Core Insights - The report from Nomura Securities highlights the urgent need for advanced cooling solutions in AI chips due to rapidly increasing thermal design power (TDP) levels, with projections indicating that TDP for mainstream AI chips will rise from 600-700W in 2023 to potentially over 3500W by 2027 [3][4][10]. AI Chip Cooling Demand - The TDP of AI chips is expected to escalate significantly, with Nvidia's Blackwell series reaching 1000-1400W by 2025 and the Rubin series potentially hitting 2300W in 2026, and 3500W in 2027 [3][4]. - Traditional single-phase liquid cooling solutions are nearing their limits, necessitating new technological breakthroughs to handle TDPs above 2000-3000W [4]. Microchannel Cold Plates (MCL) - MCL is identified as the most practical solution for cooling chips exceeding 3000W post-2027, integrating heat spreaders and cold plates to reduce thermal resistance [5][7]. - MCL maintains compatibility with existing supply chains, utilizing current cooling fluids and components, unlike two-phase liquid cooling which requires extensive redesign [7]. - There are three main challenges to MCL mass production: design precision of microchannels, manufacturing capabilities, and supply chain coordination [8][9][10]. Thermal Interface Materials (TIM) - Upgrading TIM is crucial, with current materials like graphite films being insufficient for future TDPs; alternatives like indium TIM show promise but face challenges in assembly and interface treatment [10][11]. Other Technologies - Emerging technologies such as TSMC's Si integrated microcoolers and Microsoft's embedded microfluidics are considered less likely to be implemented in the short term due to scalability issues [11]. Market Opportunities for Traditional Cooling Manufacturers - Traditional cooling manufacturers like AVC and Auras are expected to see growth due to overlooked liquid cooling demands for non-core chips and the overall acceleration of liquid cooling adoption in AI servers [12][13]. - The market for liquid cooling components in AI servers is projected to grow from $1.2 billion to $3.5 billion between 2025 and 2027, with a compound annual growth rate exceeding 60% [12]. Investment Targets - Jentech is highlighted as a leading player in the microchannel market, with expected revenue from MCL contributing significantly to its overall earnings by 2028 [15]. - AVC and Auras are also recommended for investment, with AVC being a key supplier for Nvidia and Auras having advantages in manifold components [15].

Core Viewpoint - The article discusses the comparison of Scale Up solutions in AI servers, focusing on the UALink technology promoted by Marvell and the current mainstream Scale Up approaches in the international market [1][3]. Comparison of Scale Up Solutions - Scale Up refers to high-speed communication networks between GPUs within the same server or rack, allowing them to operate collaboratively as a large supercomputer [3]. - The market for Scale Up networks is projected to reach $4 billion in 2024, with a compound annual growth rate (CAGR) of 34%, potentially growing to $17 billion by 2029 [5][7]. Key Players and Technologies - NVIDIA's NVLink technology is currently dominant in the Scale Up market, enabling GPU interconnection and communication within server configurations [11][12]. - AMD is developing UALink, which is based on its Infinity Fabric technology, and aims to transition to a complete UALink solution once native switches are available [12][17]. - Google utilizes inter-chip interconnect (ICI) technology for TPU Scale Up, while Amazon employs NeuronLink for its Trainium chips [13][14]. Challenges in the Ecosystem - The current ecosystem for Scale Up solutions is fragmented, with various proprietary technologies leading to compatibility issues among different manufacturers [10][22]. - Domestic GPU manufacturers face challenges in developing their own interconnect protocols due to system complexity and resource constraints [9]. Future Trends - The article suggests that as the market matures, there will be a shift from proprietary Scale Up networks to open solutions like UAL and SUE, which are expected to gain traction by 2027-2028 [22]. - The choice between copper and optical connections for Scale Up networks is influenced by cost and performance, with copper currently being the preferred option for short distances [20][21].

