最近花了很多时间在调研谷歌光交换机OCS，查了很多资料，结合跟业内专家沟通的结果，写成了 本篇超过1.4万字的内容，本篇文章包含6个章节，深度分析OCS的技术、发展、合作商和价值量的 拆解，文中给的数据都非常详细，相信看完这篇文章，对OCS技术和整个产业链的认识提高不止一 个level。 1、谷歌AI的强劲势头 本来这篇文章是讲AI的Networking方面的内容，但谷歌AI近期的表现实在是过于亮眼，我们就小篇 幅提一下。 2025年8月26日，谷歌推出Gemini 2.5 Flash Image（代号Nano Banana），其图像生成与编辑功能在细 节和语义一致性上表现优异，覆盖美式证件照、像素风游戏、微缩模型等多种应用场景。上线不到 一个月，新增2300万用户，生成超过5亿张图片，助力Gemini APP在9月14日于美国、印度、加拿大 等多国应用商店登顶下载榜。此外，5月至8月，谷歌发布多项多模态大模型更新，包括Veo 3音视频 同步生成、Genie 3实时交互世界、Imagen 4图片生成以及Pixel 10系列AI硬件整合，凸显其在AI研发 上的领先优势。Gemini 3.0预计于2025年底推出 ...

以下文章来源于More Than Semi ，作者猫叔 More Than Semi . More Than SEMI 半导体行业研究 从昨天开始，大家都在讨论英伟达Rubin的这个新的液冷方案，热度非常高。盘中JS科技因为3d打印微 通道液冷板送样的消息一度还涨停了.. 但其实这件事几天就有媒体在报道了，我们星球中三天前就发过这个信息，只是当时市场对这个讨论度 并不高。 今天突然就有几个球友都在问我关于这个微通道盖板的情况，下面我们先看下JP Morgan和Morgan Stanley是如何分析这个液冷方案，再结合今天实际跟液冷厂的朋友聊的情况来分析一下。 1、投行的观点 大摩和摩根大通的这两份报告还是写了不少分析的，也给了几个比较明确的数据。 首先，什么是微通道盖板？ 价值量的增加 image-20250916224831268 按照JP Morgan给出的解释，microchannel lid是封装级直接集成了热扩散器（heat spreader）和冷板 （cold plate）的功能，能高效实现热量传导与散发 。它通过在刻有微通道的铜基板上加上盖板，并利 用歧管分配冷却液，将芯片产生的热量带走，属于 ...

以下文章来源于More Than Semi ，作者猫叔 More Than SEMI 半导体行业研究 今天的新闻比较多，一个个来看。 川总表示谈判很顺利，具体谈了啥咱不知道，从市场表现也能看出来，今晚美股开盘后，中概都涨 的不错。这是一个好的迹象，合作才能共赢。 More Than Semi . 再来看花旗上调胜宏这事，好几个群里都在讨论。 花旗的报告质量确实很高，都会直接给出很多他们自己调研到的核心数据，像出货量、产能、价格 等，之前大家转发比较多的那篇光模块的报告也是出自花旗。 其实花旗给到什么价，意义并不大，哪怕再调高100，也只是吸引一下眼球，我们还是要关注报告中 更有意义的部分。 1、PCB的价值量，花旗认为每GPU的PCB价值量，讲从GB200/GB300的375美元，上升到VR200的 863美元。跟我们在上周的文章中给的数据基本一致，说明业内对这块的看法基本一致。 英伟达Rubin CPX 的产业链逻辑 再补充个背景知识，可能有些对Rubin不太了解的同学不清楚上面提到的VR 和VR Ultra的区别。 这里的VR机架就是指Rubin NVL144 机架，也就是VR200，而VR Ultra是 ...

More Than Semi . More Than SEMI 半导体行业研究 昨天晚上，根据台湾经济日报的消息，美光暂停报价，可能要涨价20%-30%。几个月前我们分析过 DDR4的那波涨价，星球中我们每个月也会跟踪UBS的Memory monthly，关于Memory（DDR、 HBM、NAND）的需求、产能等的详细数据。 以下文章来源于More Than Semi ，作者猫叔 存储芯片是用途非常广，2024 年存储市场在手机、PC、服务器等需求复苏或提升的推动下创造了 1670亿美元的历史新高，差不多1.2万亿人民币，其中 NAND Flash 市场规模达696 亿美元，DRAM 市场规模达 973 亿美元。这篇文章我们聊一聊存储器的概述、市场和产业链。 存储芯片的概述 存储芯片是现代电子设备的基础部件，它负责数据的临时存储或长期保存。根据是否断电丢失数 据，存储芯片主要分为易失性存储和非易失性存储两大类。易失性存储断电后数据会消失，比如常 见的内存条；非易失性存储则能保留数据，比如U盘或固态硬盘。 先谈易失性存储，主要有三种常见类型。 第一种是静态随机存储器（SRAM），它的优点是速度非常快，因为内部 ...

