根据沙利文研究数据，以销售收入计，全球PCB市场规模从2020年的620亿美元增长至2024年的750亿美 元，2020年至2024年期间的复合年增长率为4.9%。 AI服务器单台PCB价值量远高于传统服务器，高性能PCB的需求将大幅增长。 从材料角度考虑，低介电常数及低介质损耗因数的材料最合适，Q布性能远优于二代布，或颠覆原来玻 纤布原材料路线。 机构强烈推荐M9上下游及相关PCB厂商。 建滔积层板(01888)：建滔积层板今年上半年实现营业额95.88亿港元，同比增加11%；纯利9.33亿港元， 同比增加28%。此外，近期覆铜板涨价潮起，建滔、威利邦、宏瑞兴8月15日齐涨5—10元/张。开源证 券指出，2025H2公司覆铜板价格已率先调涨，PCB需求强劲或支撑价格刚性，下半年业绩表现有望受 益；2026年高端CCL及物料产能布局提速，AI期权有望步入兑现期，驱动估值中枢上移，实际报表贡献 或在2027年集中释放。 据相关产业消息，英伟达确定在新一代产品Rubin使用M9 材料，在明年下半年发售的Rubin系列中， CPX和midplace的PCB都将使用M9 CCL， 由于Q布（石英布/三代布）极度紧缺， ...

智通财经APP获悉，据相关产业消息，英伟达确定在新一代产品Rubin使用M9 材料，在明年下半年发售 的Rubin系列中，CPX和midplace的PCB都将使用M9 CCL， 由于Q布（石英布/三代布）极度紧缺，英伟 达正在评估compute和swithc tray是否也采用M9，其中swithc tray评估将在11月底确定。 27年的Rubin Ultra#确定使用正交背板代替铜缆，M9材料，3块26层板合成78层板。 仅CPX，midplane和正交背板市场空间近千亿。 中金发布研报称，Rubin CPX是英伟达(NVDA.US)发布的一款专为处理超长上下文AI推理任务设计的 GPU，其采用了创新的解耦式推理架构。Rubin CPX在硬件端带来了较大变化，其中，连接器/PCB方 面，采用无线缆架构，同时新增Paladin B2B连接器与位于机箱中间的PCB中板(mid plane)相连。 Rubin CPX在PCB、连接器、散热等架构上带来显著创新，有望大幅增加其硬件端市场规模，中金公司 预计2027年英伟达AI PCB市场规模有望达69.6亿美元，较2026年增长142%。 华金证券发布研报称，人工 ...

Summary of NVIDIA VR200 NVL144 CPX Rack Design Changes and Beneficiaries Company and Industry - **Company**: NVIDIA (NVDA US) - **Industry**: Semiconductor and PCB (Printed Circuit Board) Design Key Points and Arguments Introduction of VR200 NVL144 CPX Rack - NVIDIA has launched the VR200 NVL144 CPX rack, generating significant interest among investors [1] - The new rack features PCB design changes aimed at optimizing performance and cost [1] Processing Stages of Large Language Models - The processing of large language models is divided into prefill and decode stages, each with distinct load characteristics [2] - The prefill stage is compute-bound, while the decode stage is memory-bound, leading to performance degradation when run concurrently [2] Design Changes in the VR200 NVL144 CPX Rack - The rack maintains the traditional design of 18 compute trays and 9 switch trays but incorporates changes to the Oberon architecture [5] - Each compute tray includes 2 Vera CPUs, 4 Rubin GPUs, 8 Rubin CPX GPUs, and 8 CX9 NICs [5] - The Rubin CPX uses GDDR7 memory instead of HBM, focusing on optimizing prefill scenarios [2][7] PCB Value Increase - The PCB value in the VR200 NVL144 CPX rack has nearly tripled compared to the GB200/300 generation, with an estimated total PCB value of over $3000 per compute tray [19] - The Bianca board in the VR200 has seen a 30% increase in value compared to previous generations [18] Cooling Solutions - The power of Rubin GPGPU has increased from 1800W to 2300–2500W, necessitating the adoption of Micro-Channel Lid (MCL) for improved cooling efficiency [23][24] - MCL is supplied exclusively by Taiwan's Jentech, making it a key beneficiary of the power increase [25] Memory Module Changes - NVIDIA is switching from LPDDR5X memory to SOCAMM, which offers higher flexibility and maintainability [26][29] - This change benefits suppliers of memory module peripheral chips, such as Rambus, as SOCAMM modules require additional components [29] Supply and Demand for HVLP4 Copper Foil - Demand for HVLP4 copper foil is expected to exceed supply, with NVIDIA's VR200 racks and AWS Tranium 3 racks fully adopting HVLP4 [31][33] - By 2027, total demand for HVLP4 is projected to be 15,000 tons, while supply is estimated at 13,000 tons, resulting in a 10% supply shortage [33] MEC as a Hidden Beneficiary - MEC, a Japanese company specializing in surface treatment chemicals, is positioned to benefit from the increased demand for HVLP4 copper foil due to its CZ series products [35][36] - MEC's revenue from the adoption of its products in AI server PCBs could increase significantly, contributing to 50% of its total revenue by 2027 [48] Other Important Points - The PCB design changes include a midplane to replace overpass cables, enhancing connectivity within the compute tray [12][15] - The NVSwitch tray boards have also seen a value increase, with a new design that includes more layers and higher-grade materials [20] - The transition to SOCAMM memory modules is NVIDIA's second attempt, raising concerns about potential warpage issues [29] This summary encapsulates the critical developments and implications of NVIDIA's new VR200 NVL144 CPX rack, highlighting the potential beneficiaries and market dynamics within the semiconductor and PCB industries.

