中兴通讯 20251010 摘要 超节点已成国产算力标配，英伟达通过 NVLink 闭环和华为引领的临朐 总线是主要技术路线。英伟达的 GB200、NVL72 等产品和华为的 Cloud Metric 384 及 Atlas 900 SuperPod 是代表性方案，分别采用 全互联架构和零区总线技术。 系统能力而非单卡能力是 AI 算力发展的关键。随着大模型训练对算力需 求攀升，跨服务器、跨机柜高速互联成为行业热点。高带宽、低时延、 可靠性及能效目标依赖于系统级协同，具备强大系统能力的公司将占据 优势。 通信系统厂商如华为、中兴通讯等在算力领域具有互联优势。华为能构 建大规模集群得益于其在通信系统的积累。英伟达通过收购 Mellanox 和 Cumulus Networks，强化 Networking 能力，并通过 NVLink 和 CUDA 生态延伸计算能力。 中兴通讯推出智算超节点系统并获得年度重大突破成果奖，该系统基于 GPU 卡间高速互联、开放架构及自研凌云 AI 交换芯片。Nebula 64 超 节点系统支持多种主流互联协议，实现大规模互联。 Q&A 什么是超节点时代？为什么说它已经到来？ 超节点时 ...

Group 1 - The report from Dongfang Securities highlights that process technology upgrades drive chip performance improvements, while advanced packaging serves as another key driver for enhancing chip capabilities [1] - In the context of slowing process technology upgrades, increasing the die area can enhance transistor count and computational power, with Nvidia's H100 die area nearing the reticle limit of approximately 800-900 mm² [1] - Nvidia's B200 adopts advanced packaging with dual die integration, achieving 208 billion transistors in a single package, which is more than double the 80 billion transistors in the H100 [1] - According to Nvidia's roadmap, the Rubin Ultra will integrate four dies in a single package, targeting a computational power of 100PF FP4 per card [1] Group 2 - System assembly is emerging as a new driver for AI server performance enhancement, as wafer manufacturing and advanced packaging may not keep pace with the growing demand for AI computing power [2] - The number of GPUs in AI servers is expected to increase from 8 per server to 72 per cabinet, with projections for the VR Ultra NVL576 cabinet in 2027 to support 144 GPUs, each with four die, totaling 576 die [2] - The increase in GPU count raises cooling requirements and complicates system assembly, exemplified by the production ramp-up challenges faced by GB200 NVL72 due to assembly difficulties [2] - Leading companies in the industry are likely to benefit from the rising technical barriers and improved competitive environment in system assembly [2] Group 3 - In terms of investment targets, companies related to AI server system assembly are maintained, including Industrial Fulian, which has significantly optimized GB200 series product testing and reduced cabinet debugging time [3] - Industrial Fulian has expanded capacity globally and introduced fully automated assembly lines, expecting strong growth in GB200 shipments, primarily driven by large North American cloud service providers [3] - Haiguang Information's merger with Zhongke Shuguang is anticipated to create vertical integration capabilities encompassing CPU, DCU, and system assembly [4] - Lenovo is expected to launch various servers based on Blackwell Ultra starting in the second half of 2025, as indicated by Nvidia [4] - Huaqin Technology, a core ODM supplier for domestic internet firms' AI servers, benefits from the capital expenditure expansion of downstream cloud companies [4]

