看过我们之 前介绍电力的文章的读者应该都会知道一个现状：美国现 在非常缺电。老美那边也是意识 到了这么个问题，所以卖方的分 析师们也是一直在写这方面的警告。原因我们之前也都讲过，再简 单说一下，一是美国的基础设施确实比中国落后很多，现在已经跟 不上数据中心的建设了；而且美国那边再新建电力效率太低，而现 在电力支出中很大一部分都是在维护老设备，而不是买新设备。 这两天瑞银发了一个分析，认为数据中心会转向SOFC，来解决一部分电力问题。 关于SOFC的文章之前也发过，但讲的不是很深，星球中我们提过很多次Bloom Energy这家公司，就 是SOFC的龙头，从今年7到现在4个月时间，已经翻了快10倍了.. 这篇文章，我们再来深入分析一下SOFC产业链。 AI 算力需求 AI 算力的爆发催生了能源领域的变革，据国际能源署（IEA）发布的《能源与人工智能（AI）报 告》，2024 年数据中心的电力需求约占全球的 1.5%，达 415 太瓦时；预计到 2030 年，数据中心用 电量将增长超一倍，达到约 945 太瓦时，略高于日本当前的用电总量。人工智能是推动这一增长的 主要因素，同时其他数字服务需求也在持续上升。IEA ...

今天要说的比较多，一个个来说。 1、中美谈判 今天大家最关注的，还是领导谈判的结果，我们只提跟芯片相关的，由于川普之前预热了很久，说 是要谈英伟达的Blackwell，国内外都在传B30A要放开的事，结果会后又发了下面内容： Trump: China Will Be Discussing Chips With Nvidia Trump: We're Not Talking About the Blackwell 真的是让人搞不懂是啥意思，Hopper已经停产，不应该再谈Hopper的单子了，有朋友说可能还是 B30A，他指的Blackwell应该是B200/B300，不管怎样，我们也只能等待后面的消息。 2、万亿OpenAI准备上市 根据路透社的消息， OpenAI 正启动IPO筹备工作，初步估值瞄准 1 万亿美元，计划融资规模不低于 600 亿美元。公司已明确将 2027 年作为上市目标年份. 关于OpenAI上市的事，最近其实传了挺久了，1万亿目标，跟老黄说的也是一样的，说明老黄上个月 投资OpenAI时就谈过这个计划了，但投资人可能更想看到一份完整的财务报告。 2025 年OpenAI营收预计约 130 ...

More Than Semi . More Than SEMI 半导体行业研究 芯片制造行业，长期以来似乎被一种惯性所主导，头部企业的技术决策往往受限于我们一直都是这么做 的，对任何可能的技术回归或改变都抱有极大的恐惧。这种惯性渗透在生产的每一个细节里，比如光刻 车间的照明颜色，即便晶圆厂早已明确，传统的黄色灯光对光刻胶不再存在任何影响风险，却依然被保 留至今，没人愿意迈出改变的第一步。这种对既有路径的依赖，在近年来愈发明显：即便现有技术的缩 放速度不断放缓、成本持续飙升，芯片制造商们仍倾向于在原有技术框架内迭代，而非探索全新路径。 以下文章来源于More Than Semi ，作者Nico ASML的一款 roadmap 工具就是典型例子 ，该公司甚至公开承认，这款工具可能并不具备经济可行 性，但行业内的企业却很难轻易放弃既有的技术体系。毕竟，当前的光刻工具和晶圆厂能带来巨额利 润，一台售价 2.25 亿美元的 EUV 工具，一年就能生产出价值超过 6.5 亿美元的完整晶圆（尽管生产这 些晶圆还需承担除 EUV 工具外的诸多成本）。这种高投入高回报的现状，让大多数企业不愿冒险打破 平衡，却也为敢于创新的后来 ...

