编者荐语： 日经中文网与36氪开展内容交换合作。将精选36氪的精彩独家财经、科技、企业资讯，与读者分享。 以下文章来源于36氪 ，作者邱晓芬 36氪 . 36氪是服务中国新经济参与者的卓越品牌和开创性平台，提供新锐深度的商业报道，强调趋势和价值，我们的slogan是：让一部分人先看到未来。 平头哥自研PPU已成为中国新增AI算力市场的主力芯片之一。 文 ｜ 邱晓芬 封面来源 ｜ 日经中文网 1月22日，接近市场人士称，阿里巴巴集团已决定支持旗下芯片公司平头哥未来独立上市。平头哥是阿里巴巴全资子公司，于2018年成立。 成立8年以来，公司已在计算、存储、网络等领域推出多款性能业界顶级的芯片，据悉，平头哥自研PPU已成为中国新增AI算力市场的主力芯片之一。 互联网公司里， 自研芯片的早期探索者 芯片一直被认为是科技领域最难坑的骨头之一，而平头哥创立之前，阿里巴巴的硬件研发「履历」几乎空白，互联网公司造芯在国内更无先例，平头哥研 发芯片的难度不亚于阿里投入云计算。 虽然现实很骨感，但在这场技术和资本的游戏里，阿里的战略决心不亚于对阿里云的长期投入。 公司成立后，内部芯片研发的路线和模式非常清晰，即重点围绕云端数据中 ...

Core Viewpoint - The independent progress of major internet companies in chip manufacturing has accelerated, with several companies preparing for IPOs, indicating a shift towards commercialization in the chip industry [5][6]. Group 1: Industry Background - The chip manufacturing wave among internet giants began in 2018, driven by geopolitical tensions and the surge in AI computing demand [7][14]. - Initially, these companies focused on creating chips to support their internal operations, but now they face the challenge of proving their viability as independent entities [8][22]. - The shift from internal cost-saving measures to providing market supply of computing power marks a significant evolution in the chip manufacturing strategy of these companies [23][30]. Group 2: Company Developments - Baidu was the first major internet company to enter the chip market, launching its first AI chip, Kunlun, in 2018 [11]. - Alibaba's chip company, Pingtouge, is preparing for an independent IPO, while Baidu's Kunlun chip has also submitted a listing application [5][22]. - Tencent's Suiyuan Technology is also planning to go public, indicating a broader trend of internet companies moving towards independent chip commercialization [5][6]. Group 3: Market Position and Financial Projections - Kunlun's revenue is projected to grow from approximately 2 billion RMB in 2024 to over 3.5 billion RMB in 2025, with expectations of achieving breakeven [23][25]. - Alibaba's Pingtouge has developed a comprehensive product line, including AI inference and general-purpose CPUs, with significant market penetration [27]. - The domestic AI chip market is rapidly evolving, with local manufacturers increasing their market share, projected to reach a 30% penetration rate by 2024 [29]. Group 4: IPO Trends and Valuations - A new wave of IPOs in the semiconductor sector is anticipated, with around 30 companies expected to enter the A-share market by 2025, aiming to raise nearly 100 billion RMB [34]. - The valuation of Kunlun is estimated to be around 21 billion RMB, with market rumors suggesting it could reach 100 billion HKD upon IPO [37]. - Morgan Stanley predicts Pingtouge's potential valuation could range from 25 billion to 62 billion USD, representing 6% to 14% of Alibaba's current market value [38]. Group 5: Challenges and Investment Needs - Despite the optimistic market outlook, many domestic AI chip companies face profitability challenges due to high R&D costs, which often exceed their revenue [39][43]. - The reliance on external funding through IPOs is becoming essential for these companies to sustain their R&D efforts and maintain their growth trajectory [43].

