阿里云栖大会观展反馈 20250928 摘要 国产算力商业闭环于 2024 年 10 月形成并强化，受益于豆包模型接近 GPT-4 能力及华为、寒武纪芯片突破，带动 AI 硬件崛起，国内推理和应 用端需求显著增长，字节跳动 token 消耗量超 Google，第三方 IDC 企 业如世纪互联业绩提升。 阿里云栖大会显示，阿里数据中心装机容量到 2032 年将达 2022 年的 10 倍，每年新增 6-7GW。阿里超节点盘久采用 CPU 与 GPU 分离架构， GPU 可使用阿里平头哥，CPU 选择多样，并公开 A Link 协议，类似华 为 UB 协议，促进计算存储网络互通。 国产 AI 硬件生态迅速发展，中兴发布凌云 SKY UP 交换芯片，字节跳动 或将推出相应芯片，表明国内交换芯片及模型训练与推理需求旺盛。阿 里钉钉平台拥有数千万中小企业客户，其 OA、业务流及供应链系统部 署于阿里云与钉钉之上，为 AGI 商业闭环提供沃土。 阿里、华为、海光公开互联协议，有助于 GPU 和交换芯片企业集成，实 现高效互联互通。阿里 Alink 系统实现小于 150 纳秒的 PTP 低延迟和单 柜 PB 级带宽，采用 ...

沐曦摩尔上市加快，国产算力加速腾飞 20250928 摘要 华为在单卡性能优化和大规模集群核心技术方面取得进展，为国产算力 发展奠定基础。阿里巴巴宣布三年投入 3,800 亿人民币建设 AI 技术， 预示数据中心功耗将大幅提升。 摩尔线程已通过上交所上市审核，预计 2025 年 11 月中下旬至 12 月中 上旬挂牌；沐曦半导体二轮问询函审核意见已公告，预计最快 2025 年 12 月中上旬至年底上市，或延至 2026 年一季度。 沐曦半导体 C500 芯片对标国内 A800 水平，主要应用于政府主导的智 算中心项目，如上海国家训练场。2025 年 C500 均价因大客户采购降 至 4 万-4.5 万元。 截至 2025 年 9 月初，沐曦已获 14.3 亿元订单，C500 系列板卡为主。 正大力推进与互联网企业和运营商的合作，预计年底与客户 B 小批量下 单，已进入中国电信集采短名单。 沐曦主力产品 C500 已量产，预计 2026 年仍是主要收入来源。星云 C600 已完成回片，采用国产供应链，支持 FP8 计算精度和液冷，兼容 CUDA 生态。C700 研发中，对标英伟达 H100。 Q&A 国产算力领 ...

Group 1 - Alibaba announced a "three-year 380 billion yuan AI infrastructure construction plan" and a plan to "increase global data center energy efficiency by 10 times by 2032," reinforcing the long-term logic of domestic computing power investment [1] - During the economic recovery, China aims to maintain reasonable growth in key industries while focusing on structural optimization and efficiency improvement [1] - Plans for the electronic information, automotive, power equipment, light industry, steel, and building materials sectors have been rolled out, with continued attention on the progress of anti-involution [1] Group 2 - Rare earth prices continued to stabilize this week, and food delivery subsidies have decreased, leading to ongoing recommendations for the rare earth and internet sectors [1]

Investment Rating - The report maintains a "Buy" rating for the company [1] Core Insights - The "Star Gate" project in China is emerging, indicating a gradual realization of a closed-loop industry for domestic computing power. The Yangtze River Delta hub in Wuhu has successfully integrated four national-level computing platforms and six data centers, including the Jiangsu Zhixin (Zhongke Shuguang Intelligent Computing Center) [9][10] - The domestic intelligent computing power is steadily increasing, projected to rise from 3% in 2016 to 35% by 2025. Over 250 intelligent computing centers are currently built or under construction in China [10] - The demand for computing power from internet clients is continuously rising, with an accelerated adoption of domestic AI chips. Major companies like Alibaba and Tencent are significantly increasing their investments in AI infrastructure [11][13] Financial Performance and Projections - The company is expected to achieve revenues of 156.63 billion, 218.04 billion, and 285.72 billion yuan for the years 2025, 2026, and 2027 respectively, with net profits of 28.43 billion, 40.98 billion, and 58.02 billion yuan for the same years [14][15] - The earnings per share (EPS) are projected to be 1.94, 2.80, and 3.97 yuan for 2025, 2026, and 2027 respectively, with corresponding price-to-earnings (P/E) ratios of 61.98, 43.01, and 30.37 [14][15] Market Position and Competitive Advantage - The company is building a comprehensive competitive advantage in the domestic computing power sector through national-level project orders, regional penetration, and the implementation of proprietary technologies [14] - The launch of the Shuguang AI supercluster system marks a significant advancement, capable of supporting extensive AI computing needs with high efficiency [13][14]