Group 1 - The report from Dongfang Securities highlights that process technology upgrades drive chip performance improvements, while advanced packaging serves as another key driver for performance enhancement [1] - In the context of slowing process technology upgrades, increasing the die area can enhance transistor count and computing power, with Nvidia's H100 die area nearing the reticle limit of approximately 800-900 mm² [1] - Nvidia's B200 adopts advanced packaging with dual die integration, achieving 208 billion transistors in a single package, which is more than double the 80 billion transistors in the H100 [1] Group 2 - System assembly is emerging as a new driver for AI server performance enhancement, as wafer manufacturing and advanced packaging may not keep pace with the growing demand for AI computing power [2] - The number of GPUs in AI servers is expected to increase significantly, from 8 GPUs per server to 72 GPUs per cabinet, and potentially to 144 GPUs in the VR Ultra NVL576 cabinet by 2027 [2] - The increase in GPU count raises challenges in heat dissipation and system assembly complexity, with industry leaders likely to benefit from the elevated barriers to entry and improved competitive environment [2] Group 3 - In terms of investment targets, companies related to AI server system assembly are maintained, including Industrial Fulian, which has optimized GB200 series product testing and reduced system-level cabinet debugging time [3] - Industrial Fulian is expanding capacity globally and has introduced fully automated assembly lines, expecting strong growth in GB200 shipments, primarily driven by large North American cloud service providers [3] - Haiguang Information's merger with Zhongke Shuguang is anticipated to create vertical integration capabilities encompassing CPU, DCU, and system assembly [4] Group 4 - Lenovo is expected to launch various servers based on Blackwell Ultra starting in the second half of 2025, as indicated by Nvidia [4] - Huaqin Technology, a core ODM supplier for domestic internet companies' AI servers, is set to benefit from the capital expenditure expansion of downstream cloud companies [4]

Summary of Key Points from the Conference Call Industry Overview - The conference call focuses on the **Datacenter Market** within the **Hardware Technology** sector, specifically regarding **ODM (Original Design Manufacturer) Direct** server shipments in the Asia Pacific region. Core Insights 1. **ODM Shipments Growth**: In 2Q25, ODM shipments increased by **3% quarter-over-quarter (q/q)** and **10% q/q** in value, driven by a growing mix of AI servers [1][2] 2. **Server Shipment Statistics**: Global ODM direct server shipments reached **1.91 million units** in 2Q25, reflecting a **3% q/q** and **46% year-over-year (y/y)** growth. This accounted for **45.4%** of total global server shipments [2] 3. **Market Share Dynamics**: - **Wiwynn** led the ODMs with a **23.7%** market share, followed closely by **Quanta** at **23.2%** and **Inventec** at **19.8%**. - **AMD** regained market share to **44.2%**, while **Intel** saw a decline to **38.1%** [2][10] 4. **Regional Performance**: Central & Eastern Europe exhibited the highest growth in shipment units at **105% y/y**, with the US following at **64% y/y**. The US remains the largest market, accounting for **76%** of aggregate ODM direct shipments [3][10] Financial Metrics 1. **Aggregate ODM Direct Server Shipment Value**: The total shipment value was **US$65.4 billion**, marking a **10% q/q** and **201% y/y** increase, primarily due to AI servers [2] 2. **Price Trends**: The average selling price (ASP) of ODM servers increased by **7% q/q** in 2Q25, attributed to improved yields and increased GB200 shipments [4] Company Preferences - The report expresses a preference for specific ODM companies, highlighting **Hon Hai/FII**, **Wistron**, and **Quanta** as favorable investment opportunities [1][4] Additional Insights 1. **AI Server Demand**: The demand for AI servers continues to be a significant driver of growth in the ODM sector, contributing to the overall increase in shipment value [2][4] 2. **Market Challenges**: The report notes uncertainties around tariff implementations that could impact future growth, yet strong demand from cloud services has supported unit share gains for key players [4] Conclusion - The conference call provides a comprehensive overview of the ODM direct server market, highlighting growth trends, competitive dynamics, and financial metrics that indicate a robust demand for AI servers and a favorable outlook for selected ODM companies in the Asia Pacific region.

