March 23, 2026 01:27 AM GMT Greater China Semiconductors | Asia Pacific AI Supply Chain Tracker: Key Investor Feedback from GTC/ OFC We visited NVIDIA's GTC and the OFC last week along with investors. GTC raised more questions than it answered, while investors generally left OFC more confident on optics. We remain positive on cloud semis and CPO outlook. We revise Aspeed's PT, given rack disaggregation suggests more BMC demand. GTC – New server rack designs and chip schedule ( Exhibit 1 ): Investors asked i ...

本次大会有几个关键亮点： 第一，Agent AI 的应用开启了智能体的新拐点， 预计将成为未来的关注焦点，因为它能更广泛地替代人类工作。 第二，新发布 的芯片迭代速度加快，并显示出强烈的走向 CPO 的趋势。 第三，新产品在算 力大幅提升的同时，token 成本相较前一代产品降低了近 90%，约为原来的十 分之一。 第四，自研的 VeraWell CPU 平台相比 Grace 平台有显著提升，其 效率相比英特尔和 AMD 最新的 CPU 高出近两倍，核心数也翻倍至 88 核。未 来 VeraWell CPU 可与 GPU、LPU 等组合，面向推理服务器市场。 第五，散 热方面实现了 100%液冷，采用 45 度的温水进行冷却，降低了因使用压缩机 制冷而产生的高昂电力成本。 第六，产品模块化和集成度显著提高，例如 NV Switch 和 Rubin 节点。这使得服务器的组装时间大幅缩短，从原先的两天缩 短至约两个小时。 GTC 大会新架构与核心技术要点解读 20260322 摘要 Feynman 架构采用 1.6nm 制程，通过 CPO 交换机互联实现带宽密度 近 10 倍提升，旨在解决大规模 AIG 集群柜间 ...

Token需求爆炸式增长，云厂商纷纷提价。 2026年，全球云计算行业延续近二十年的"只降不升"定价惯例被集体打破。 1月，AWS率先 对GPU训练 实例 上调约15%，谷歌云随即宣布数据传输服务最高涨价100%。 3月，国内云厂商密集跟进，腾讯云率先上调自研大模型价格，涨幅最高达463%；阿 里云与百度智能云于本周宣布AI算力及存储产品涨价，最高涨幅34%。 这一切均将原因指向" 全球AI需求爆发、核心硬件成本 显著上涨" 。 这一轮云厂商涨价潮，表面上是成本传导，本质上是： 算力在当下正 在从基础设施变成稀缺资源。 过去几年，AI创业者习惯了一个相对宽松的算力环境，价格持续下行，资源随取随用。但 在2026年的当下，这个前提正在悄然失效。 超大规模云服务商的集群资源被牢牢锁定，小 型团队几乎无从批量获取 。 云服务厂商2026年数据中心资本支出预期较一年前大幅增长甚至翻倍，仍被市场认为"不 够用"。 这不是周期性的供需波动，而是结构性的产能短缺 。 锦 秋基金认为： 2026年当算力从"成本项"变成"战略资源"，它就不再只是财务决策，也关乎产品节奏、商 业模式乃至公司生死的核心变量。 谁能在正确的时间窗口 ...

在英伟达GTC大会的主题演讲中，CEO黄仁勋表示，AI Token正在成为一种新的货币形式，将在人才招 聘、预算制定和生产力提升方面发挥重要作用。大会于周四在加利福尼亚州圣何塞举行。 黄仁勋指出，AI Token也将越来越多地影响公司的发展进程和盈利能力。"Token是新的商品"，他说 道，随后补充："计算过去是基于检索的，现在是生成式的。" Token是现代AI计算的核心，就像比特曾经是基于CPU的传统计算的单位一样。AI正被集成到大多数新 的软件产品中。 "我们将把Token生成率从2200万提高到7亿——增长350倍"，黄仁勋说道。 这些芯片适用于英伟达所称的AI工厂，生成Token来帮助公司实施AI计划。"AI工厂的收入等于每瓦 Token数。在功耗限制下，每个未使用的瓦特都是收入损失"，黄仁勋表示。 到目前为止，在GTC大会上，英伟达的大部分Token信息都围绕推理而非训练展开，但推理不会涉及"巨 额"的Token成本，J. Gold Associates首席分析师Jack Gold表示。 "这个概念......将结构化数据与生成式AI融合，将在一个又一个行业中重复出现"，黄仁勋说道。 Token也 ...