Summary of Key Points from the Conference Call Industry Overview - The focus is on the semiconductor industry, specifically the CoWoS (Chip on Wafer on Substrate) technology utilized by companies like Nvidia and Broadcom [2][3][4]. Core Insights and Arguments 1. **Demand Forecasts for CoWoS**: - Nvidia's CoWoS demand estimates have been raised by 5% for 2025 and 26% for 2026, driven by higher production units of the Blackwell architecture, which is expected to increase by 30% quarter-over-quarter in Q3 2025 [2]. - Total GPU production units for Nvidia at TSMC are projected to be 6.9 million in 2025 and 7.4 million in 2026, up from previous estimates of 6.5 million and 6.7 million respectively [2]. 2. **Rubin Production at TSMC**: - Nvidia's Rubin production at TSMC is on track, with an increase in production units estimated from 1.3 million to 2.3 million in 2026 [3]. - Trial production of Rubin chips is expected to conclude soon, with sample shipments to supply chain partners anticipated this quarter [3]. 3. **New Product Launch - Rubin CPX**: - The introduction of the Rubin CPX SKU, which features a single GPU die and utilizes GDDR7, is expected to drive additional CoWoS demand [4]. - CoWoS demand for Nvidia is forecasted to rise from 444,000 units in 2025 to 678,000 units in 2026 due to the CPX and Rubin ramp-up [4]. 4. **Stock Recommendations**: - TSMC is recommended as a leading foundry for Cloud/Edge AI, benefiting from its advanced packaging capabilities [5]. - ASE is also highlighted as a key beneficiary in the advanced packaging and testing sector [5]. Additional Important Insights 1. **Valuation Comparisons**: - A detailed valuation comparison table shows TSMC with a market cap of $1,225,437 million and a target price of NT$1,570, while ASE has a market cap of $25,042 million with a target price of NT$189 [6]. 2. **Revenue Projections**: - CoWoS/2.5D packaging revenue is projected to grow significantly, reaching $13,423 million by 2026, with Nvidia's revenue expected to hit $7,970 million in the same year [9]. 3. **Market Dynamics**: - The report emphasizes the competitive landscape and the rapid technological changes in the semiconductor industry, which pose both opportunities and risks for companies involved [18][19][20]. 4. **Investment Risks**: - The report outlines various risks associated with tech investing, including volatility in financial results and challenges in valuation due to the dynamic nature of the market [18][19][20]. 5. **Analyst Certifications**: - Analysts involved in the report have certified that their views reflect their personal opinions and were prepared independently [23]. This summary encapsulates the critical insights and projections regarding the semiconductor industry, particularly focusing on Nvidia and Broadcom's CoWoS technology and the implications for TSMC and ASE.

ab 8 October 2025 Global Research UBS Global I/O Semiconductors Stronger CoWoS demand outlook from Nvidia and Broadcom Raising our CoWoS estimates for Nvidia and Broadcom Following the upward revision of our TSMC CoWoS capacity estimate from 100kwpm to 110kwpm by end-2026 in our TSMC preview, we raise our bottom-up demand forecasts for Nvidia and Broadcom. We raise our Nvidia CoWoS demand estimates 5%/26% for 2025/2026 to factor in: 1) higher Blackwell production units, which we believe could be up 30% QoQ ...