Core Viewpoint - Nvidia has completely exited the Chinese market due to U.S. export controls, resulting in a market share drop to zero [1][8]. Group 1: Nvidia's Market Exit - In 2022, the U.S. first implemented AI chip export restrictions, with Nvidia holding over 90% market share in China [4]. - In 2024, Nvidia shipped 600,000 to 800,000 units of the H20 chip to China, which had only 15% of the performance of the H100 [5]. - By April 2025, the H20 chip was included in export controls, leading Nvidia to stop sales and recognize a $4.5 billion inventory loss [6]. - In August 2025, the H20 received an export license but was abandoned by Chinese customers due to security reviews [7]. - By October 2025, Nvidia's revenue from China plummeted from $17.1 billion to negligible levels [8]. Group 2: Domestic Market Dynamics - Despite the exit from the AI chip market, desktop GPUs, except for a few high-end models, can still be traded in China [9]. - Nvidia's recent DGX Spark can still be purchased in China, indicating that some products are still available despite restrictions [10]. Group 3: Cambricon's Q3 Report - Cambricon reported Q3 revenue of 1.727 billion yuan, with a net profit of 567 million yuan and a net profit margin of 32.8%, down from 36.08% in the first half of the year [11]. - The market expected higher revenue, with projections of 2.4 billion yuan for Q3 based on previous guidance of 5-7 billion yuan for the year [11]. - The launch of the 690 chip occurred faster than anticipated, indicating strong R&D capabilities, but the average selling price increased, leading to a decline in overall shipment volume [11]. Group 4: Inventory and Client Base - Cambricon's inventory rose from 2.69 billion yuan to 3.7 billion yuan, likely including a significant amount of HBM [12]. - Contrary to the belief that ByteDance is Cambricon's only major client, the company also serves other CSPs, national supercomputing centers, leading security firms, and several automotive companies [14]. Group 5: Industry Outlook - The ban on Nvidia's restricted AI chips is expected to benefit domestic GPU/NPU companies, including Huawei and Cambricon [14]. - By 2027, it is projected that China's GPU self-sufficiency rate could reach 82% [15]. - The long-term outlook for domestic AI chips remains positive [16].

Core Insights - The article emphasizes the importance of "Hardware-Assisted Verification" (HAV) and "Shift-Left Verification" strategies in the development of complex System-on-Chip (SoC) systems, highlighting that these approaches are essential for improving development efficiency and reducing hardware and software failure risks [1]. Group 1: HAV Technology Overview - Siemens has launched the Veloce™ CS system, which includes three core platforms: Veloce™ Strato CS (hardware emulation platform), Veloce™ Primo CS (enterprise-level prototyping platform), and Veloce™ proFPGA CS (software prototyping platform) [3]. - Strato CS and Primo CS operate on a highly consistent architecture, sharing the same operating system (Veloce OS) and applications (Veloce Apps), enabling seamless switching between the two and significantly enhancing verification efficiency, with a potential increase of up to 3 times and a reduction in total ownership costs by approximately 6 times [3]. Group 2: Modular Design and Scalability - The Veloce proFPGA CS hardware system features a modular design that allows users to combine various components, ranging from a single FPGA with 80 million gates to a configuration of 180 FPGAs with a total capacity of 14.4 billion gates [4]. - proFPGA CS shares front-end tools and some VirtuaLAB resources with Strato CS and Primo CS, facilitating easy transitions between different platforms for users [4]. Group 3: Upcoming Events and Presentations - A series of HAV technology seminars are scheduled, including sessions on improving SoC and system design verification efficiency using the Veloce CS ecosystem, enhancing hardware prototyping methodologies with proFPGA CS, and accelerating high-performance RISC-V SoC verification [5][6]. - The seminars will also cover the role of Strato CS in supporting efficient hardware-software co-verification for Arm Neoverse CSS and introduce the next-generation virtual platform, Innexis, which empowers SoC design verification [6].

