losses." A capability lead that has to be won again with each model release, while the company runs at a heavy loss, is not yet a moat.

類型: 無

None of the usual moat sources is established in the record. The intellectual property is given away by design: DataNorth (https://datanorth.ai/news/zhipu-ai-releases-glm-5-2, 2026-06-18) reports GLM-5.2 is "released under an MIT license with no regional limits, which permits self-hosting, fine-tuning, and commercial use". The API side shows no lock-in of Zhipu's own: KrASIA recalls that MaaS "was once believed to be a space where large model APIs would have limited stickiness" because "developers only needed to change a few lines of code to replace the underlying model or switch cloud platforms", and although it reports IDC data showing the leader's share holding and says MaaS is becoming enterprise infrastructure "with stronger stickiness", it puts that leader's hold down to call volume, inference engineering and model capability. The on-premise line, where installed deployments would create lock-in, is shrinking: the interim results report enterprise-level general-purpose large model revenue down 54.6% and say some of those customers "gradually migrated to Open Platform and API services". Scale belongs to others in China's MaaS market, where IDC put Volcano Engine's 2025 share at 49.5%. What Zhipu does have is the R&D capability to ship leading open-weight models repeatedly, which is a capability lead rather than a structural moat.

領先地位: 快速追隨者

Independent trackers put its models at the top of the open-weight field, but not at the top of the market. Let's Data Science (https://www.letsdatascience.com/news/chinese-models-narrow-gap-with-anthropic-and-openai-98f17c4b, 2026-06-27), summarising Artificial Analysis's Intelligence Index, says GLM-5.2 "is now the top-ranked open-weights model overall and sits fourth on the full leaderboard behind Claude Fable 5, Opus 4.8, and GPT-5.5". Reuters, via The Standard (https://www.thestandard.com.hk/finance/article/328208/Chinas-Zhipu-posts-132-percent-rise-in-annual-revenue-on-AI-boom, 2026-03-31), calls Zhipu "One of the leading players in China's crowded artificial intelligence sector" and names MiniMax, Moonshot AI, DeepSeek, ByteDance and Alibaba as competitors. The prospectus cites Frost & Sullivan, its commissioned industry consultant, ranking it "first among China's independent developers and second among all developers of general-purpose large models with a market share of 6.6% in terms of revenue in 2024", behind a non-independent provider at 9.4%. That ranking was commissioned and covers 2024, and in China's MaaS market IDC put Volcano Engine at 49.5% in 2025 (KrASIA). Independent trackers put its model at the front of the open-weight field but behind three US closed models, and the only ranking that puts Zhipu near the top of the market by revenue was commissioned and covers 2024, so the band is fast follower rather than co-leader.

定價權: 弱

Zhipu reports raising prices from a low base, but the margin record does not yet show pricing power. The interim results announcement says that "the average API selling price increased by approximately 101%, and Coding Plan subscription prices also went up" while MaaS token invocation volume "had increased over 40 times from the beginning of 2026", and that open platform and API gross margin rose to 24.6%. Those are the company's own figures, an average selling price the announcement does not break down by model, and they coincide with a supply gap: Let's Data Science reports GLM-5.2 "landed days after U.S. export controls forced Anthropic to pull Claude Fable 5 and Mythos 5 worldwide in mid-June". The product is still sold at a discount: DataNorth reports an API price "which Z.ai positions at roughly one sixth the cost of comparable frontier models". Group gross margin fell to 26.4% from 50.0%, Reuters reports rivals "cutting prices", and the prospectus warns of "continued pricing pressures".

依所引來源做出的人工整理護城河判斷。 來源：www1.hkexnews.hk

與其他公司比較…

產業鏈位置

層級: foundation models

• 上游（供應商）： —
• 下游（客戶）： —
• 競爭對手： —

Zhipu is a model layer between compute and applications. Upstream, its prospectus says its suppliers are primarily "providers of computing resources, such as computing hardware and computing services", and records that on January 16, 2025 the company and nine subsidiaries were added to the US Entity List, which "restricts our ability to purchase or otherwise access certain goods, software and technology that are subject to the EAR without a license from the BIS"; its interim results say it now runs large-scale inference "across a cluster of over 100,000 domestic chips". Downstream it sells model access through its MaaS platform, Coding Plan subscriptions and on-premise deployments to enterprises, public-sector bodies and developers, and publishes its model weights openly.

由這家公司已對應的供應鏈關係及其所屬產業層級推導而來。

訊號貢獻 編輯先驗

護城河先驗在長週期（約 6 個月）投出 -0.06 票，權重為 0.20——該週期上的群體判斷是 bullish（分數 0.52，信念度 65）。

編輯先驗——主觀、未經回測。它有界（±0.6）、低信賴度，且從不對短週期與中週期投票。 檢視它究竟如何投票 →

產品線

產品與服務，附編輯判定的市佔區間與技術門檻區間——這些是與所引依據綁定的定性判斷，絕非杜撰的數字。

你正在看的是什麼

這些圖表繪製的是 Zhipu AI 在其 SEC 申報文件中已申報的損益表、現金流量表與資產負債表（截至 2025-12-31）。每一個數字都是申報原值 XBRL——沒有任何一項是預估。

在已申報的區間內，營收成長 1156%，達到 $103.5M（FY2022→FY2025）; 毛利率從 55% 變為 41%; FY2025 淨損 $674.4M。

我們的模型在中週期上對這檔股票偏向 Low-confidence · Neutral——這是對市場方向的判斷，不是對獲利的預測。檢視「訊號」分頁。

申報文件

• 2025Q4 · 申報於 2025-12-31 來源：finance.yahoo.com
• FY2025 · 申報於 2025-12-31 來源：finance.yahoo.com
• 2025Q2 · 申報於 2025-06-30 來源：finance.yahoo.com
• FY2024 · 申報於 2024-12-31 來源：finance.yahoo.com
• FY2023 · 申報於 2023-12-31 來源：finance.yahoo.com

關鍵比率與同業比較

正在與同業比較…

機構持有人（13F）

尚無被追蹤的 13F 持有人。

這可能表示該公司的 CUSIP 尚未為 13F 比對整理過，而不是沒有機構持有它。

策略性企業投資方

• Alibaba Group — Among the investors in Zhipu's 2023 fundraising totaling RMB 2.5B (~$342M), alongside Tencent, Ant, Xiaomi, HongShan and Meituan (individual amount undisclosed) · 2023-10-23 來源：datacenterknowledge.com

尚未追蹤 Zhipu AI 的任何催化劑。