Core Insights - Alibaba's new AI chip "Zhenwu 810E" has been launched, marking a significant breakthrough in domestic AI chip performance and application scale [1][5] - Alibaba is now one of the only two companies globally, alongside Google, to achieve a full-stack AI integration of "AI + Cloud + Chip" [3][15] - The global cloud computing market is dominated by four major players: Microsoft, Amazon, Google, and Alibaba, collectively holding over 80% market share [4] Group 1: AI Chip Development - The "Zhenwu 810E" chip features impressive specifications, including 96GB HBM2e memory, 700GB/s interconnect bandwidth, and a PCIe 5.0×16 interface, surpassing NVIDIA's A800 [5][10] - Alibaba's chip development journey began in 2018, with the establishment of Pingtouge, aimed at overcoming chip supply challenges [9][10] - The chip is a culmination of Alibaba's 17 years of technological investment in building a "Cloud + AI + Chip" ecosystem [6][10] Group 2: Strategic Positioning - Alibaba's full-stack self-research approach contrasts with Microsoft and Amazon's reliance on external technologies, providing a more stable long-term strategy [4][5] - The company plans to invest at least 380 billion RMB over the next three years in cloud computing and AI infrastructure, exceeding its total investment in the past decade [14] - Alibaba's "Tongyun Ge" AI triangle, consisting of Pingtouge, Alibaba Cloud, and Tongyi Laboratory, creates a synergistic effect that enhances both chip performance and cloud service efficiency [10][15] Group 3: Market Impact - The launch of the "Zhenwu" chip and the "Tongyun Ge" AI triangle positions Alibaba as a key player in the global AI infrastructure competition [12][15] - Alibaba's advancements in AI technology are expected to lead to a shift in global AI standards, with the company emerging as a significant competitor against U.S. tech giants [12][15] - The collaboration between chips and cloud services allows for rapid iteration and cost advantages, enhancing Alibaba's competitive edge in the AI market [15]

1月29日，阿里旗下平头哥官网上线一款名为"真武810E"的高端AI芯片，此前被央视曝光的阿里自研芯 片PPU正式亮相。 至此，"通云哥"也就是通义千问大模型、阿里云和平头哥半导体，共同构成阿里巴巴AI能力黄金三角， 在模型、云、芯片层面完成全栈布局。这也是一种追求AI最佳表现的路径选择，记者了解到，阿里巴 巴正在将"通云哥"打造成一台AI超级计算机，寻求在芯片架构、云平台架构和模型架构上的协同创新， 从而实现在云上训练和调用大模型时达到最高效率。 平头哥在2018年成立，在2018年阿里云栖大会上正式官宣。阿里把此前收购的中天微和达摩院自研芯片 业务整合成"平头哥半导体有限公司"，由集团全资控股，推进云端一体化的芯片布局。 平头哥成立以来，发布过多款芯片，比如含光800(AI推理芯片NPU)、倚天710(服务器芯片CPU)，以及 存储和网络芯片、端侧和IOT芯片等。2020年，新一轮AI技术爆发前夕，平头哥在内部启动PPU研发工 作。 据平头哥官网介绍，"真武"PPU采用自研并行计算架构和片间互联技术，配合全栈自研软件栈，实现软 硬件全自研。其内存为96G HBM2e，片间互联带宽达700GB/s，可应用于 ...

继传出预计"独立上市"之后，阿里旗下芯片公司"平头哥"日前再传大消息。 2026年1月29日上午，平头哥官网悄然上线一款名为"真武810E"的高端AI芯片，此前被央视《新闻联播》曝光的阿里自研芯片PPU正式亮相。行 业资深人士认为，这也意味着通义实验室、阿里云和平头哥组成的阿里巴巴AI黄金三角"通云哥"首次浮出水面。 《中国经营报》获悉，阿里巴巴正在将"通云哥"打造成一台AI超级计算机，它同时拥有全栈自研芯片平头哥、亚太第一的阿里云，以及全球最 强的开源模型"千问"，可以在芯片架构、云平台架构和模型架构上协同创新，从而实现用户在阿里云上训练和调用大模型时达到最高效率。 据悉，"真武"PPU已在阿里云实现多个万卡集群部署，服务国家电网、中科院、小鹏汽车、新浪微博等400多家客户。 "真武"PPU的正式亮相，显示了平头哥在芯片领域积累多年的实力。阿里巴巴2009年创建阿里云，2018年成立平头哥芯片公司，2019年启动大模 型研究，经过长达17年的战略投入和垂直整合，终于实现"通云哥"全栈AI的完整布局。 阿里巴巴长期观察人士认为，目前整个阿里AI布局，包括阿里云、平头哥、千问大模型及千问APP及C端硬件四大板块 ...