Core Viewpoint - The article emphasizes a sustained optimism towards AI investments, highlighting that current spending is not a bubble but rather a significant growth opportunity in the sector [1][2]. AI Investment Trends - By 2025, annual AI-related spending in the U.S. is projected to reach approximately $300 billion, with a notable increase of $277 billion compared to the average in 2022 [2][3]. - The growth in AI spending has been partially driven by tariff policies, leading to preemptive equipment purchases, although overall spending remains high [2]. Technological Support for AI Investment - AI is expected to enhance labor productivity by 15% over a decade if widely adopted, with many companies reporting productivity increases of 25%-30% post-AI deployment [3][6]. - The demand for computational power to train large language models is growing at an annual rate of 400%, significantly outpacing the 40% annual decrease in computing costs [7]. Macroeconomic Context - Current AI investment levels, while nominally high, represent less than 1% of GDP, indicating room for growth compared to historical peaks in infrastructure and technology investments [7][8]. - The potential economic value generated from AI productivity improvements is estimated to be between $5 trillion and $19 trillion, far exceeding current investment levels [8][9]. Market Structure and Competition - The AI market exhibits varying levels of competition across different layers, with hardware providers like Nvidia enjoying a dominant position, while application layers face intense competition [10][11]. - The rapid pace of technological change in AI may diminish the advantages of early adopters, complicating the landscape for long-term winners [10][11]. Industrial Growth and Financial Projections - Industrial companies like Hon Hai Precision Industry (Industrial Fulian) are expected to see a compound annual growth rate (CAGR) of 45% in net profit from 2025 to 2027, driven by the AI server business [12][15]. - High expectations for revenue and profit growth are reflected in adjusted financial forecasts, with projected revenues for 2026 reaching ¥1.47 trillion and net profits of ¥564.32 billion [15][16]. Energy Demand and Supply Challenges - By 2030, global data center electricity demand is expected to increase by 175%, significantly impacting energy consumption patterns in the U.S. [21][22]. - The report outlines six key factors influencing electricity demand, including AI's pervasiveness, productivity of computing resources, and the impact of energy prices [22][23][24]. Investment Opportunities - Companies focusing on ensuring reliable electricity and water resources, meeting new electricity demands, and enhancing efficiency are highlighted as key areas for investment [26][27].

Core Viewpoint - The article emphasizes the importance of power and energy efficiency in the second half of AI data centers, highlighting the ongoing electricity shortages in the U.S. and the impact of data centers on electricity costs [2][4]. Group 1: AI Data Center Transformation - The traditional computing centers are evolving into AI factories, making power infrastructure a critical factor for deployment and scalability [7]. - NVIDIA proposes an 800VDC power distribution system combined with multi-time scale energy storage to address the explosive power demands of AI workloads [7][10]. Group 2: Technical Innovations - The shift from traditional low-voltage systems to an 800VDC architecture eliminates unnecessary AC-DC conversions, enhancing overall efficiency to over 90% [10][12]. - The new architecture supports high-density GPU clusters, allowing for scalability exceeding 1 megawatt per rack while reducing copper cable usage by 157% [12][13]. Group 3: Industry Collaboration - Building the 800VDC ecosystem requires collaboration across the industry, with NVIDIA partnering with various silicon suppliers and power system component partners [11]. - The Open Compute Project (OCP) is facilitating the establishment of open standards for voltage ranges and connectors [11]. Group 4: Solid-State Transformer (SST) Technology - SST technology is identified as a key solution for the next generation of data centers, with increasing demand in North America and significant market potential [21][22]. - Major companies like NVIDIA, Google, and Microsoft are actively developing SST solutions, with NVIDIA's Rubin architecture expected to adopt SST as a standard [21][22]. Group 5: Market Potential and Projections - The global market for SST could reach 800-1000 billion yuan by 2030, assuming a 20% penetration rate in new AI data centers [23]. - The demand for efficient power solutions is driving the rapid adoption of SST and HVDC technologies, with significant advancements expected by 2026 [22][24].

其中Strato CS和Primo CS运行在高度一致 的架构上,共享同一操作系统 (Veloce OS)、 解决方案与应用(Veloce Apps)，这使得 Strato CS 与 Primo CS 能实现无缝切换,大 幅提升验证效率,缩短学习与迁移成本。此 外,这种统一架构还能加速调试与任务部署 ――验证效率提升多达 3倍,总拥有成本可 降低约6倍。 在现代系统复杂性的不断上升的当下,"软 硬件协同验证"与"验证左移"(Shift-Left Verification)已成为推动复杂SoC系统开 发的关键策略。Hardware-Assisted Verification(HAV)技术,尤其是在 SoC 系 统验证中,已成为不可或缺的核心工具。任 何 SoC 开发团队都必须在设计初期就慎重 选择 HAV 工具和方法,以提高开发效率并降 低后期软硬件故障风险。 立即报名 /eloce CS >>>>> =位一体、统一架构的 HAV 方案 西门子最新推出的 Veloce™ CS 系统 包括三大核心平台: | ◆ Veloce™ Strato CS | 硬件仿真平台 | | --- | --- | | ◆ Velo ...

