星球中每日还会更新Seeking Alpha、Substack、 stratechery的精选付费文章， 现在星球中领券后只需要 390元，即可每天都能看到上百篇外资顶尖投行科技行业的分析报告和每天的精选报告，无论是我们自 己做投资，还是对行业有更深入的研究，都是非常值得的。 还有专注于半导体行业分析的SemiAnalysis的全部分析报告： 想要看外资研报的同学，给大家推荐一个星球，在星球中每天都会上传几百篇外资顶尖投行的原文研 报：大摩、小摩、UBS、高盛、Jefferies、HSBC、花旗、BARCLAYS 等。 ...

Core Viewpoint - The report from Morgan Stanley highlights a significant increase in the global production of GB200 NVL72 racks, driven by the surging demand for AI computing, particularly in cloud computing and data center sectors [1][2]. Group 1: Production Forecast - The global total production of GB200 NVL72 racks is estimated to reach 2,000 to 2,500 units by May 2025, a notable increase from the previous estimate of 1,000 to 1,500 units in April [1]. - The overall production for the second quarter is expected to reach 5,000 to 6,000 units, indicating a robust supply chain response to market demand [1]. Group 2: Company Performance - Quanta shipped approximately 400 GB200 racks in May, a slight increase from 300 to 400 units in April, with monthly revenue reaching about 160 billion New Taiwan Dollars, a year-on-year increase of 58% [2]. - Wistron demonstrated a strong growth trajectory, shipping around 900 to 1,000 GB200 computing trays in May, a nearly sixfold increase from 150 units in April, with revenue growth of 162%, reaching 208.406 billion New Taiwan Dollars [2]. - Hon Hai shipped nearly 1,000 GB200 racks in May, with a forecast of delivering 3,000 to 4,000 racks in the second quarter, despite some decline in its cloud and networking business due to traditional server shipment slowdowns [2]. Group 3: Market Dynamics - The actual delivery volume of GB200 racks may be lower than the reported shipment figures due to the need for further assembly of Wistron's L10 computing trays into complete L11 racks, which involves additional testing and integration time [3]. - Morgan Stanley ranks the preference for downstream AI server manufacturers as Giga-Byte, Hon Hai, Quanta, Wistron, and Wiwynn, with Giga-Byte being favored for its potential in GPU demand and the server market [3]. - A report from Tianfeng Securities indicates that major hyperscale cloud providers are deploying nearly 1,000 NVL72 cabinets weekly, with the shipment pace continuing to accelerate [3].

Core Viewpoints - The article discusses the rapid growth and potential of the computing power leasing industry, particularly through the lens of CoreWeave, a significant player in this sector [2][11]. Company Overview - CoreWeave was established in 2017, originally as a cryptocurrency mining company, and has since pivoted to focus on AI cloud and infrastructure services, operating 32 data centers by the end of 2024 [2][3]. - The company has deployed over 250,000 GPUs, primarily NVIDIA products, and is a key provider of high-performance infrastructure services [2][3]. Business Model - CoreWeave offers three main services: bare-metal GPU leasing, management software services, and application services, with a focus on GPU leasing as the core offering [3][4]. - Revenue is generated primarily through two models: commitment contracts (96% of revenue) and on-demand payment, allowing flexibility for clients [4][5]. Financial Performance - In 2024, CoreWeave's revenue reached $1.915 billion, a year-over-year increase of over seven times, with Q1 2025 revenue at $982 million, reflecting a fourfold increase [8][9]. - The company has a remaining performance obligation of $15.1 billion, indicating strong future revenue potential [8]. Competitive Advantages - CoreWeave has optimized GPU utilization rates and efficiency, achieving significant performance improvements in AI training and inference tasks [7]. - The company has established strong relationships with NVIDIA, ensuring priority access to cutting-edge chips and technology [6][7]. Market Outlook - The AI infrastructure market is projected to grow from $79 billion in 2023 to $399 billion by 2028, with a compound annual growth rate of 38%, highlighting the industry's potential [11]. - The computing power leasing sector is expected to play a crucial role in the digital economy, driven by increasing demand for AI capabilities [11][14]. Future Growth Strategies - CoreWeave plans to expand its customer base, explore new industries, and enhance vertical integration with strategic partnerships [10]. - The management aims to leverage existing contracts and maintain a low leverage asset structure to support growth [10].

