Core Insights - The AI era is facing a critical challenge with the shortage of high-speed interconnect networks, which is essential for the efficient operation of large-scale computing clusters [1][10] - Domestic companies are making strides in developing their own computing chips, but the core technology for high-speed interconnects remains dominated by Nvidia, posing a significant risk to the industry [1][10] Group 1: Industry Challenges - The transition from GPU to high-speed interconnects is becoming a new bottleneck as the scale of computing clusters increases from thousands to tens of thousands of nodes [1][4] - The communication time in distributed training can account for 30-50% of the total time, indicating that a significant portion of the investment in computing power is wasted on data transfer rather than computation [4][5] - The demand for high-speed networks has increased by 10 to 20 times as servers now require multiple network cards to support GPU-centric architectures [6] Group 2: Technological Landscape - There are two main technological routes in the high-speed network domain: RoCE and InfiniBand, with the latter being the preferred choice for high-performance computing due to its superior performance metrics [7][10] - InfiniBand networks are used in approximately 60% of the world's high-performance computing systems, and they are almost standard in the largest AI training clusters [10] Group 3: Domestic Developments - In response to the challenges posed by the dominance of foreign technology, companies like Zhongke Shuguang have developed their own high-speed network solutions, such as the scaleFabric, which is fully self-developed [2][11] - The decision to pursue a fully self-developed InfiniBand system was driven by the inadequacy of available commercial IPs and open-source solutions to meet the performance and reliability requirements for large-scale clusters [12]

Core Viewpoint - Nvidia remains a leader in AI hardware, with significant growth driven by demand for its AI chip systems, but its stock has underperformed compared to the broader semiconductor market [1][2][3] Nvidia - Analysts forecast a 73% increase in Nvidia's earnings this year, with a 70% jump in revenue to $367 billion [2] - Nvidia's stock has gained only 2% in six months, while the PHLX Semiconductor Sector index has appreciated by 27% during the same period [3] Alphabet - Alphabet is positioned as a comprehensive AI stock, integrating AI across its services like Google Search, Gemini chatbot, Google Cloud, and YouTube [6] - The Gemini app has over 750 million monthly users, and AI Mode queries in Google Search are three times longer than traditional searches [8] - Alphabet's Google Cloud backlog increased by 55% sequentially to $240 billion, following a 48% year-over-year revenue jump [10] - The company has a potential $900 billion revenue opportunity from selling custom AI chips (TPUs) to third parties [12] - Alphabet's stock trades at 9 times sales, compared to Nvidia's 20 times, indicating more upside potential for Alphabet [14] Snowflake - Snowflake operates a cloud-based data platform that supports AI tools for data analysis and sharing [15] - The company has over 9,100 customers using its AI solutions, more than double the previous year's figure [17] - Snowflake's customer base grew by 21% year-over-year, with remaining performance obligations (RPO) increasing by 42% to $9.77 billion [18] - The company anticipates exceeding a 27% growth in product revenue for fiscal 2027, with an expected improvement in operating margin to 12.5% [19][20] - Snowflake's sales multiple of 13 is lower than Nvidia's, suggesting potential for greater upside as growth accelerates [22]

Economic Vulnerability - The South Korean economy is highly vulnerable to oil imports, which has been highlighted by recent currency movements [1] - The retail sector remains resilient despite current uncertainties, with investors showing interest in overseas markets, particularly US equities, contributing to currency weakness [2] Market Sentiment and Government Policy - The South Korean government is pushing for a vibrant capital market, influenced by previous administrations and modeled after Japan's market policies [4][5] - There is a significant outflow of capital from Korean retail investors to US markets, with over $100 billion invested abroad, indicating a need to attract investment back home to stimulate a bull market [6] Market Dynamics and Risks - Concerns are growing about the sustainability of the current bull market, with potential risks resembling past boom-bust cycles [7] - Current earnings projections suggest that Korean equities are not overvalued, with the semiconductor sector experiencing significant growth, although reliance on these projections is crucial [8] Sector-Specific Insights - The tech sector is somewhat insulated from the immediate impacts of rising oil prices, but prolonged high prices could eventually affect the sector [9] - Supply chain risks, particularly concerning helium for semiconductor manufacturing, are emerging as a concern for the industry, necessitating greater independence from global supply chains [10][11] Long-Term Considerations - The macroeconomic impact of rising oil prices is immediate, but there are broader implications for various industries that need to be considered in the long term [12][13]