根据英伟达GTC大会黄仁勋的最新披露，以及结合市场的数据，Hopper系列GPU已出货400万张， 创收1000亿美元；Blackwell系列GPU在前3.5个季度出货600万张。 展望未来5个季度，Blackwell和Rubin系列GPU预计总出货量将达2000万张，总收入约5000亿美元， 较H系列增长5倍。 光模块 扣除已出货的600万颗，未来5个季度预计新增出货1400万颗，平均每季度约280万颗。假设Rubin与 Blackwell出货比例为1:2，即Rubin约370万颗，Blackwell约630万颗，且均采用1.6T光模块。根据花旗 的数据，GB300中，GPU芯片与1.6T光模块配比1:2.5，而到了Rubin中，变成了1:5. 2）3Q业绩低于预期的点是Oracle的Delay，新易盛是Oracle的一供，影响大一些 3）4Q继续扩产，公司觉得市场需求没问题。今天总经理在去泰国路上 PCB 在之前的文章中我们就提到过，目前Rubin的中板、背板、CPX 板这三块板卡，已经基本确定是采用 M9了。 所以B系列中光模块需求约1575万只；Rubin中光模块需求约1850万只，总计1.6T光模块 ...

Core Points - The "2025 Super Node Data Center Industry Summit and High-Density Data Center Developer Forum" will be held from November 26-28 in Hangzhou, Zhejiang, focusing on topics such as super node data center architecture design, interconnection technology, optical module packaging trends, liquid cooling technology, and high thermal conductivity materials [2] - The forum is organized by Cheqian Information and Thermal Design Network, expecting over 40 speakers and 500 industry experts to attend [2] - The event will feature a main conference and two specialized sessions for in-depth discussions [2] Group 1: Main Conference Topics - Core technologies and challenges of super node data centers, along with new product releases, will be discussed [3] - Presentations include topics from major companies like H3C, Intel, Huawei Cloud, and Shuguang Data Infrastructure Innovation Technology [3][4] - Innovations in AI chip interconnection protocols and the evolution of super node optical interconnection will be highlighted [4] Group 2: Specialized Session Topics - Session one will focus on optical interconnection and optical module technology, featuring companies like Zhongji Xuchuang and Fenghuo Communication Technology [5] - Session two will cover thermal management technologies for chip-server-data center integration, with discussions on GPU packaging and liquid cooling technologies [5][6] - Key challenges and strategies for high-speed optical modules in the AI era will also be addressed [6] Group 3: Additional Presentations - Various companies will present on topics such as liquid cooling solutions, energy-saving technologies, and the reconstruction of traditional data centers by super nodes [18][21] - The event will also explore the opportunities and challenges posed by AI computing on optical interconnection [16][18] - The agenda is subject to change, with final topics to be confirmed on-site [19]

Core Viewpoint - The article discusses the recent performance of the A-share market, particularly the Shanghai Composite Index reaching the 4000-point mark for the first time in nearly a decade, and the implications of this milestone for investors [1]. Group 1: Company Performance - Victory Giant Technology (VGT) reported a significant increase in revenue and net profit for the first three quarters of 2023, with revenue reaching CNY 14.117 billion, a year-on-year increase of 83.40%, and net profit of CNY 3.245 billion, up 324.38% [2][3]. - In Q3 alone, VGT achieved revenue of CNY 5.086 billion and net profit of CNY 1.102 billion, reflecting year-on-year growth of 78.95% and 260.52%, respectively [2][3]. Group 2: Industry Trends - The company’s performance of increasing revenue but declining net profit is attributed to several factors, including production line adjustments due to product iterations, increased labor costs from factory expansions, and rising R&D expenses [3]. - Analysts, including those from Nomura, maintain a positive outlook on VGT, expecting profit margins to recover in Q4 and anticipating significant growth in new product iterations in the coming years [3]. Group 3: Product and Market Position - VGT specializes in high-precision multilayer PCBs, HDI PCBs, and flexible printed circuits, with applications in aerospace, automotive electronics, next-generation communication technologies, and medical devices [6][10]. - The company has successfully entered NVIDIA's AI server supply chain and is expanding its market share in the GB200/GB300 product series [6][21]. - VGT's production capabilities include the ability to manufacture over 100-layer high-density PCBs and has been recognized as one of the first companies to achieve mass production of 24-layer HDI+6 products [6][30]. Group 4: Future Growth and Investments - VGT is investing heavily in expanding its production capacity, with projects in Vietnam and Thailand aimed at increasing HDI and HLC production capabilities significantly by 2027 [31][13]. - The company plans to increase its HDI production capacity from 930,000 square meters in 2024 to 2.15 million square meters by 2027, and HLC capacity from 8.65 million square meters to 9.43 million square meters [31].