Group 1 - The article recommends a platform where users can access hundreds of top-tier foreign investment bank research reports daily, including those from firms like Morgan Stanley, UBS, Goldman Sachs, Jefferies, HSBC, Citigroup, and Barclays [1] - The platform also provides comprehensive analysis reports focused on the semiconductor industry from SemiAnalysis, along with selected paid articles from Seeking Alpha, Substack, and stratechery [3] - The subscription to the platform is currently available for 390 yuan, offering access to a wealth of technology industry analysis reports and selected articles daily, which is deemed valuable for both personal investment and deeper industry research [3]

Core Viewpoint - The article discusses the current market situation of the NVIDIA RTX 5090 graphics card, focusing on its price, rental market, computing power, power consumption, performance, heat generation, and networking capabilities since its release in January 2025 [1]. Pricing - The initial expected price of the RTX 5090 was over 40,000 yuan, but it has dropped to just over 20,000 yuan within four months, with some brands listed as low as 23,000 yuan on platforms like JD.com. This price decline is attributed to concerns over chip overheating, rumors of performance bottlenecks in multi-card setups, initial high pricing by manufacturers, and the competitive appeal of the previous generation RTX 4090 [2]. Rental Market - The high initial price of the RTX 5090 (over 30,000 yuan) led to slow development in the rental market. It wasn't until May, when prices fell, that some data centers began to offer RTX 5090 models for rent. Currently, the investment payback period for an 8-card machine is approximately four years, which may be too long for AI companies given the rapidly changing demand for computing power [3][6]. Computing Power - The RTX 5090 excels in computing power, particularly in AI training and inference scenarios, with a single card achieving 419 TFLOPS and an 8-card machine reaching about 3.4 PFLOPS. A cluster of 300 RTX 5090 cards can form a computing cluster capable of trillions of floating-point operations, making it advantageous for large language model training and high-performance computing tasks [4]. Power Consumption - The RTX 5090 has a rated power consumption of 575W, with peak consumption reaching up to 900W. An 8-card machine consumes approximately 6kW, leading to monthly electricity costs of around 3,600 yuan based on a rate of 0.6 yuan per kWh. This high power consumption increases operational costs and necessitates robust cooling and power supply systems [7]. Performance - In AI inference scenarios, the RTX 5090 supports low-precision calculations (FP8 and FP4), significantly enhancing efficiency. It shows about a 50% faster inference speed compared to the previous generation RTX 4090. In gaming, it outperforms the 4090 at 4K resolution, but optimal performance requires targeted optimization, especially in low-precision inference [8]. Heat Generation - The RTX 5090 faces heat issues primarily related to the chip and power connectors, particularly the 12V-2x6 connectors. Although such overheating incidents are rare, they require attention. Solutions include limiting peak power through driver or BIOS settings, using liquid cooling or turbo fans, and employing original power cables to avoid compatibility issues [9][10]. Networking - Initial concerns about potential "lock card" issues or performance bottlenecks in multi-card setups have not been substantiated in practical tests. Actual tests showed no such problems, and many companies using the RTX 5090 reported stable performance in NVLink and PCIe networking, making it suitable for building high-performance AI clusters [11].

最近TechInsights发布了一篇《Semiconductor Forecast Update and Top Supplier Rankings》的报告，里 面详细分析了对半导体市场的预测，和对目前半导体顶级供应商的排名，这篇文章我们结合这篇报 告的内容，看下TechInsights给出的顶级供应商的排名如何。 2025 年第一季度，全球经济——尤其是半导体行业——经历了一个动荡的开端。许多人担忧的事情 变成了现实：川普兑现了竞选前的承诺，作为其"解放日"计划的一部分，实施了广泛且显著的关税 政策。 2025 年第一季度宣布的几项关税直接或间接针对半导体行业，包括芯片、设备和材料。几乎每天都 在变化的关税政策使得半导体市场预测变得极为困难，堪比 COVID-19 疫情早期的挑战。 除了扰乱全球贸易外，关税还导致美国与其贸易伙伴之间的不和谐和紧张关系。在伯克希尔·哈撒韦 年度股东大会的问答环节中，即将离任的首席执行官沃伦·巴菲特表示："贸易不应该被用作武器。当 75 亿人都不喜欢你时，这是一个巨大的错误。我认为这既不正确，也不明智。" 下面进入正文，看下TechInsights给出的半导体的各个细分领域的 ...