先从第二层的芯片说起。这次最核心的硬件发布是Vera Rubin系统，但它更像一个"芯片联合国"。整套 系统横跨五个机架，内部集成了七种不同的芯片：Vera CPU负责高单核性能的通用计算，Rubin GPU主 打并行计算，Groq 3 LPU则专攻低延迟推理。这里值得多说一句Groq LPU的角色——黄仁勋花了不少 时间解释"低延迟与高吞吐是天敌"这个命题。 【文/观察者网专栏作者 心智观察所】 2026年3月17日凌晨，圣何塞SAP中心的灯光徐徐熄灭，空旷的场馆里响起一段乡村音乐。台下坐着一 万八千名观众，屏幕前还有数百万人在等待同一个人——黄仁勋。 几分钟后，他穿着那件熟悉的黑色皮夹克走上舞台，没有惊喜，也不需要。这件夹克，已经成了一种符 号：只要它出现在台上，就意味着未来一年全球AI产业的资源流向、技术路线，甚至地缘格局，都将 被重新定义。 这并非夸张。回顾整场GTC 2026主题演讲，表面上看是琳琅满目的新品发布——DLSS 5、Vera CPU、 Groq LPU、Vera Rubin NVL72、OpenClaw智能体操作系统、Nemotron联盟，甚至还有一个太空数据中 心项目。但如果你只盯着 ...

Investment Rating - The report maintains a "Recommended" rating for the electric power equipment industry [1]. Core Insights - The AI industry is transitioning from generative AI to inferential AI, with a significant increase in computational demand, approximately 1 million times over the past two years [2]. - Key performance indicators for AI factories include throughput efficiency and inference speed, with token generation speed being crucial for AI intelligence [2]. - NVIDIA anticipates that the revenue from its Blackwell and Rubin flagship chip product lines will exceed $1 trillion by 2027, a significant increase from the previously projected $500 billion [2]. - The Vera Rubin full-stack AI computing platform was officially launched, featuring a substantial increase in computing power, projected to rise by 40 million times over the next decade [2]. - The power supply for the Vera Rubin NVL72 is expected to be significantly enhanced, with a total power supply of 440KW, representing an increase of over 60% compared to previous models [2][3]. Summary by Sections Industry Progress - The AI industry is evolving towards inferential and intelligent AI, with a focus on optimizing AI infrastructure costs [2]. - NVIDIA's 2025 is defined as the "Year of Inference," aiming to become the most cost-effective and reliable AI infrastructure platform globally [2]. Product Developments - The Vera Rubin platform includes multiple advanced chips and has achieved a significant reduction in deployment time from two days to two hours [3]. - The platform's cooling system utilizes 100% liquid cooling, which enhances energy efficiency and reduces cooling costs [3]. Investment Recommendations - Suggested companies to watch include: - For cabinet power: Megmeet, Euron, Newray, and others [3]. - For external power: Jinpan Technology, Igor, and others [3]. - For liquid cooling solutions: Invec, Shunling Environment, and others [3]. - For data center storage: Sungrow, CATL, and others [3]. - For backup power solutions: Yiwei Lithium Energy, and others [3].

数据来源：聚源 -38% -19% 0% 19% 38% 58% 2025-03 2025-07 2025-11 2026-03 电子 沪深300 行 业 研 究 2026 年 03 月 18 日 投资评级：看好（维持） 行业走势图 相关研究报告 《OpenClaw 热潮加速端侧 Agent 渗 透，推理算力需求激增—行业点评报 告》-2026.3.16 《OpenClaw 催化 AI 终端热度，英伟 达 GTC 大会召开在即—行业周报》 -2026.3.15 《存储高景气度逐步向上游扩散， OpenClaw 登顶 GitHub 开启端侧 AI Agent 新形态—行业周报》-2026.3.8 Groq 3 LPU 与 GPU 协同作战，系统架构如期升级 陈蓉芳（分析师） 刘琦（分析师） chenrongfang@kysec.cn 证书编号：S0790524120002 liuqi1@kysec.cn 证书编号：S0790525020001 Groq 3 LPU 超预期：推理性能倍增，放量节奏提前 单芯片性能跃升：Groq 3 LPU 集成 500MB 的 SRAM，提供 150TB/s 带宽，是 HBM（2 ...

