1月8日，智谱登陆港股，首日股价较发行价上涨超13%，成功拿下"大模型全球第一股"。从北京到香 港，从高校实验室到股票交易所，智谱在资本市场的一小步，也是大模型在公众视野的一大步。 尽管大模型成熟期言之尚早，其成长历程不可能只有上市这一次考试，但大考毕竟广受关注。 大考智谱走过的路，是中国大模型产业从技术探索走向商业验证的缩影。智谱董事长刘德兵说，IPO最 重要的是对"技术逻辑能否跑通商业逻辑"的实战检验。 智谱上市踩中了人工智能资本化的风口期。近段时间，国产GPU摩尔线程和沐曦股份上市接连引爆市 场。国产大模型赛道的智谱敲钟，再度推高市场对AI企业的商业化预期。 智谱上市，恰逢人工智能从参数竞赛转向场景落地的关键节点，是从大模型的浅水区游向深水区。水浅 意味着近岸和安全，我们见识了诸多"小龙""小虎"的热闹非凡；水深象征着危险，有时候需要持久的潜 行，但只有深水区才能承载楼船巨舰。 如何将宏大的技术叙事，转化为稳健的财务数字和清晰的盈利前景。投资者的耐心是有限的，他们追问 的不再只是"模型有多聪明"，还关心"成本如何下降""盈利何时到来"…… 智谱一家诠释不了，行业必须共同回答：在解决了"从0到1"的技术突 ...

行业研究丨深度报告丨软件与服务 [Table_Title] 从智谱与 MiniMax 看大模型商业化路径 %% %% %% %% research.95579.com 1 丨证券研究报告丨 报告要点 [Table_Summary] 2025 年中国大模型行业从百模大战逐步收敛至头部玩家梯队，智谱 AI 与 MiniMax 作为不同商 业化路线的代表，将于 2026 年 1 月 8 日、9 日相继在港 IPO，作为全球第一批上市的大模型 公司，对产业具有风向标意义。当前时点，大模型的发展范式正式迈向下一阶段。模型玩家从 早期的参数竞赛与任务性能比拼，转向以"场景-数据-模型"正向循环为核心的新阶段。谁能在 真实场景中扎根更深、产品体验更顺、生态协同更紧，谁就能在下一个竞争阶段赢得主动。 分析师及联系人 [Table_Author] 宗建树 郭敬超 宋浪 SAC：S0490520030004 SAC：S0490525120002 SFC：BUX668 刘思缘 请阅读最后评级说明和重要声明 2 / 28 %% %% research.95579.com 2 软件与服务 cjzqdt11111 [Table_Titl ...

Core Viewpoint - The launch of Zhipu's IPO marks a significant step for the AI large model industry, transitioning from heavy R&D investments to capital market engagement [1] Group 1: IPO Details - Zhipu officially started its Hong Kong IPO process on December 30, aiming to become the "first global large model stock" [1] - The subscription period will last until January 5, 2026, with trading set to begin on January 8 under the stock code "2513" [3] - The company plans to issue over 37 million H-shares at a price of HKD 116.2 per share, aiming to raise approximately HKD 4.3 billion, with an expected market capitalization exceeding HKD 51.1 billion [3] Group 2: Investor Confidence - A notable aspect is the strong backing from cornerstone investors, with 11 institutions committing to subscribe for HKD 2.98 billion, accounting for nearly 70% of the total offering [5] - This indicates a significant level of confidence from both international and domestic long-term capital in Zhipu [5] Group 3: Industry Context - Zhipu initially aimed for an A-share listing but shifted to Hong Kong due to delays in regulatory feedback from the China Securities Regulatory Commission [6] - The Hong Kong market appears to be more accommodating for unprofitable tech companies, with faster review processes, making it a more suitable choice [8] Group 4: Industry Dynamics - The IPO highlights the current state of differentiation within the large model industry, with some players exiting the foundational model competition to focus on vertical applications [10] - Zhipu primarily targets enterprise clients in sectors like finance, internet, and healthcare, while competitors like MiniMax focus on individual users [10] Group 5: Financial Performance - Zhipu reported revenues of less than CNY 200 million in the first half of the year, with losses exceeding CNY 2.3 billion, primarily due to R&D costs nearing CNY 1.6 billion [13] - This reflects a common trend in the large model industry, where even leading companies face ongoing financial challenges [12] Group 6: Technological Foundation - Zhipu's technological foundation is built around its GLM framework, with continuous iterations since its first proprietary pre-trained model was released in 2021 [15] - The company has launched multiple versions of its models, achieving significant milestones in the open-source community and expanding its offerings to include multimodal models [17] Group 7: Business Model and Challenges - Zhipu has adopted a MaaS (Model as a Service) business model since 2021, packaging its model capabilities into services that support both local and cloud deployments [19] - Despite having a solid technological base and a clear commercialization path, Zhipu faces challenges due to increasing competition and the difficulty of scaling its enterprise-level services [19] Group 8: Future Outlook - The funds raised from the IPO will address immediate financial needs, but the real test will be converting technological advantages into sustainable profitability [21] - The IPO serves as an important exploration for the large model industry, assessing how much capital markets are willing to invest in long-term AI initiatives [21]