前段时间有朋友让我写写关于先进封装的东西，其实手头跟封装相关的资料还是蛮多的，可写的内 容也比较多，我们这篇文章就先介绍一下半导体封装，主要参考海力士在官网的一篇科普文章，以 及部分Yole的报告。 半导体封装工艺的四个等级 电子封装技术与器件的硬件结构有关。这些硬件结构包括有源元件（如半导体）和无源元件（如电 阻器和电容器）。因此，电子封装技术涵盖的范围较广，可分为0级封装到3级封装等四个不同等 级。下图展示了半导体封装工艺的整个流程。首先是0级封装，负责将晶圆切割出来；其次是1级封 装，本质上是芯片级封装；接着是2级封装，负责将芯片安装到模块或电路卡上；最后是3级封装， 将附带芯片和模块的电路卡安装到系统板上。从广义上讲，整个工艺通常被称为"封装"或"装配"。 然而，在半导体行业，半导体封装一般仅涉及晶圆切割和芯片级封装工艺。 有源元件 ：一种需要外部电源才能实现其特定功能的器件，就像半导体存储器或逻辑半导体。 无源元件 ：一种不具备放大或转换电能等主动功能的器件。 电容器（Capacitor） ：一种储存电荷并提供电容量的元件。 封装通常采用细间距球栅阵列（FBGA）或薄型小尺寸封装（TSOP）的形式， ...

今天大家都在谈MS的这篇DeepSeek R2分析的报告，提前曝光了R2的性能和参数，我们简单总结一 下这个报告的核心内容： DeepSeek R2 使用了多达 1.2 万亿个参数，采用了新颖的架构，实现了运行成本的显著降低。其采用 混合专家混合（MoE）架构，有 780 亿个活跃参数。 并且R2 使用华为的 Ascend 910B 芯片进行训练，而非 NVIDIA 的芯片。 R2 增强了多语言覆盖能 力，能流畅处理非英语语言；扩展了强化学习，利用更大的数据集，使模型能够进行更具逻辑性和 更像人类的推理；增加了多模态功能，能够处理文本、图像、语音和视频数据；实现了推理时的缩 放，通过采用通用奖励模型（GRM），在推理过程中增加计算资源，从而提高了输出质量。 R2 具有高成本效益，输入成本为每百万代币 0.07 美元，输出成本为每百万代币 0.27 美元，而 R1 的 输入成本为 0.15-0.16 美元，输出成本为 2.19 美元。 由于这篇报告讲的人已经很多了，我们就不赘述了，而且报告也放到了星球中，有兴趣的朋友可以 到星球中看原文。 今天这篇文章来看另一篇AI的分析，Artificial Analysis ...

最近关于 HBM 的讨论越来越多，尤其是在涉及 AI 芯片领域时。HBM 即高带宽内存（High Bandwidth Memory），是一种特殊的 DRAM，通过垂直堆叠并利用硅片内名为 TSV（硅通孔，Through-Silicon Vias）的微小导线与处理器连接。TSV 技术允许直接连接多个 HBM DRAM 芯片，从而提升整体内存带 宽。 在生成式 AI 时代，内存带宽至关重要，因为模型训练往往受限于带宽而非计算能力。Transformer 模型 中的注意力机制需要存储和计算所有 token 之间的关系，内存需求与序列长度成二次方增长。类似地， 推理阶段内存也成为更大瓶颈，因为需要处理更长的上下文窗口和 Transformer 模型中扩大的键值缓存 （KV 缓存）——KV 缓存的内存消耗随 token 数量线性增长。 高速性能 ：HBM 的带宽可达每秒数 TB，比常规 DDR 内存快 20 倍以上。 低功耗 ：由于紧邻逻辑芯片，数据传输距离短，大幅节省能源。 面积效率 ：单位面积内容量密度最高。 为何 HBM 如此重要？ HBM 的核心优势如下： image-20250604192803128 SK ...

还有专注于半导体行业分析的SemiAnalysis的全部分析报告： 星球中每日还会更新Seeking Alpha、Substack、 stratechery的精选付费文章， 现在星球中领券后只需要 390元，即可每天都能看到上百篇外资顶尖投行科技行业的分析报告和每天的精选报告，无论是我们自 己做投资，还是对行业有更深入的研究，都是非常值得的。 想要看外资研报的同学，给大家推荐一个星球，在星球中每天都会上传几百篇外资顶尖投行的原文研 报：大摩、小摩、UBS、高盛、Jefferies、HSBC、花旗、BARCLAYS 等。 ...