Core Viewpoint - The article discusses the performance of Hong Kong stock indices and highlights the launch of Alibaba's new AI chip "Zhenwu 810E," which is part of its AI supercomputer initiative, showcasing its advancements in AI technology and chip development [1][4][6]. Group 1: Stock Market Performance - On January 29, Hong Kong's three major indices opened lower, with the Hang Seng Index briefly turning positive before stabilizing [1]. - Major tech stocks experienced declines, with JD.com down nearly 2%, Alibaba down approximately 0.7%, Baidu down 0.45%, and Tencent down 0.16% [1]. Group 2: Alibaba's AI Developments - Alibaba is developing the "Tongyun Ge" AI supercomputer, integrating its self-developed chip, Alibaba Cloud, and the open-source model "Qianwen" to achieve high efficiency in training and deploying large models [4]. - The "Zhenwu" PPU has been deployed in multiple clusters on Alibaba Cloud, serving over 400 clients, including major organizations like State Grid and Xpeng Motors [4][6]. - The "Zhenwu" PPU features a 96GB HBM2e memory and a 700 GB/s inter-chip bandwidth, making it suitable for AI training, inference, and autonomous driving applications [6]. Group 3: Competitive Positioning - The performance of the "Zhenwu" PPU reportedly exceeds that of NVIDIA's A800 and is comparable to NVIDIA's H20, indicating a strong competitive position in the market [6]. - The launch of the "Zhenwu" PPU reflects years of strategic investment and vertical integration by Alibaba, marking a significant milestone in its AI and chip development journey [6]. Group 4: AI Model Achievements - On January 26, Tongyi Laboratory released the Qianwen flagship inference model Qwen3-Max-Thinking, achieving multiple global records and performance comparable to GPT-5.2 and Gemini 3 Pro [7]. - The Qianwen open-source model has surpassed 200,000 derivative models and over 1 billion downloads, maintaining its position as the largest AI open-source community globally [7].

据介绍，"真武"PPU采用自研并行计算架构和片间互联技术，配合全栈自研软件栈。其内存为96G HBM2e，片间互联带宽达到700 GB/s，可应用于AI训练、AI推理和自动驾驶。 阿里巴巴已将"真武"PPU大规模用于千问大模型的训练和推理，并结合阿里云完整的AI软件栈进行深度 优化，为客户提供一体化产品和服务。 智通财经APP获悉，阿里巴巴(09988)平头哥官网上线高端AI芯片"真武810E"产品，该芯片实现软硬件 全自研，已在阿里云实现多个万卡集群部署。每颗真武810E芯片配备7个ICN片间互联端口，配合平头 哥自研互联加速库，实现多卡协同工作，从而高效支持大模型训练及推理需求。 公开资料显示，平头哥半导体有限公司成立于2018年9月，作为阿里巴巴集团的全资半导体芯片业务主 体，其拥有端云一体全栈产品系列(涵盖数据中心芯片、IoT芯片等)，实现芯片端到端设计链路全覆 盖。平头哥已推出多款数据中心及物联网芯片，包括AI芯片"PPU"、推理芯片"含光800"、CPU"倚天 710"、SSD主控芯片"镇岳510"及IoT芯片"羽阵"。 此前媒体传出阿里巴巴(09988)正准备推进旗下AI芯片制造企业"平头哥"( ...