Group 1: AI Hardware and Chip Demand - The report from Tianfeng Securities highlights optimism regarding AI hardware driving chip demand and opportunities for domestic substitution [1][3] - The semiconductor, domestic computing power, and self-controllable sectors are expected to be long-term trends [1][3] - In the context of ongoing uncertainties in US-China trade policies regarding AI computing power chips, domestic large model development companies and internet platforms are likely to gradually increase their procurement and usage of domestic chips [1][3] Group 2: Consumer Electronics and Market Trends - Apple is set to release new features and products, including the iPhone 17 series, AirPods, and Apple Watch, indicating a positive outlook for the Apple supply chain's innovation cycle and valuation recovery [2] - Meta has launched four AI glasses priced between $379 and $799, covering both everyday use and professional sports, creating a comprehensive product matrix [2] - The GPU and AISC markets are expected to expand, with the launch of GB300 servers and the ramp-up of GB200 driving a new growth phase for AI infrastructure [2] Group 3: Semiconductor Market Dynamics - AI SoC chip companies are benefiting from the increasing penetration of AI hardware, with significant growth reflected in the first half of 2025 [3] - Rising storage prices and the demand from AI servers, PCs, and mobile devices are driving rapid upgrades in storage capacity, with high-value products like HBM, eSSD, and RDIMM seeing increased penetration [3] - The domestic storage market is expected to accelerate growth further [3]

就在几天前，英伟达与OpenAI宣布了一笔价值千亿美元的世纪交易，以构建下一代AI的超级算力集群。 而今天，战场的另一端便给出了同样量级的回应——阿里巴巴在云栖大会上宣布，将为其AI基础设施投入3800亿人民币，并持续追加。 至此，继OpenAI、谷歌、Meta以及我们此前分析过的xAI之后，阿里正式成为加入这场全球"算力大战"的最新一位巨头。 从硅谷到杭州，全球最顶尖的科技公司们，不约而同地将筹码压在了同一个方向上。 这一系列动辄千亿美金级别的豪赌，传递出一个无比清晰的信号： 在通往通用人工智能（AGI）乃至超级智能（ASI）的道路上，先进的算法模型是入场券，而庞大、高效的算力基础设施，则构成了这场竞赛最核心的战场与赢得未 来的战略基石。 那么，当所有巨头都意识到必须建立自己的"算力壁垒"时，一个更深层次的战略问题浮出水面： 构建一个成功的算力帝国，究竟需要什么？ 这是一条无法仅靠金钱铺就的道路，它考验着一家公司对未来的预判能力、极致的工程执行力、对系统架构的颠覆式思考，以及构建开发者生态的战略定力。 最近，英伟达宣布向英特尔投资50亿美元，双方将联手开发定制数据中心和PC产品。这一合作在业界引起了巨大反响 ...

谷歌链未来值得高度重视。谷歌不仅可以对外开放其 TPU，同时微软、 OpenAI、Meta 甚至苹果都有意愿采购谷歌 TPU。谷歌在大模型方面实力强劲， Q&A AI 在当前市场中的发展趋势如何？尤其是三季度的业绩预期？ AI 目前进入了一个再进攻的状态。此前经历了一段时间的震荡，现在正值三季 报窗口期，预计会像二季度一样出现板块性全面超预期的情况。例如，工业富 联在业绩预告中表现领先，随后互电等公司创新一胜天符。在三季度，GB200 和低价产品拉货量非常大，包括谷歌 TPU 的拉货量也很可观。因此，对于一线 龙头标如易中天、PCB 领域中的沪电、胜宏、深南等公司的同比业绩非常值得 期待。此外，Oracle、英伟达、CPX 以及英伟达投资 OpenAI 等事件密集发生， 对市场产生了很大的催化作用。龙头标如旭创，其合理估值在 20 倍以上，如 果牛市持续，其估值可能达到 30 倍甚至更高。 谷歌链和其他科技巨头在 AI 领域的发展前景如何？ Al 之光一新架构.新技术.新时代：光模块更新 20250923 摘要 三季度 GB200 和低价产品拉货量巨大，谷歌 TPU 需求可观，预计易中 天、沪电、胜宏、深南等 ...

