岁末年初，中国算力产业在资本领域的运作骤然提速。 1月8日，上海天数智芯半导体股份有限公司（下称"天数智芯"，09903.HK）在港股挂牌上市，其此次 公开发售获超400倍认购，充分显示了资本市场的热情。 与GPU厂商的二级市场热潮相呼应，国产存储芯片厂商也在2025年末完成了关键一跃。2025年12月30 日，长鑫科技集团股份有限公司（以下简称"长鑫科技"）正式向上交所递交科创板招股书，并披露了其 2025年前三季度320.84亿元的营收数据，这一数字直观展示了国产DRAM（动态随机存取存储器）产能 释放的规模。 更早之前的2025年9月25日，长江存储科技控股有限责任公司（以下简称"长存集团"）完成了股份制改 革，其1600亿元的估值一度刷新了半导体独角兽的纪录，也标志着这家NAND Flash（闪存）龙头进入 了发展新阶段。 从芯片设计到存储颗粒，从科创板到港交所，一浪接一浪的资本热潮间接宣告了国产算力产业的快速进 步。当然，在资本市场的热度之外，国产芯片在实际的智算中心建设与应用适配中，仍面临着复杂考 验。 各司其职 此前不久，另外两家国产GPU头部企业也已接连登陆科创板：2025年12月5日，摩尔线 ...

Core Insights - Nvidia and Broadcom are identified as top AI stocks to consider for investment in 2026 and beyond, with a significant majority of investors planning to buy or hold AI stocks [1][2]. Nvidia - Nvidia's market capitalization stands at $4.6 trillion, with a current stock price of $187.47 and a gross margin of 70.05% [3][4]. - The company has experienced remarkable growth, with revenue increasing from $5.9 billion in fiscal Q3 2023 to $57 billion in fiscal Q3 2026 [3][4]. - Nvidia's GPUs are the primary drivers of AI workloads, and its data center networking revenue surged by 162% to $8.2 billion [4]. - The CUDA software platform has been pivotal in establishing Nvidia's dominance, as it has trained a generation of developers to optimize AI applications on its chips [5][6]. Broadcom - Broadcom has a market capitalization of $1.6 trillion, with a current stock price of $343.77 and a gross margin of 64.71% [7][8]. - The company specializes in data center networking and ASIC design, providing essential products for managing data flow in AI clusters [8][9]. - Broadcom is positioned to benefit from the growing demand for custom AI chips, with analysts predicting its AI revenue will increase from $20 billion in fiscal 2025 to over $50 billion in fiscal 2026, and $100 billion in fiscal 2027 [11]. - ASICs, while having high upfront costs and less flexibility than GPUs, are becoming a viable option for data center operators as inference workloads grow [10][12].

Core Viewpoint - Marvell Technology, Inc. has announced the acquisition of XConn Technologies for approximately $540 million, marking a strategic move to enhance its position in the AI infrastructure sector and accelerate its growth in data center connectivity solutions [2][3]. Group 1: Acquisition Details - The acquisition will be financed through a combination of 60% cash and 40% stock, with the stock price based on Marvell's weighted average price over the past 20 days, resulting in the issuance of about 2.5 million shares [3]. - This transaction is expected to significantly strengthen Marvell's Ultra Accelerator Link (UALink) team by integrating XConn's leading PCIe and CXL products, while also bringing in experienced engineers from XConn [3][5]. Group 2: Market Context and Strategic Importance - The demand for AI computing power is driving data centers to evolve from single-rack deployments to large-scale multi-rack configurations, necessitating high-bandwidth and ultra-low-latency networks [4]. - UALink, supported by major tech companies, aims to break NVIDIA's NVLink monopoly and is designed to connect up to 1,024 AI accelerators in a single cluster, positioning it as a key technology for next-generation AI infrastructure [4]. Group 3: Market Potential and Financial Impact - The global market for PCIe switches is projected to reach $2 billion by 2029, while CXL technology is expected to approach $16 billion by 2028, indicating significant growth opportunities for Marvell [5]. - Following the acquisition announcement, Marvell's stock price rose over 2%, reflecting positive market sentiment towards its strategic direction, especially after a challenging year where its stock fell over 23% due to competition [5][6]. - XConn's products are anticipated to contribute to revenue starting in the second half of fiscal year 2027, with projected revenues of approximately $100 million in fiscal year 2028 [5]. Group 4: Future Outlook - The transaction is expected to close in early 2026, pending regulatory approvals, and will solidify Marvell's core strategy of integrating UALink, optical interconnects, and switching chips [6]. - Marvell's ongoing strategic acquisitions are aimed at enhancing its competitive edge in the AI infrastructure market, positioning the company as a key player in reshaping industry dynamics [6].