Summary of Key Points from the Conference Call Industry and Company Overview - **Industry**: AI Supply Chain, specifically focusing on semiconductor manufacturing and AI ASICs - **Key Companies**: TSMC (Taiwan Semiconductor Manufacturing Company), NVIDIA, Meta, Google, AWS, Alchip, MediaTek, GUC (Global Unichip Corp) Core Insights and Arguments 1. **CoWoS Capacity and Demand**: - TSMC's 2026 CoWoS capacity plan is currently under-supplying based on anticipated demand from key customers, particularly NVIDIA [2][10] - NVIDIA's RTX Pro 6000 forecast remains strong, indicating robust inference demand in China [1][2] - TSMC's current offering of 590k CoWoS-L wafers in 2026 is projected to be 20% below NVIDIA's demand [2] 2. **NVIDIA's AI ASIC Developments**: - The Rubin CPX GPU is expected to adopt TSMC's CoWoS-S packaging, with a performance of 30 PFLOPS and 128GB of GDDR7 memory [3][39] - NVIDIA's AI capacity costs are estimated at $50-60 billion per GW, with $35-40 billion allocated to NVIDIA, suggesting potential revenues of $350-400 billion starting in 2H26 [2][11] 3. **AI ASIC Volume Forecast**: - AI ASIC volumes could reach approximately 5.7 million units in 2026 and 8 million units in 2027, with Google and AWS being the primary players [47] - MediaTek's chances of winning the Meta MTIA project are diminishing due to strong competition from Broadcom and Marvell [57][58] 4. **Market Dynamics and Competitor Analysis**: - GUC is expected to benefit from increased demand for Google Axion CPUs, potentially contributing around $250 million to its revenue in 2026 [59] - Alchip is projected to generate approximately $1.8 billion in revenue from Trainium3 in 2026, despite a back-loaded shipment schedule [56] 5. **Capacity Expansion and Future Projections**: - TSMC's clean room space is ready for expansion to 110-120k if demand is confirmed, with a revised CoWoS capacity assumption of 100kwpm by 2026 [2][12] - The overall demand for CoWoS is expected to grow significantly, with NVIDIA's demand forecasted to increase by 61% in 2026 [20] Additional Important Insights - **NVIDIA's Strategic Positioning**: - NVIDIA's Rubin architecture is designed to optimize inference performance, potentially delivering a 30x to 50x return on investment [39] - The demand for NVIDIA's Blackwell GPU, including RTX Pro 6000, is strong among Chinese customers, with expected sales of 1.5 million to 2 million units in 2H25 [14] - **Competitive Landscape**: - SMIC is aggressively expanding its 7nm node capacity to meet domestic GPU and AI ASIC demand in China [14] - The competition in the AI ASIC market is intensifying, with various players vying for market share and technological advancements [57][58] This summary encapsulates the critical insights and projections regarding the AI supply chain, particularly focusing on TSMC and NVIDIA's roles within the semiconductor industry.

公众号记得加星标⭐️，第一时间看推送不会错过。 过去几年，云厂商为了训练大模型投入巨资购买芯片，如今也到了利用推理实现变现的时候了。根据麦肯锡报告，全球AI推理市场规模预计2028年将达1500亿美 元，年复合增长率超40%，远高于训练市场的20%。推理支撑着各类应用的实时推理需求，包括智能推荐、内容生成、虚拟助手等。可以说，推理阶段才是实现实 际应用和商业化的关键。 这场推理之战，随着华为、英伟达和谷歌三大巨头相继发布了各自的推理芯片之后，已经将正式打响！ 华为Ascend 950PR： 成本优化下的推理利器 9月18日，在2025年华为全联接大会上，华为宣布了昇腾芯片的规划和进展。未来3年，也就是到2028年，华为在开发和规划了三个系列，分别是Ascend 950系列、 Ascend 960、Ascend 970系列。华为表示，将以几乎一年一代算力翻倍的速度，同时围绕更易用，更多数据格式、更高带宽等方向持续演进，持续满足AI算力不断 增长的需求 以往每年9月，都是手机发烧友的狂欢月，因为这时期苹果、小米、华为等都会发新机。然而，今年的9月，一个更深层次的产业变革正在暗流涌动。当所有人都在 对iphone ...