9月17日上海见~ 9月17日， 2025全球AI芯片峰会 将在上海浦东喜来登由由大酒店正式举行！ 本次峰会由智一科技旗下 智猩猩 与 芯东西 共同举办，将以 "AI大基建 智芯新世界" 为主题，聚焦AI 大时代的新基建热潮，解码大模型下半场中国芯破局。 从2018年3月举办国内首场AI芯片峰会至今，七年来，除了2021年受疫情影响外，全球AI芯片峰会基本 上保持每年一届的节奏，共邀请 180+ 位产学研大咖分享前沿研究、创新洞见、落地进展与行业趋势， 成为了解国内外AI芯片发展动态的重要窗口，也是目前国内在AI芯片领域里最为火热且最具影响力的产 业峰会之一。 2025全球AI芯片峰会为期一天，由 主论坛+专题论坛+技术研讨会+展览区 组成。其中， 主论坛 将于 上午开幕， 大模型AI芯片专题论坛、AI芯片架构创新专题论坛 将于下午在主会场、分会场一分别进 行； 存算一体AI芯片技术研讨会、超节点与智算集群技术研讨会 将在分会场二上下午先后进行，主要 面向持有闭门专享票、贵宾通票的观众开放。 同时，今年的峰会也将在会场外设置 展览区 ，以标展为主，将有超摩科技、奎芯科技、特励达力科、 Alphawave、芯来 ...

Core Viewpoint - The article discusses the rapid advancements in the AI chip development landscape, particularly focusing on the self-developed ASICs by major companies like Google, Meta, Amazon, and OpenAI, highlighting their production forecasts and strategic initiatives [2][4]. Group 1: Google - Google is expected to ship 2.7 million TPU units by 2026, a significant increase from the previous estimate of 1.8 million [5]. - The upward revision in Google's TPU shipment forecast is attributed to consistent monthly increases in expected output and strong demand from both Google and Broadcom, which is projected to reach between 2.7 million and 2.8 million units [5]. Group 2: Meta - Meta is actively pursuing ASIC development, with a notable increase in average selling price (ASP) by five times and nearly doubling shipment volume from 5nm to 3nm ASICs [6]. - Meta has initiated two projects for 2nm ASICs, with the high-end project named "Olympus" expected to start mass production in the second half of 2027, featuring advanced specifications [6][7]. - The second project, aimed at mid-range ASICs, will utilize a customer-owned technology (COT) model, with ongoing selection of external partners for backend design [6]. Group 3: Other Companies - Amazon Web Services (AWS) maintains its ASIC shipment forecast for 2026, with potential adjustments depending on capacity releases from other clients and product structure changes [8]. - OpenAI's ASIC is projected to begin mass production in Q4 2026, with an initial shipment volume of 136,000 units [8]. - Apple faces delays in its ASIC development due to internal disagreements, making the likelihood of mass production by 2026 very low [8]. - Oracle's ASIC is expected to start mass production between 2027 and 2028, potentially targeting a Chinese cloud customer [9].

Core Viewpoint - The article discusses the significance of Nvidia's Rubin CPX, highlighting its tailored design for AI model inference, particularly addressing the inefficiencies in hardware utilization during the prefill and decode stages of AI processing [1][2][3]. Group 1: AI Inference Dilemma - The key contradiction in AI large model inference lies between the prefill and decode stages, which have opposing hardware requirements [2]. - Prefill requires high computational power but low memory bandwidth, while decode relies on high memory bandwidth with lower computational needs [3]. Group 2: Rubin CPX Configuration - Rubin CPX is designed specifically for the prefill stage, optimizing cost and performance by using GDDR7 instead of HBM, significantly reducing BOM costs to 25% of R200 while providing 60% of its computational power [4][6]. - The memory bandwidth utilization during prefill tasks is drastically improved, with Rubin CPX achieving 4.2% utilization compared to R200's 0.7% [7]. Group 3: Oberon Rack Innovations - Nvidia introduced the third-generation Oberon architecture, featuring a cable-free design that enhances reliability and space efficiency [9]. - The new rack employs a 100% liquid cooling solution to manage the increased power demands, with a power budget of 370kW [10]. Group 4: Competitive Landscape - Nvidia's advancements have intensified competition, particularly affecting AMD, Google, and AWS, as they must adapt their strategies to keep pace with Nvidia's innovations [13][14]. - The introduction of specialized chips for prefill and potential future developments in decode chips could further solidify Nvidia's market position [14]. Group 5: Future Implications - The demand for GDDR7 is expected to surge due to its use in Rubin CPX, with Samsung poised to benefit from increased orders [15][16]. - The article suggests that companies developing custom ASIC chips may face challenges in keeping up with Nvidia's rapid advancements in specialized hardware [14].