3家AI芯片上市公司进入华东大厂采购名单 寒武纪、海光、壁仞3家国产AI芯片上市公司进入华东互联网公司采购名单，虽然这三家的订单还未最终 落地，但数量都不小，据说单家采购超数万张，或订单金额高达数十亿人民币。 不过获得华东大厂采购并非是这几家国产卡产品力获得认可的"金牌认证"，消息人士表示，国产卡被采购 的一个条件是有明确的客户，华东大厂采购回来后再给明确的客户提供国产算力。国产芯片公司如何拿下 互联网大厂？更多信息交流添加微信 BENSONEIT 。 北京大厂低价"扫货"，H200集群预算压至5.1万 春节前后，某北京大厂正疯狂扫货服务器，"有多少拿多少"。有服务器厂商透露，其给出的H200集群租 金预算仅为5.1万元，远低于目前市面上6.3万-6.4万元的平均租金水平。 业内分析认为，这是该厂针对渠道方的定价策略，通过抛出极低预算、且在多个渠道放出消息，以此拉低 整体定价预期，在最终成交时获得价格优势。 与此同时，市场近期出现约200台规模的A100、4090集群需求，传闻买家实为OpenAI旗下的子公司。大 厂算力缺口到底有多大？欢迎添加 Ericazhao23 交流。 华北上市AI芯片公司拟绑定专属服 ...

Core Viewpoint - Himax Technologies (HIMX) shares experienced a significant surge following rumors that the company may become a supplier for NVIDIA (NVDA) and Apple [1] Group 1 - The recent rumors have led to a major outbreak in the stock price of Himax [1]

事实上，就在 Mac mini 被疯抢的同一周，阿里云、腾讯云、火山引擎、百度智能云几乎同时上线了 OpenClaw云端一键部署方案，最低 68 元一年就能开一台预装好的轻量应用服务器。云端的一台小虚拟机，同样可以是"小龙虾"的宿主。 OpenClaw是一个开源 AI 智能体平台，用户可以在本地部署一只始终在线的"AI 小龙虾"，接入微信、飞书、钉钉等入口，自动完成网页 浏览、文件操作、命令执行等任务。要跑这类应用，前提是有一台 24 小时开机、并具备系统级权限的设备。 Mac mini 很适合这个角色。它功耗低、体积小、运行安静，放在角落里就能长期在线。macOS 还原生集成了日历、备忘录、iMessage、 快捷指令等能力，Agent 获得权限后可以直接调用，这一点是 Windows 和 Linux 很难完全替代的。 还有一个现实原因：大多数用户并不放心把一个可能"犯错"的 AI 放在自己的主力电脑上。Mac mini 刚好可以作为一台独立、低成本 的"隔离沙箱"。 OpenClaw（俗称"龙虾"）的火爆意外引爆了Mac mini的市场。 但是Mac mini 的走红只是 OpenClaw 早期扩散的表象 ...

March 15, 2026 10:54 PM GMT China Equity Strategy | Asia Pacific M Idea US-China Presidential Summit: Scenario Framework & Investment Implications We see the upcoming US-China Presidential Summit as a catalyst-laden event, with the Iran situation complicating possible outcomes as well as potential equity market reactions. We expect a measured, index-level impact vs. 2025, and provide lists of stocks that could outperform indices in each case. With the US–China Presidential Summit set to take place against a ...

[Table_ReportDate] 2026 年 03 月 16 日 [Table_Author] 分析师：李时樟 执业证书：S1030522060001 电话：18065826333 邮箱：lisz@csco.com.cn 分析师：罗晴 执业证书：S1030524110001 电话：13603091122 邮箱：luoqing@csco.com.cn 公司具备证券投资咨询业务资格 证券研究报告 TMT [Table_Title] OpenClaw 政策出台与商业布局同步提速 TMT 行业周报（3 月第 2 周） [Table_S 行业观点： ummary] 请务必阅读文后重要声明及免责条款 4) 风险提示：AI Agent 等应用进展不及预期。 [Table_Industry] [Table_ReportType] 1) 周度市场回顾。TMT 板块内一级行业上周（3/9-3/13）涨跌幅 为：通信（-0.12%）、计算机（-0.92%）、电子（-1.23%）、 传媒（-3.23%），TMT 板块整体跑输沪深 300（0.19%）。板块 内涨幅靠前的三级子行业分别为通信线缆及配套（4.89%）、 印制电路板 ...