开源证券近日发布通信行业周报：硬件方面，发布磐久AI Infra2.0128卡超节点AI服务 器、高性能网络HPN8.0、自研飞天+CIPU架构等。巨头持续加大AI资本投入，通过柜内超 节点、柜外scale across等方式构建万卡或10万卡等更大算力规模集群，弥补单芯片算力及显 存不足的差距，持续看好国产AI大模型及应用发展，国内算力产业链或持续受益。 阿里发布7款大模型，不断丰富Qwen模型家族，性能持续提升 阿里正式发布旗舰模型Qwen3-Max，是通义千问家族中最大、最强的基础模型，预训练 数据量达36T，总参数超过万亿，性能超过GPT5、Claude Opus4等，跻身全球前三，包括指 令（Instruct）和推理（Thinking）两大版本，在Coding编程能力和Agent工具调用能力上有 较大提升；发布下一代基础模型架构Qwen3-Next，模型总参数80B仅激活3B，性能即可媲美 千问3旗舰版235B模型，模型训练成本较密集模型Qwen3-32B降低超90%，长文本推理吞吐 量提升10倍以上；多模态模型中，Qwen3-VL开源，拥有极强的视觉智能体和视觉Coding能 力，在32项核心能力 ...

记者丨骆轶琪 编辑丨张伟贤 9月26日，摩尔线程在披露招股书（6月30日）后仅88天，顺利进入上会阶段。公告显示，上海证券交易 所上市审核委员会认定，摩尔线程符合发行条件和上市条件。 与之招股时间接近的沐曦股份，则在9月21日披露第二轮问询函内容。虽然沐曦股份暂时未定下一步动 态，但这些都意味着，国产GPU芯片企业的上市正事实上进入加速阶段。 21世纪经济报道记者综合梳理发现，目前，两家GPU芯片公司均尚未实现盈利，但支撑其登陆资本市场 底气的，是二者在2024年均实现营业收入大幅增长，同时面向大模型时代的智算相关产品正逐渐获得市 场认可。 公司方面指出，上半年随着市场对大模型训练、推理部署、GPU云服务等需求大幅提升，加之公司第四 代GPU芯片"平湖"实现商业化，营业收入大幅增长，同时加强对成本费用的管理，实现了亏损改善。 从发展可持续性来看，两家公司不仅在手订单量远超营收，向以互联网、运营商为代表的AI智算关键 客户拓展也见成效。 此外，这已经不是单点突破，国内AI芯片产业链公司近些年来正在持续构建生态化能力，两家公司也 是国产AI算力建设的重要贡献者。 收入大增，智算芯片受热捧 2024年是国产AI芯片 ...

记者丨骆轶琪 编辑丨张伟贤 9月26日，摩尔线程在披露招股书（6月30日）后仅88天，顺利进入上会阶段。公告显示，上 海证券交易所上市审核委员会认定， 摩尔线程符合发行条件和上市条件 。 与之招股时间接近的沐曦股份，则在9月21日披露第二轮问询函内容。虽然沐曦股份暂时未定 下一步动态，但这些都意味着，国产GPU芯片企业的上市正事实上进入加速阶段。 21世纪经济报道记者综合梳理发现，目前， 两家GPU芯片公司均尚未实现盈利，但支撑 其登陆资本市场底气的，是二者在2024年均实现营业收入大幅增长 ，同时面向大模型时 代的智算相关产品正逐渐获得市场认可。 | 项目 | 2025 年 1-6 月 | 2024 年度 | 2023 年度 | 2022 年度 | | --- | --- | --- | --- | --- | | 营业收入 | 7.02 | 4.38 | 1.24 | 0.46 | | 浄利润 | -2.71 | -14.92 | -16.73 | -18.40 | 从发展可持续性来看， 两家公司不仅在手订单量远超营收，向以互联网、运营商为代表的AI 智算关键客户拓展也见成效 。 此外，这已经不是单点突破 ...