Core Viewpoint - The article discusses the potential implications of the B30A chip, designed by NVIDIA as a downgraded version of its flagship B300 chip, particularly in the context of U.S. export controls to China and the ongoing AI computing race [5][16]. Group 1: AI Computing Race and Export Controls - The U.S. government faces a complex decision regarding the export of the B30A chip to China, which could significantly enhance China's AI computing capabilities despite being a lower-performance version of the B300 [5][6]. - The Trump administration's AI action plan aims to maintain U.S. leadership in AI by restricting access to advanced AI computing resources, with the U.S. currently leading China in AI supercomputing capabilities by approximately five times [7]. Group 2: Hardware Configuration and Performance - The B30A chip has peak performance and memory bandwidth that are 50% lower than the B300, with a single B30A card priced at approximately $22,500 compared to the B300's $45,000 [8][12]. - A server with eight B30A GPUs consumes only 40% of the power of a B300 server, making it more energy-efficient [8]. Group 3: Cluster Cost Analysis - To achieve the same total computing power as a B300 cluster, a B30A cluster requires double the number of chips, leading to a 24% higher initial investment cost, although this is mitigated by Chinese government subsidies [11]. - The overall amortized cost of a B30A cluster, including server, network, and energy costs over five years, is approximately 1.24 times that of a B300 cluster, indicating a 20% higher cost [13]. Group 4: Strategic Implications of B30A Export - If the B30A is allowed for export, it could significantly narrow the AI computing gap between the U.S. and China, potentially reducing the disparity from over 31 times to below 4 times [14]. - The introduction of B30A could pressure domestic Chinese chip manufacturers, as its performance exceeds that of local alternatives while being more cost-effective [14][15]. Group 5: Global Supply Chain Impact - Allowing the export of B30A could disrupt the global chip supply chain, as NVIDIA's production capacity is limited, potentially leading to longer wait times for other markets [15]. - The B30A's established supply chain and controllable procurement costs make it an attractive option for China, representing a "low investment, high return" scenario [15]. Group 6: Technical and Geopolitical Interplay - The decision to allow B30A exports is complicated by geopolitical considerations, as it could undermine U.S. core advantages in AI while providing NVIDIA with significant revenue [16]. - The AI computing race is not solely a technological competition but also a geopolitical struggle, with the potential for U.S. market restrictions to accelerate China's domestic technology development [16].

Global Insights - Ongoing US-China negotiations are expected to delay certain policies [4] - Arkham Intelligence reported that SpaceX transferred a total of 1,215 bitcoins to multiple addresses [4] - Citigroup highlighted Anthropic's agreement with Google to procure up to 1 million TPUs, indicating potential benefits for TSMC, MediaTek, and KYEC due to increased TPU chip shipments [5] China Insights - Alibaba upgraded Qwen deep research with one-click webpage and podcast generation, achieving accuracy on par with Gemini, outperforming ChatGPT and Grok [7] - Goldman Sachs noted the increasing adoption of Horizon Journey 6 platform in new vehicle models, with mass production of J6P and H5D starting in EXEED models [7] - Xiaoma Zhixing announced the mass production of 300 units of the seventh-generation Robotaxi, aiming for a fleet of 1,000 by the end of 2025 [8] - Citigroup reported that GoerTek's Q3 2025 performance met expectations, with revenue growth of 4% year-on-year to 30.6 billion RMB [8] - Xiaomi is expected to announce Q3 2025 results on November 18, with a projected net profit of 1.2 billion RMB, reflecting a year-on-year increase of 64% [9] - Raycus Laser's Q3 2025 revenue grew by 10% year-on-year to 841 million RMB, with a significant net profit increase of 107% to 50 million RMB [10] - Nomura identified Shenghong Technology as a key beneficiary of Nvidia's innovation cycle, projecting a revenue and earnings CAGR of 40% and 45% respectively from FY2025 to FY2027 [11]