Core Viewpoint - The company Zhipu has officially launched its IPO, aiming to become the "first stock of global large models," marking a significant milestone for both itself and the domestic large model industry, transitioning from "money-burning R&D" to "capital market validation" [2][3]. Industry Overview - The IPO of Zhipu signifies the beginning of a "淘汰赛" (elimination round) in the domestic large model industry, with increasing competition expected as more players enter the capital market by 2026 [3]. - The AI sector has seen a significant decline in financing, with a 31.2% year-on-year drop in Q1 2025, leading to a situation where IPOs become a crucial funding channel for companies like Zhipu [9]. Financial Performance - Zhipu's revenue is projected to grow from 57.4 million yuan in 2022 to 312.4 million yuan in 2024, reflecting a compound annual growth rate (CAGR) of 130% [3][5]. - However, adjusted net losses are expected to increase from 97 million yuan in 2022 to 2.466 billion yuan in 2024, indicating a 25-fold increase in losses over three years [5]. - In the first half of 2025, Zhipu reported a net loss of 1.752 billion yuan, nearly matching the total loss for 2024 [5]. R&D Investment - Zhipu's cumulative R&D investment reached 4.4 billion yuan from 2022 to the first half of 2025, with 2.195 billion yuan spent in 2024 alone [5][9]. - A significant portion of R&D costs, approximately 70%, is attributed to computing power expenses, which amounted to 1.553 billion yuan in 2024 [5]. Business Model and Market Position - The majority of Zhipu's revenue (84.8% in the first half of 2025) comes from localized deployments for B-end clients, while cloud services account for only 15.2% [6]. - Despite having a strong technical foundation, Zhipu's business efficiency is low, with an average daily revenue per employee of only 1,189 yuan, significantly lower than competitors [6][10]. - Zhipu holds a mere 6.6% market share in the domestic general large model market, which is limited when compared to major players like Baidu, Alibaba, and Tencent [6][10]. Competitive Landscape - The competitive environment has shifted from a focus on talent to a "算力军备竞赛" (computing power arms race), with major companies planning substantial investments in AI infrastructure [9][10]. - Major players like Baidu and Tencent are expected to spend between 30 billion to 100 billion yuan on AI in 2025, dwarfing Zhipu's total R&D investment [9]. Future Challenges - Post-IPO, Zhipu will face increased scrutiny and competition, with quarterly performance comparisons becoming a norm [11]. - The company must quickly find a profitable business model to avoid being integrated by larger firms, similar to trends seen in the food delivery and ride-hailing industries [13][14].