Group 1 - The core viewpoint of the article highlights the significant opportunities for the domestic computing power industry chain due to favorable policies and technological advancements, particularly with the successful listing of Moore Threads, marking a milestone for domestic GPU development [1][2] - The Chinese government has released multiple policies to accelerate the construction of computing power infrastructure, emphasizing the urgency and commitment to developing domestic computing capabilities [2][3] - The "East Data West Computing" initiative aims to strengthen the supply of computing resources and build a national integrated computing network, which will drive demand for hardware across the entire computing infrastructure industry chain [3][4] Group 2 - The investment logic in the domestic computing power industry chain is evolving, with a shift from single-function chips to full-function GPUs, expanding application scenarios and market space [3][4] - The integration of AI with various industries, such as transportation, is expected to create diverse and customized demands for computing resources, providing opportunities for companies with industry-specific solutions [4][5] - Key companies benefiting from the recent developments include Cambrian (寒武纪), Haiguang Information (海光信息), and Zhongke Shuguang (中科曙光), which are positioned to capitalize on the growing demand for AI chips and computing infrastructure [7][8][9]

Summary of Semiconductor Testing Equipment Conference Call Industry Overview - The conference call focuses on the semiconductor testing equipment industry, particularly in relation to AI chip development and advanced packaging technologies [1][3]. Key Points and Arguments 1. **AI Chip Development Impact**: The rapid development of AI chips is driving demand for advanced packaging technologies such as HBM (High Bandwidth Memory) and CoWoS (Chip on Wafer on Substrate), creating opportunities for domestic equipment manufacturers [1][3]. 2. **Increased Complexity in Testing**: The complexity of SoC (System on Chip) and storage devices is increasing, leading to a significant rise in demand for testing machines. The market share for SoC testing equipment rose to 60% in 2022, while storage devices dropped to 20% [1][13]. 3. **Advanced Packaging Equipment Demand**: The shift towards HBM and CoWoS as mainstream packaging solutions is expected to increase the demand for advanced packaging equipment [1][8]. 4. **Market Share Dynamics**: In 2022, the combined market share of SoC and storage testing equipment reached 80%, highlighting the close relationship between AI chip development and testing equipment demand [1][13]. 5. **Testing Process Complexity**: HBM testing involves complex processes, including wafer-level testing of DRAM and logic chips, which increases requirements for equipment precision and current-carrying capacity [1][19]. 6. **Key Equipment in Advanced Packaging**: Die Bond machines, bonding machines, and dicing machines are critical in the advanced packaging process, accounting for 30%, 23%, and 30% of the workflow respectively [1][25]. Additional Important Insights 1. **Domestic Companies to Watch**: Companies such as Huafeng Measurement & Control, Changchuan Technology, and Maiwei Co. are highlighted as key players in the context of growing high-process demand [2][32]. 2. **Testing Equipment Pricing**: The price range for storage testing machines is approximately $1 million to $3 million, with a low domestic production rate [10]. 3. **Market Competitors**: In the global market, Advantest holds about 60% of the market share for SoC and storage testing machines, followed by Teradyne with around 30% [23]. 4. **Technological Evolution**: The testing process for semiconductor chips has evolved significantly from 2010 to 2020, necessitating new testing equipment due to increased transistor counts and complexity [17]. 5. **Investment Logic for Domestic Computing Power**: The investment logic focuses on the growth of testing machine demand due to the increasing complexity of SoC and storage devices, as well as the rising need for advanced packaging equipment [8][9]. This summary encapsulates the critical insights from the conference call, emphasizing the impact of AI chip development on the semiconductor testing equipment industry and the opportunities for domestic manufacturers.

在全球经济格局深刻变革、国内宏观政策持续发力扩内需的背景下，供应链的稳定与高效运转已成为提 升产业竞争力的关键。京东供应链金融科技历经12年深耕，以"金融+科技"双轮驱动，精准切入国家补 贴、国产算力、跨境贸易三大新兴场景，通过全链路金融服务破解企业融资难题，为产业链整体升级注 入新动能。 国补政策精准滴灌：双端激活撬动消费与产业双增长 2025年，国家"大规模设备更新和消费品以旧换新"政策持续加码，为企业带来市场机遇的同时，也考验 着政策红利的落地效率。京东供应链金融科技依托"消费金融+供应链金融"协同模式，构建"需求激活供 给保障"闭环，成为政策与市场间的"转换器"。 在供给端，针对国补政策下企业面临的"垫资规模大、结算周期长"等痛点，京东供应链金融科技的"京 保贝""订货贷""货押贷"等产品推出商家专属政策，实现从采购、生产到销售的全链条资金覆盖。以华 为核心分销商国信通达为例，通过"订货贷"，企业在2024年华为Mate70上新时，在预订货阶段就获得融 资，快速锁定货源。2025年国补期间，使用京东供应链金融科技服务后，资金周转效率大幅度提升。 在消费端，京东白条通过"免息+国补"叠加，显著降低用户决 ...