以下文章来源于More Than Semi ，作者猫叔 Oracle给出未来5年AI云收入指引： 2026E：180 亿美元 2027E：320 亿美元 2028E：730 亿美元 2029E：1140 亿美元 2030E：1440 亿美元 也就是说，到2030年，再翻8倍。 More Than Semi . More Than SEMI 半导体行业研究 甲骨文的业绩指引 而财报中，4550亿美元的RPO（未执行订单）是财报的最大亮点，表明未来3-5年的收入已有强力保 障。 RPO 增长主要来自AI相关的云基础设施合同，与OpenAI、xAI、Meta等公司合作。 也许是受这个消息的刺激，今天AI算力基本都涨的不错，PCB、光模块、液冷。 昨天的文章中，我们给出了国内液冷龙头一些业务细节，如海外业务进展、规划，在昇腾产业链中 的占比。 液冷龙头的海外业务与规划 持续看好AI板块 前段时间谷歌、Meta、微软、英伟达和博通都出了最新的季报，几个CSP的业绩都非常炸裂，无论 营收还是净利，同比都增长了非常多，而且在财报出来后，都调高了今年的资本开支指引。 | 公司 | 财报季度 | 总收入 | 净收入 | 同比变 ...

Core Viewpoint - The article discusses the rapid growth of the domestic liquid cooling leader Y's overseas business, particularly its collaboration with Meta, which is a key driver for its growth in the liquid cooling industry [2][3]. Group 1: Collaboration with Meta - Liquid cooling leader Y's overseas business has seen significant growth, especially through its partnership with Meta, which is expected to have a demand of $800-900 million in liquid cooling-related fields over the next two years [2]. - The demand for pure liquid cooling cabinets is estimated to account for about 25% of Meta's total needs, with Tianhong Technology responsible for procurement [2][3]. - Tianhong Technology's own demand for liquid cooling switch cabinets is projected to reach 3-4 billion RMB, contributing significantly to liquid cooling leader Y's business growth [2][3]. Group 2: Supply Chain and Market Dynamics - Liquid cooling leader Y maintains direct communication with Meta, but the actual supply process is primarily through integrators like Tianhong Technology and Quanta [3]. - The agreement with Meta covers at least a two-year period, with potential for increased future demand [3]. - Although there are rumors of liquid cooling leader Y being Meta's exclusive supplier, it is currently the only point of contact for Tianhong Technology, with the possibility of other suppliers being introduced later [4]. Group 3: Collaboration with NVIDIA - Liquid cooling leader Y has begun supplying some connector products to integrators like Quanta and has entered NVIDIA's supplier list [6]. - The company has made initial progress on new projects with NVIDIA, including manifold and CDU products, although actual supply has not yet commenced for some items [6][7]. - The competitive landscape for CDU products includes major players like Cool Master, but liquid cooling leader Y believes there are still significant opportunities through procurement from NVIDIA and customer specifications [6][7]. Group 4: Product Details and Market Share - The value of the manifold for NVIDIA's NVL72 is estimated at $7,800-8,200, with the total value including connectors around $25,000 [7]. - Liquid cooling leader Y aims to achieve a market share of x% in the CDU, manifold, and cold plate markets by 2026, although it is expected to be lower than Cool Master [7]. - The company plans to optimize production management to narrow the gap with Taiwanese manufacturers, with a net profit margin of approximately xx% for its liquid cooling business [7][8]. Group 5: Future Plans and Market Outlook - The company aims for overseas revenue to reach XX billion by 2026, with an expected average gross margin of over 40% [8]. - The advancement of NVIDIA's Rubin architecture is anticipated to reshape the liquid cooling market, although liquid cooling will remain mainstream for the time being [8]. - The domestic liquid cooling market is expected to grow rapidly as chip supply issues are resolved, with liquid cooling leader Y positioned to benefit from this growth [8][11]. Group 6: Competitive Landscape - The domestic liquid cooling market is highly competitive, but liquid cooling leader Y maintains a leading position due to its comprehensive supply chain and experience [11]. - The company has accumulated a total capacity of 1.5-1.6 GW in liquid cooling application projects, ranking among the top in the country [11]. - Liquid cooling leader Y's products are recognized for their quality, with a focus on production control to ensure high yield rates [11].

Core Viewpoint - The article provides an in-depth analysis of the value breakdown of liquid cooling systems for GB200 and GB300, highlighting the changes in components, pricing, and supplier dynamics in the semiconductor industry [2][4]. Value Breakdown of Liquid Cooling Systems - The total value of the cooling plates for GB200 is calculated at $29,250, while for GB300, it is $31,770, reflecting a shift in design and pricing strategy [6][7]. - The total value for the cabinet liquid cooling system for GB200 is $35,680, and for GB300, it is $44,650, indicating an increase in overall system value due to component changes [10]. - The external liquid cooling system maintains a value of approximately $750,000 per cabinet, with future designs potentially increasing costs due to new requirements [11]. Supplier Dynamics - In the GB200 era, the external supplier landscape was dominated by a single player, Viavi, but the GB300 era is expected to see 5-6 suppliers, including Cooler Master and Delta, indicating a shift towards a more competitive and diversified supply chain [11]. - The internal cooling plate suppliers have also changed, with Cooler Master becoming the main supplier for GB300, while previously, AVC and Shuanghong were the primary suppliers for GB200 [12]. Story of Inveke - Inveke initially aimed to supply cooling plates directly to NVIDIA but shifted strategy to collaborate with Cooler Master as a secondary supplier to avoid direct competition [13]. - Inveke is recognized for its top-tier CDU products and is expanding into ASIC and switch liquid cooling solutions, aligning with industry trends towards full liquid cooling systems [13].