据华尔街见闻此前文章， 英伟达近日与Groq达成了一项非独家的技术许可协议。 按照披露，英伟达将把Groq的AI推理技术整合进未来产品体系中，而Groq创始人兼首席执行官Jonathan Ross、总裁Sunny Madra以及部分核心工程人员将加 入英伟达。Groq公司本身仍保持独立运营，其云业务Groq Cloud也将继续对外提供服务。 然而，如果只把它理解为普通的技术合作，显然过于表面。技术可以授权，但一家芯片公司的创始人和核心架构团队，很少作为"附带条款"整体迁移。 英伟达真正看中的，从来不是Groq的收入规模，而是它背后的架构思想。 而这套思想，与谷歌TPU高度同源。 业内普遍认为， 随着AI竞争重心从训练转向推理，GPU长期建立的统治优势开始出现松动，TPU在效率与成本结构上的优势正逐步显现，并有望成为谷歌云 未来十年的关键护城河 ，这一背景下，黄仁勋第一次显露出被逼到墙角的焦虑。 可以肯定的是，一旦英伟达借助这次技术引入在推理架构上追近甚至抹平与谷歌TPU的差距，原本在谷歌与OpenAI／英伟达阵营之间不断扩大的技术与生态裂 口，很可能会迅速收敛，竞争格局也将重新回到拉锯状态。 AI叙事正在从训 ...

作者 | 吴梅梅 排版 | 王梓璇 来源 |IT桔子 封面 |豆包AI生成 中国AI大模型头部公司的基调在2025年发生了转变：从早期的宏观规模叙事转向切实的经营盈利压力。 与此同时，其组织架构、高管人事变动也颇为频繁。IT 桔子曾盘点过这两年从大模型独角兽中离职、再次创业的公司已经有多家。（参见） 2025年12月，中国AI大模型领域的资本市场迎来关键节点。12月19日，智谱AI（Zhipu AI）正式提交港股招股书并通过上市聆讯，计划募资约3亿美元， 最快于2026年初挂牌交易。 仅两天后，MiniMax也于12月21日顺利通过港股上市聆讯，拟在2026年1月5日前后启动招股，计划募资6-7亿美元，估值约40亿美元。 另据多家媒体报道，月之暗面计划2026年下半年启动IPO，官方未确认具体时间表，但知情人士称上市筹备已在进行，正在评估港交所与纽交所双重上市 的可能性。 智谱AI（2019年成立，清华系背景）与MiniMax（2021年成立）都是非常年轻的大模型创业公司，距今成立不过五六年的时间。 他们选择此时冲刺港股IPO，是多重因素交织下的战略决策，核心可归结为"VC融资模式难以为继、资本化窗口迫在眉睫 ...

经历三年的军备竞赛，大模型行业迎来了资本价值兑现的关键节点。 12月19日，北京智谱华章科技股份有限公司（下称"智谱"）正式披露招股书，向港交所IPO发起冲击。不到48小时，上海稀宇科技有限公司（下 称"MiniMax"）同样向港交所披露了招股书。更早之前，这两家AI独角兽相继通过了港交所聆讯。 如此密集的资本动作，瞬间将"谁将是AI大模型第一股"的悬念拉满。 值得关注的是，二者虽然目标一致，但走出了截然不同的发展路径。智谱以B端、G端服务为核心；MiniMax以C端订阅为支柱，深耕全球化用户市场。 两种路径的背后，是技术理念、商业逻辑与估值逻辑的全方位碰撞。一直以来，商业化的效率之争都是行业的比拼重点，一定程度上决定了大模型企业的 估值和想象力。 当大模型行业从技术探索迈入商业化深水区，这场资本市场的争夺，更像是一场商业路径可行性的测试。 另一个变化在于，AI大模型赛道已从两年前的群雄逐鹿，逐步聚焦到少数具备核心竞争力的头部玩家。头部玩家投入逐年增大，争夺资本市场的关注与 资源，就是争夺市场话语权，也是为接下来的扩张收集弹药。 压力和动力总是结伴而行。成功上市只是第一步，更关键的是后续资本故事的支撑，而这些 ...