Core Viewpoint - The article discusses the market potential of Solid Oxide Fuel Cells (SOFC) in the context of the evolving electricity market in the United States, highlighting the challenges faced by data centers and the advantages of adopting new technologies like SOFC and Solid State Transformers (SST) [1][2]. Group 1: Background of U.S. Electricity Market Reform - The U.S. electricity market reform began in the 1970s due to the global energy crisis, leading to the establishment of the Department of Energy (DOE) and the Federal Energy Regulatory Commission (FERC) [3]. - The Public Utility Regulatory Policies Act (PURPA) of 1978 marked the start of market reform by encouraging non-utility companies to participate in electricity generation, breaking the monopoly of traditional utility companies [3][4]. Group 2: Key Milestones in Reform - The Energy Policy Act of 1992 allowed non-utility companies to access the grid and sell electricity in the wholesale market, establishing the initial framework for competition [4]. - FERC's orders in 1996 mandated the separation of generation and transmission businesses, introducing Independent System Operators (ISO) to ensure fair competition [4][5]. - The Energy Policy Act of 2005 further strengthened FERC's regulatory role, leading to the establishment of a wholesale market system centered around ISOs and Regional Transmission Organizations (RTOs) [5]. Group 3: Retail Market Opening - The retail side of the U.S. electricity market reform focused on breaking the monopoly of traditional utility companies, allowing competitive electricity service providers to enter the market [6]. Group 4: SST and SOFC Technologies - SST simplifies the power architecture and enhances construction efficiency by directly converting high-voltage electricity to the required output for data centers, reducing equipment and wiring complexity [8][9]. - SOFC is an efficient technology that converts fuel into electricity through electrochemical reactions, offering high efficiency and low emissions, making it suitable for data centers [10]. Group 5: Economic Viability and Reliability - SST and SOFC can significantly shorten construction cycles and reduce capital costs for data centers, which face challenges in power supply stability and grid connection [11]. - The reliability of SST and SOFC addresses the high demand for continuous power supply in data centers, mitigating risks associated with grid instability [11]. Group 6: Market Space and Valuation Outlook - The North American data center construction is projected to grow significantly from 20GW in 2026 to 53GW in 2030, indicating a substantial market opportunity for SST and SOFC [12]. - The estimated market size for SST in North America is expected to grow from 16 billion RMB in 2026 to 265 billion RMB by 2030, while SOFC's market size is projected to increase from 25 billion RMB to 198.7 billion RMB in the same period [12][13]. Group 7: Company Insights - Companies like Sifang Co. and Siyuan Electric are positioned to benefit from the growing demand for SST, with projected net profits exceeding 1.2 billion RMB and 4 billion RMB by 2027, respectively [13][14]. - Sifang Co. is estimated to achieve a market value increase of over 34 billion RMB from SST by 2027, while Siyuan Electric could see an increase of over 45.5 billion RMB [14].

Core Viewpoint - The article discusses the recent developments in the U.S. power generation landscape, highlighting the transition towards natural gas and renewable energy sources, and the implications for the semiconductor and AI industries, particularly in relation to solid-state transformers (SST) [3][9][16]. U.S. Power Generation Status - The U.S. power generation structure is shifting towards natural gas and renewable energy, with solar and battery storage growing rapidly [9]. - As of the end of 2024, coal-fired power generation capacity is projected to be 188.95 GW, down from 314 GW in 2011, indicating a significant decline in the coal industry [10]. - Natural gas generation capacity is expected to reach 571.23 GW by 2024, constituting 43.12% of the total capacity, but new installations are limited due to supply chain issues [11]. - Nuclear power capacity stands at 103.58 GW, with a 2.9% year-on-year increase, but new nuclear plants have not been built in nearly 40 years [12][13]. - Renewable energy capacity is projected to reach 313 GW by the end of 2024, with wind and solar accounting for 80% of new installations, although their effective contribution is limited compared to traditional power sources [14][15]. SST Industry Development Trends - The demand for solid-state transformers (SST) is increasing as they address the challenges faced by data centers, particularly in reducing power consumption and space [16]. - The SST market is currently dominated by non-mainland manufacturers, with companies like Eaton and Delta leading the way [17]. - The estimated market size for SST in North America from 2026 to 2030 is projected to grow significantly, reaching 265 billion RMB by 2030 [18]. Valuation Analysis - The SST industry is expected to have a net profit margin of around 30-40% from 2026 to 2030, with projected net profits increasing from 6.4 billion RMB in 2026 to 79.5 billion RMB in 2030 [19]. - The estimated market capitalization for the SST industry could reach up to 1.59 trillion RMB by 2030, based on average valuation multiples [19]. A-Share Core Company Valuation - Companies like Jinqiao Technology and Kingpan Technology are positioned to benefit from the SST market, with projected net profits of over 8 billion RMB and 17 billion RMB respectively by 2027 [21][23].