Core Insights - Nvidia's $5 billion investment in Intel marks a significant partnership in the semiconductor industry, described as a "surprising alliance" that could reshape the competitive landscape [1][2] - Nvidia's CEO Jensen Huang is noted for his bold, future-oriented investment strategies, which are seen as crucial for maintaining leadership in the AI era [1][2] - The collaboration aims to develop customized data center and PC products, reflecting a shift in market dynamics and potential benefits for both companies [1][12] Nvidia's Strategic Characteristics - Nvidia's core strategic features include intuitive, aggressive decision-making, rapid execution capabilities, and ecosystem investments to diversify its customer base [2][4] - Huang's leadership style is characterized by a willingness to take significant risks, often betting the company's resources on unproven technologies or markets [6][64] - The company has successfully maintained a high market share by delivering products on the first attempt, avoiding the common iterative issues faced by competitors [86] Market Dynamics and Competition - The investment is expected to enhance Nvidia's position against major cloud providers like Microsoft, Amazon, and Oracle, which are also ramping up their AI infrastructure spending [3][14] - Nvidia's pricing strategy, which offers equal pricing to all customers, aims to foster relationships with both large clients and emerging cloud service providers [3][14] - The anticipated capital expenditures for super cloud companies are projected to reach $450-500 billion by 2026, significantly higher than Wall Street's expectations, with a large portion directed towards Nvidia [3][21] Implications for Competitors - The partnership between Nvidia and Intel poses a significant threat to AMD and ARM, as it consolidates Nvidia's influence in the CPU market and disrupts existing competitive dynamics [17][18] - AMD's struggles in the market are exacerbated by this alliance, as it faces increased competition from a united Nvidia and Intel front [17][18] Future Outlook - Nvidia's future strategy may involve further investments in AI infrastructure and data centers, leveraging its strong cash flow to capitalize on market opportunities [88][89] - The company is positioned to benefit from the anticipated explosive growth in AI-related spending, with projections suggesting annual expenditures could reach trillions of dollars [30][31] - Nvidia's ability to adapt and innovate in response to market demands will be critical in maintaining its competitive edge in the evolving tech landscape [88][89]

编者按： 一阵消费风向的转变，能让昔日的"硬通货"跌落神坛。曾经的"公务宴请标配"、价格稳居高位的飞天茅台，批发价也跌到历史新低。 在经济趋势研究专家时寒冰看来，稀缺性并不是永恒的，而是受时间、供需与替代品影响。投资意义上的稀缺，指在某一阶段需求持续而供给跟不上、且 缺少可替代品。它可以来自地域，比如贵州茅台的独特产地；来自技术，如微软操作系统、英伟达的AI芯片优势；也可以是不可再生资源，如铟和可 可。 所以说，无论是做投资还是做实业，在消费降级的巨大威力面前，都必须顺势而为，及时进行调整，调整越早、越快，就越主动，反应越慢、越迟钝，就 越被动。 *文章摘自《全球视野下的投资机会》一书，作者为时寒冰。 如何定义稀缺性？ 物以稀为贵——以稀缺性为目标，是获取高盈利的捷径——但必须考虑时间因素，时间是成就、强化稀缺性的帮手，同时也是摧毁、击败稀缺性的利器。 据智通财经的报道，酒类行情监测平台"今日酒价"数据显示，2025年53度500ml飞天茅台散瓶批发参考价报1770元，再创上市以来的新低；与 平台上线时的2220元、2255元相比，两者累计跌幅均已超过20%。图源：新华社 经济学上给稀缺性下的定义是这样的：稀 ...

Group 1 - Mizuho Securities raised Micron Technology's target price from $155 to $182 while maintaining an "outperform" rating, ahead of the company's Q4 earnings report on September 23 [1] - The upgrade is primarily based on an improved valuation of Micron's High Bandwidth Memory (HBM) and optimistic expectations for the DRAM and NAND flash markets [1] - The pricing environment for NAND and DRAM is expected to improve due to recent capacity cuts by major suppliers and increased supply constraints driven by AI demand, potentially leading to a 20-30% increase in DRAM prices at current levels [1] Group 2 - Micron is expected to remain one of the only two qualified suppliers for HBM4, as Samsung Electronics currently lacks the necessary qualifications [2] - SK Hynix and Micron have begun delivering initial samples of HBM4 for NVIDIA's Rubin, with mass production expected to commence between late Q2 and early Q3 of 2026 [2] - The demand for nearline SSDs is benefiting from HDD delivery times extending to over a year, further supporting the overall positive trend in the memory market [2]