在现代系统复杂性的不断上升的当下,"软 硬件协同验证"与"验证左移"(Shift-Left Verification)已成为推动复杂SoC系统开 发的关键策略。Hardware-Assisted Verification(HAV)技术,尤其是在 SoC 系 统验证中,已成为不可或缺的核心工具。任 何 SoC 开发团队都必须在设计初期就慎重 选择 HAV 工具和方法,以提高开发效率并降 低后期软硬件故障风险。 立即报名 /eloce CS >>>>> =位一体、统一架构的 HAV 方案 西门子最新推出的 Veloce™ CS 系统 包括三大核心平台: | ◆ Veloce™ Strato CS | 硬件仿真平台 | | --- | --- | | ◆ Veloce™ Primo CS | 企业级原型平台 | | ◆ Veloce™ proFPGA CS | 软件原型平台 | 其中Strato CS和Primo CS运行在高度一致 的架构上,共享同一操作系统 (Veloce OS)、 解决方案与应用(Veloce Apps)，这使得 Strato CS 与 Primo CS 能实现无缝切换,大 幅提升验证效率,缩 ...

Core Viewpoint - The article discusses the impact of AI large language models, particularly GPT-5, on the demand for memory components such as HBM, DRAM, and NAND, suggesting a potential memory supercycle driven by AI inference workloads [4][8]. Memory Demand Analysis - The demand for HBM and DRAM is primarily driven by the inference phase of AI models, with GPT-5 estimated to require approximately 26.8 PB of HBM and 9.1 EB of DRAM if a 50% cache hit rate is assumed [8][10]. - NAND demand is significantly influenced by retrieval-augmented generation (RAG) processes, with an estimated requirement of 200 EB by 2025, considering data center capacity adjustments [8][11]. Supply and Demand Dynamics - The global supply forecast for DRAM and NAND indicates that by 2025, the supply will be 36.5 EB and 925 EB respectively, with GPT-5's demand accounting for 25% and 22% of the total supply [9]. - The article highlights a shift from oversupply to a shortage in the NAND market due to increased orders from cloud service providers, leading to price increases expected in late 2025 and early 2026 [11][12]. Beneficiary Companies - Companies such as KIOXIA and SanDisk are identified as key beneficiaries of the NAND price increases, with KIOXIA having the highest price elasticity but facing debt risks, while SanDisk is expanding its enterprise segment [12]. - Major manufacturers like Samsung and SK Hynix are positioned to benefit from both HBM and NAND markets, although their valuations may already reflect some of the positive outlook [12]. Market Outlook - Analysts predict that the current cycle is in its early stages, with profitability expected to begin in Q4 2025 and a potential explosion in demand in 2026, particularly for companies like SanDisk [13]. - The article notes several risk factors that could impact the sustainability of this cycle, including potential overestimation of cloud orders and the possibility of increased NAND production leading to oversupply by 2027 [13].

英伟达将在10月15日正式发售DGX Spark，今天他又亲自送给了马斯克一个。 2016年，老黄向马斯克交付了第一款为AI优化的GPU，当时马斯克还是OpenAI的投资人。而9年后，老 黄又交给马斯克英伟达最小的超算处理器。 简介 NVIDIA DGX Spark 是英伟达推出的一款革命性 AI 桌面超级计算机，被誉为"世界上最小的 AI 超级计 算机"。它将数据中心级别的计算能力浓缩到紧凑的桌面设备中，专为 AI 开发者和研究人员设计，帮助 他们在本地高效运行大型 AI 模型，而无需依赖云端资源。这款产品于 2025 年 CES 展会上首次亮相， 原计划 5 月上市，但因硬件优化和全球因素推迟至 10 月 15 日正式开售，起售价为 3999 美元（约合人 民币 3.5 万元）。 配置 核心规格与性能 连接性与扩展 DGX Spark 的接口丰富，支持现代办公需求： 特别值得一提的是，它允许两个 DGX Spark 通过高速网络互联，形成一个双机集群，总内存达 256GB，能处理高达 4050 亿参数 的超大规模模型，实现无缝扩展。 处理器与架构 ：搭载 NVIDIA GB10 Grace Blackwe ...