编辑部 整理自 MEET2026 量子位 | 公众号 QbitAI 在大模型参数竞赛卷到极致的今天，AI真正要跨过的门槛，已不再是"更强的能力"，而是"如何在行业里真正活起来"。 技术演进的焦点也随之从规模扩张转向一个更本质的问题：智能究竟如何在物理世界中持续产生价值。 正是在这一关键拐点上， 卓世科技合伙人、副总裁 赵 策 在量子位MEET2026智能未来大会上，提出了一个与主流截然不同的判断： 大模型的下一场竞争，不在模型本身，而在模型、终端、数据与业务流构成的自驱动闭环。 在这套闭环中，云端不再是智能的唯一舞台，终端成为感知物理世界的入口，数据回流则不断反哺模型，让智能体在企业真实业务流里"长期 在线、持续进化"。 也正是基于这套路径，他们走进 企业服务、工业制造、医疗健康、文教传媒、养老、水利、园区 等不同场景里，用一组组节能降耗、人效提 升、收益改善的数字，回答"大模型到底值不值"的问题。 为了完整体现赵策的思考，在不改变原意的基础上，量子位对演讲内容进行了编辑整理，希望能给你带来更多启发。 MEET2026智能未来大会是由量子位主办的行业峰会，近30位产业代表与会讨论。线下参会观众近1500人，线 ...

Investment Rating - The report maintains a "Positive" investment rating for the software and services industry [8]. Core Insights - On October 27, Xiyu Technology officially open-sourced and launched MiniMax M2, a model with a total parameter count of 230 billion, specifically designed for agent and code applications. The complete weights of M2 are fully open-sourced under the MIT license and are available globally for a limited time free of charge. The MiniMax Agent has also launched a domestic version and upgraded its overseas version [2][5]. - The launch of M2 opens new possibilities for open-source models in intelligent execution and enterprise applications, with the potential for accelerated commercialization of large models. The report emphasizes the importance of cost reduction effects of the models and continues to favor the domestic AI industry chain, recommending shovel stocks and major players with significant positioning advantages [2][10]. Summary by Sections Event Description - The report details the launch of MiniMax M2, which features a MoE architecture and is tailored for agent and code applications. The model's complete weights are open-sourced and available for free globally for a limited time. Additionally, the MiniMax Agent has launched a domestic version and upgraded its overseas version [5]. Event Commentary - MiniMax M2 has demonstrated exceptional performance in various benchmarks, including a SWE-bench Verified score of 69.4, placing it among the top models for real programming tasks. The model also achieved a score of 61 in the Artificial Analysis test, ranking fifth overall and first among open-source models. In terms of tool usage, it scored 77.2 in the τ²-Bench test, leading among domestic models [10]. - The model's architecture focuses on executable agent tasks, ensuring that every reasoning step has complete context visibility. The interleaved thinking format allows the model to plan and verify operations across multiple dialogues, which is crucial for agent reasoning [10]. - M2's pricing is competitive, with input costs around $0.3 per MToken and output costs approximately $1.20 per MToken, significantly lower than competitors. The model also offers a TPS (tokens per second) output of around 100, which is rapidly improving [10]. - The market response to M2 has been enthusiastic, with it ranking first on OpenRouter and HuggingFace trend charts. The model has surpassed 50 billion daily token consumption, indicating strong market interest and potential for commercial application [10].

Investment Rating - The report maintains a "Positive" investment rating for the industry [8]. Core Insights - On October 30, the AI programming platform Cursor announced the upgrade to version 2.0 and launched its first self-developed programming model, Composer, designed for low-latency coding, capable of completing most interactive tasks within 30 seconds [2][5]. - The report suggests that Cursor is transitioning from an AI programming tool to an AI development platform, with the commercialization of large models expected to accelerate [2][10]. - The report emphasizes the importance of cost reduction in token consumption as a core factor affecting the current market [2][10]. Summary by Sections Event Description - Cursor's upgrade to version 2.0 includes 15 major feature enhancements, focusing on a new interface for parallel collaboration among multiple agents [5]. Event Commentary - The Composer model balances performance and speed, completing most tasks in under 30 seconds and achieving output speeds exceeding 200 tokens per second, which is four times faster than comparable intelligent models [10]. - The model utilizes a mixture of experts (MoE) architecture and low-precision training to enhance efficiency and reduce inference costs, potentially accelerating product expansion [10]. - The new system allows for up to 8 agents to run in parallel, enhancing team collaboration and overall performance, with cloud agent reliability reaching 99.9% [10]. - The report highlights the need for further balance between cost and precision in the multi-agent model due to token consumption and management complexity [10].