Core Insights - The report highlights a strong performance in various sectors, particularly in AI, communication, and automotive industries, driving significant growth in PCB business for ShenNan Circuit [10][11] - The report emphasizes the strategic advantages of Baofeng Energy in the coal-to-olefin industry, showcasing substantial revenue and profit growth [19][20] - The report discusses the impact of geopolitical tensions on raw material prices, particularly for Foster, which is expected to benefit from rising prices in the photovoltaic sector [15][16] Group 1: Company Performance - ShenNan Circuit achieved a revenue of 236.47 billion yuan in 2025, representing a year-on-year increase of 32%, with a net profit of 32.76 billion yuan, up 74% [10][11] - Baofeng Energy reported total revenue of 480.38 billion yuan for 2025, a 45.64% increase year-on-year, with a net profit of 113.50 billion yuan, reflecting a 79.09% growth [19][20] - Foster is positioned to benefit from the rising prices of EVA and POE films, with significant price increases noted in the report [15][16] Group 2: Industry Trends - The global PCB market is projected to grow from $85.2 billion to $123.3 billion from 2025 to 2030, with a CAGR of approximately 8%, driven by demand in data centers and high-speed communication [11] - The report indicates that the AI and physical AI sectors are expected to become significant growth points, with Nvidia's new technologies enhancing performance in these areas [6][7] - The photovoltaic industry is experiencing a shift due to rising raw material costs, which may lead to a more favorable competitive landscape for leading companies [15][16] Group 3: Investment Recommendations - The report suggests focusing on companies involved in CPO chips and packaging, optical fibers, PCB materials, server assembly, and power and cooling solutions as potential investment opportunities [8][6] - Specific companies highlighted for investment include Tianfu Communication, Longfly Fiber, and ShenNan Circuit, among others [8][10] - The report maintains a "buy" rating for Baofeng Energy and ShenNan Circuit, indicating confidence in their growth trajectories [19][10]

智通财经APP获悉，根据TrendForce集邦咨询最新AI Server研究，在大型云端服务供应商(CSP)加大自研芯片力道的情况下，英伟达(NVDA.US)在GTC 2026大会改为着重各领域的AI推理应用落地，有别于以往专注云端AI训练市场。通过推动GPU、CPU以及LPU等多元产品轴线分攻AI训练、AI推理需 求，并借由Rack整合方案带动供应链成长。 TrendForce集邦咨询表示，随着以谷歌(GOOGL.US)、亚马逊(AMZN.US)等CSP为首的自研芯片态势扩大，预估ASIC AI Server占整体AI Server的出货比例 将从2026年的27.8%，上升至2030年的近40%。 观察Rubin供应链进度，预计2026年第二季存储器原厂可提供HBM4给Rubin GPU搭载使用，助力英伟达于第三季前后陆续出货Rubin芯片。至于英伟达 GB300、VR200 Rack系统出货进程，前者已于2025年第四季取代GB200成为主力，预估至2026年出货占比将达近80%，而VR200 Rack则约在2026年第三季 度末可望逐步释放出货量能，后续发展仍需视ODM实际进度而定。 另外，AI从生 ...

Core Insights - The AI industry is shifting focus from model training to inference, with companies like NVIDIA adapting to this change by introducing new products and strategies [1][3][6] - NVIDIA's CEO Jensen Huang announced the launch of the Groq 3 LPU, a dedicated AI inference chip, during the GTC 2026 event, aiming to capture a significant share of the inference chip market [1][2] - NVIDIA's revenue forecast for its Blackwell and Rubin product lines has doubled to $1 trillion by the end of 2027, indicating strong market confidence [1] Group 1: NVIDIA's Strategic Moves - NVIDIA has launched the Vera Rubin platform, which includes seven new chips, enhancing its capabilities in AI inference [2] - The Groq 3 LPU is designed to significantly increase token throughput from 100 tokens per second to 1500 tokens or more, supporting advanced AI interactions [2] - NVIDIA's acquisition of Groq's core technology assets for approximately $20 billion in December 2025 has positioned the company to leverage Groq's innovations in its product offerings [3] Group 2: Market Trends and Predictions - The market is witnessing a shift in AI chip shipments, with non-GPGPU chips expected to rise from 36% in 2024 to 45% by 2027, while GPGPU shipments will decline from 64% to 55% [3] - The demand for inference capabilities is being driven by the rise of intelligent agents, which focus more on inference rather than training [6] - NVIDIA's introduction of the LPU is a strategic response to the evolving AI compute demands, addressing the need for efficiency and lower latency in inference scenarios [3][6] Group 3: Ecosystem and Infrastructure Development - NVIDIA is enhancing its ecosystem by introducing the NeMoClaw reference architecture, which includes security and privacy features for enterprise AI systems [6] - The company has also launched the Vera Rubin DSX AI Factory reference design, aimed at optimizing AI infrastructure for scalability and performance [6][7] - Huang emphasized that in the AI era, intelligent tokens are the new currency, and AI factories are essential for generating these tokens, highlighting the importance of infrastructure in AI development [7]