Core Insights - President Trump announced a plan to make the U.S. a leader in AI and machine learning by removing restrictions on companies developing future technologies [2] - Major chip manufacturers like Nvidia, Intel, and AMD are actively developing new processors to meet increasing AI performance demands, indicating a promising market for AI chips [2] - The AI chip market is expected to grow significantly, but market maturity and consolidation may limit opportunities for new entrants [2] AI Processor Market Growth - Omdia predicts that the AI data center chip market will continue to grow rapidly, with an annual growth rate of over 250% from 2022 to 2024, but slowing to about 67% from 2024 to 2025 [4] - AI infrastructure spending is expected to peak in 2026, driven primarily by AI, before gradually decreasing by 2030 [4] - Precedence Research forecasts the AI chip market will grow from $94.31 billion in 2025 to $931.26 billion by 2034, with a compound annual growth rate of 28% [5] GPU and ASIC Dominance - GPUs remain dominant in the AI chip market due to their parallel processing capabilities, essential for training and inference tasks in data centers [5] - ASICs are expected to drive future growth in the AI chip market due to their efficiency in specific AI functions, particularly in inference-heavy environments [6] Major Developments by Chip Manufacturers - AMD launched the Instinct MI350 series GPUs, offering significant performance improvements and cost-effectiveness for AI solutions [9] - Intel introduced new Xeon 6 series CPUs designed to enhance the performance of GPU-driven AI systems [10] - Nvidia released the Rubin CPX GPU, designed for high-speed processing of large amounts of data, and integrated it into the new Vera Rubin NVL144 CPX platform [11] Cloud and Edge Computing Trends - By 2024, cloud AI processing is expected to dominate the market with a 52% share, driven by investments from major cloud providers [13] - Edge AI processing is rapidly growing due to the demand for low-latency, device-side intelligence in applications like autonomous vehicles and smart city infrastructure [12] Industry Consolidation and Collaboration - Jon Peddle Research predicts that by 2030, the AI processor market will consolidate to about 25 key players, with IoT and edge computing suppliers likely to survive due to their larger market potential [14] - OpenAI and Nvidia have formed a strategic partnership to deploy NVIDIA systems for training next-generation AI models, with Nvidia planning to invest up to $100 billion [14][15] Technical Challenges in AI Processing - The performance demands of AI processors are creating challenges for existing memory configurations, leading to the adoption of new memory designs to meet data and speed requirements [16] - Liquid cooling solutions are being explored to manage the heat generated by high-performance processors, although they add complexity and cost [17]

Core Insights - NVIDIA Corporation is recognized as one of the best stocks to buy, particularly due to its dominance in the AI and data center markets, with significant revenue growth reported [2] - The company achieved second-quarter fiscal 2026 revenue of $46.7 billion, reflecting a 56% year-over-year increase, primarily driven by demand for its Blackwell Data Center products [2] - A strategic partnership with OpenAI involves an investment of up to $100 billion, aimed at supplying advanced data center chips for next-generation AI systems, enhancing NVIDIA's role in the AI ecosystem [3][4] Financial Performance - NVIDIA's second-quarter fiscal 2026 revenue reached $46.7 billion, marking a 56% increase compared to the previous year [2] - The Blackwell Data Center products experienced a 17% sequential growth, indicating strong market demand [2] Strategic Partnerships - The partnership with OpenAI includes a commitment to invest up to $100 billion, which is expected to significantly expand NVIDIA's long-term market potential [3] - NVIDIA also announced a $5 billion investment in Intel stock to co-develop custom AI infrastructure, targeting a $50 billion market opportunity [3] Product Innovations - The introduction of the Rubin CPX GPU is designed for massive-context inference tasks, enabling advanced capabilities such as million-token coding and generative video [4] - This innovation aligns with NVIDIA's vision of transforming data centers into fully integrated "AI factories" [4]

计算机周观点第 18 期：甲骨文云计算订 单超预期，全球算力投资持续高景气 股票研究 /[Table_Date] 2025.09.29 2025-09-29 [Table_Industry] 计算机 本报告导读： 甲骨文 4550 亿美元订单超预期，全球算力投资高景气；英伟达 Rubin CPX 专攻长上 下文推理和视频生成，定制化芯片成趋势；"人工智能+"赋能智慧电网。持续看好国 内算力链、数据中心各环节以及 AI 应用在各行业落地。 投资要点： 行 业 跟 踪 报 告 证 券 研 究 报 告 请务必阅读正文之后的免责条款部分 股 票 研 究 [table_Authors] [Table_Summary] 投资建议：甲骨文 4550 亿美元订单超预期，微软与 Nebius 达成五年 174 亿算力协议，全球算力投资维持高景气度；英伟达 Rubin CPX 面向 长上下文推理和视频生成，看好专用 GPU、ASIC 与配套的 CPO/PCB/存 储环节，"人工智能+"能源高质量发展的实施意见出台，AI 应用赋能智 慧电网建设。我们维持计算机板块"优于大市"评级，推荐标的：金蝶国 际、金山办公、新大陆、朗新集团、 ...