Summary of Key Points from the Conference Call Industry Overview - The report focuses on the **U.S. Internet & Semiconductors** industry, particularly the **AI sector** and its capital expenditure (capex) trends [1][2]. Core Insights 1. **AI Capex Peak**: The analysis predicts that AI capital expenditure will peak around **2028**, exceeding **$1 trillion**, which is approximately **$300 billion** above current consensus estimates [1][4]. 2. **AI Adoption Acceleration**: There is a notable acceleration in AI adoption, with leading labs reporting annual recurring revenues (ARRs) increasing by **20-35%** in early **2026** [2][3]. 3. **Recursive Self-Improvement (RSI)**: The emergence of RSI is expected to drive faster innovation, with AI models capable of self-optimization and continuous improvement starting in **2027** [3][12]. 4. **Compute Capacity Requirements**: By **2029**, AI labs will require approximately **23 GWs** of new compute capacity, with significant investments needed to support this demand [4][19]. Financial Projections 1. **OpenAI and Anthropic Growth**: OpenAI's ARR is projected to grow from **$6 billion** in early **2025** to **$25 billion** by early **2026**, while Anthropic's ARR is expected to rise from **$1 billion** to **$19 billion** [7]. 2. **Capex vs. Operating Cash Flow**: In **2028**, AI capex is anticipated to represent around **90%** of hyperscaler operating cash flow, with some companies exceeding **100%** [4][27]. 3. **Training Compute Expense**: The training compute expense for OpenAI and Anthropic is projected to peak at **$155 billion** in **2029**, indicating a significant increase in required resources [29][30]. Market Dynamics 1. **Hyperscaler Capex Underestimation**: The report suggests that the market has consistently underestimated the capex required for AI, with Barclays' estimates being significantly higher than Bloomberg consensus [20]. 2. **Inference vs. Training Compute**: The report anticipates a shift where inference compute will become the majority of compute requirements in the 2030s, as training compute is expected to decline post-2029 [40]. Additional Considerations 1. **AI Query Growth**: The number of AI queries is expected to increase significantly, with estimates of around **4 trillion** queries in **2027**, doubling to **9 trillion** by **2030** [45]. 2. **Agentic AI Workflows**: The transition to agentic AI is expected to disrupt various sectors, with AI systems becoming more autonomous and capable of handling complex tasks [12][15]. 3. **Caveats on Forecasts**: The report acknowledges potential discrepancies in compute forecasts between OpenAI and Anthropic, suggesting that these figures may converge over time [9][30]. Conclusion - The analysis indicates a robust growth trajectory for the AI sector, with significant capital investments required to meet the increasing demand for compute capacity. The anticipated advancements in AI capabilities, particularly through recursive self-improvement, are expected to further accelerate this growth, presenting both opportunities and challenges for investors and companies in the sector [1][3][4].

交换芯片系列专家会 20260315 摘要 2026-2027 年国内 Scale-up 交换芯片需求将从 10 万颗翻倍至 20 万 颗以上，Scale-out 市场则由 26 万颗增至 30-40 万颗。 2027 年市场空间预计较 2026 年增长 7-8 倍，主因是 51.2T 芯片单价 高达 5 万元（vs 25.6T 约 1.5 万元）且渗透率将超 70%。 盛科通信在 Scale-up 市场占据绝对主导，2026 年份额预计超 90%； 在 Scale-out 市场凭借 25.6T 先发优势，2026 年份额保底 20%。 阿里、字节、腾讯 2026 年超算节点规划明确，合计交换芯片需求超 10 万颗，大规模交付集中在 2026 年下半年。 2028 年全量国产化政策驱动明确，交换芯片位列信创清单首位，运营 商及国企市场 30%份额将由国产厂商刚性承接。 竞争格局方面，盛科作为独立第三方供应商，在下游整机厂（新华三、 锐捷等）的导入优先级高于中兴微电子等自研整机厂商。 Q&A 请问 2026 年和 2027 年，Scale-up 和 Scale-out 交换芯片市场的总体空 间分别是多少？其中， ...