此前媒体传出阿里巴巴(09988)正准备推进旗下AI芯片制造企业"平头哥"(T-Head)独立上市。阿里计划将 其重组为一个部分由员工持股的业务实体，随后再寻求进行首次公开募股(IPO)。 公开资料显示，平头哥半导体有限公司成立于2018年9月，作为阿里巴巴集团的全资半导体芯片业务主 体，其拥有端云一体全栈产品系列(涵盖数据中心芯片、IoT芯片等)，实现芯片端到端设计链路全覆 盖。平头哥已推出多款数据中心及物联网芯片，包括AI芯片"PPU"、推理芯片"含光800"、CPU"倚天 710"、SSD主控芯片"镇岳510"及IoT芯片"羽阵"。 据介绍，"真武"PPU采用自研并行计算架构和片间互联技术，配合全栈自研软件栈。其内存为96G HBM2e，片间互联带宽达到700GB/s，可应用于AI训练、AI推理和自动驾驶。 阿里巴巴已将"真武"PPU大规模用于千问大模型的训练和推理，并结合阿里云完整的AI软件栈进行深度 优化，为客户提供一体化产品和服务。 阿里巴巴(09988)平头哥官网上线高端AI芯片"真武810E"产品，该芯片实现软硬件全自研，已在阿里云 实现多个万卡集群部署。每颗真武810E芯片配备7个ICN片间 ...

智通财经记者 | 程璐 智通财经编辑 | 文姝琪 就在阿里旗下芯片设计业务单位平头哥被爆出考虑启动IPO后的一周里，1月29日上午，平头哥官网悄 然上线一款名为"真武810E"的高端AI芯片，此前被央视《新闻联播》曝光的阿里自研芯片PPU正式对 外。 此前智通财经曾报道，平头哥IPO信号释放的背后，关键推力在于其产品在外部市场取得的进展。其 中，新一代自研AI加速器芯片PPU"真武"扮演了关键角色。 智通财经了解到，"真武"PPU已在阿里云实现多个万卡集群部署，并服务了国家电网、中科院、小鹏汽 车、新浪微博等400多家客户。 不过，这也意味着在商业模式上，平头哥多通过AI云服务来提供算力，其营收与阿里云深度捆绑，外 界难以清晰辨识其独立的盈利能力。未来要如随着上市规划的启动，平头哥也要加速走出去，拓展外部 销售市场，同时与母公司阿里巴巴确立更清晰的关联交易定价机制，让外界得以评估平头哥独立盈利能 力。 "真武"PPU的正式上线，是平头哥在芯片领域积累多年的能力集中输出。阿里巴巴2009年创建阿里云， 2018年成立平头哥芯片公司，2019年启动大模型研究，经过长达17年的战略投入和垂直整合，实现 了"通云哥" ...

｜GUIDE｜ ■BAT为何造芯 ■国产芯片在AI时代前景几何 然而，站上IPO风口之后，这些芯片公司正面临一个更现实的命题，即如何证明自己具备脱离母体后的生存能力？以昆仑芯为例，其正在摆脱对百度内部 需求的单一依赖，尝试在非百度场景中运行主流 AI 框架，并逐步获得运营商、政企客户的认可。 对于大厂芯片而言，从支撑内部业务的算力工具转型为极具竞争力的商业产品，将是决定其上市后估值与长期增长空间的关键变量。 BAT造芯 ■不可忽略的亏损 2026年伊始，大厂造芯业务的独立进程明显加速。 近期， 阿里巴巴也传出消息称旗下芯片公司平头哥，正筹备独立上市； 百度也分拆昆仑芯并以保密形式向港交所提交主板上市申请；腾讯重仓的燧原科 技也计划登陆科创板，其 IPO 申请已于近日获受理。 短时间内，多家大厂系芯片公司同时迈向资本市场，意味着互联网造芯正走向独立商业化阶段。 互联网大厂的造芯浪潮始于2018年。 当年在地缘政治冲击与AI算力需求爆发的双重推动下，百度、阿里、腾讯等公司陆续下场造芯，但很长时间内，大厂芯片业务主要服务于自家搜索、推 荐、云计算等核心业务。 "当初我们做搜索的时候，买别人的芯片太贵了，1 万美元一 ...