Core Viewpoint - The stock market is heavily influenced by artificial intelligence (AI), with several key stocks expected to perform well in 2026, particularly Nvidia, Broadcom, and Taiwan Semiconductor Manufacturing [1]. Group 1: Nvidia - Nvidia is recognized as the leader in AI infrastructure, with its GPUs being essential for AI data center development [2]. - The company has established a robust ecosystem around its chips, having integrated its CUDA software platform into educational institutions, which has led to widespread adoption among developers [4]. - Nvidia's recent acquisition of SchedMD enhances its software capabilities, particularly with the open-source platform Slurm, and its NVLink interconnect system provides a competitive networking advantage [6]. Group 2: Broadcom - Broadcom is becoming a preferred choice for companies seeking cost-effective alternatives to Nvidia's GPUs, focusing on ASIC technology for custom AI chip design [7]. - The company has collaborated with Alphabet to develop tensor processing units (TPUs), attracting other major clients like Meta Platforms and OpenAI, which is expected to drive significant growth [9]. - Analysts project Broadcom's AI revenue to exceed $50 billion in fiscal 2026 and reach $100 billion in fiscal 2027, a substantial increase from $20.2 billion in fiscal 2025 [10]. Group 3: Taiwan Semiconductor Manufacturing - Taiwan Semiconductor Manufacturing Company (TSMC) is positioned to benefit from the rising demand for both GPUs and AI ASICs, holding a near monopoly in advanced logic chip manufacturing [12]. - TSMC is the only foundry capable of producing smaller node chips at high yields, which is critical for the development of powerful and energy-efficient chips [14]. - The company has demonstrated strong pricing power, increasing prices by over 15% since 2019 and planning further hikes starting in 2026 [16].

Co-Packaged Optics (CPO) Book – Scaling with Light for the Next Wave of Interconnect 共封装光学（CPO）⼿册：⽤光连接下⼀波互联浪 潮 Scale-out and Scale-up CPO, CPO TCO and Power Budgets, DSP Transceivers vs LPO vs NPO vs CPO, TSMC COUPE, MZM vs MRM vs EAM Modulator Deep Dive, CPO Focused companies and CPO Supply Chain 横向扩展与纵向扩展 CPO、CPO 总拥有成本与功耗预算、DSP 收发器对⽐ LPO/NPO/CPO、台积电 COUPE 平台、MZM 与 MRM 及 EAM 调制器深度解析、CPO 领域 重点企业与产业链 DYLAN PATEL, DANIEL NISHBALL, MYRON XIE, AND 3 OTHERS DYLAN PATEL, DANIEL NISHBALL, MYRON XIE 等 3 ⼈ JAN 02, 2 ...

前几天，一个炒股的朋友跑来问我："英伟达是不是不行了？像是四面楚歌了。谷歌有TPU，OpenAI有 Triton，华为有昇腾，AMD也在后面追着喊'兄弟，等等我'。" 我说你先别急着替英伟达写悼词。你这问题，暴露了一个常见毛病：很多人把"护城河"理解成知识产 权、渠道、品牌、规模、网络效应、用户认知——这些都对，但在硬科技领域，它们往往只是"表面那 层漆"。真正能挡人的，是一种不体面、没有浪漫气息、写出来甚至有点像在给加班洗白的东西：工程 积淀。 工程积淀是什么？ 它不是某个天才的灵光一现，更不是一纸专利。它更像是：你在一条路上走了二十年，沿途遇到几十万 个坑。每个坑都不大，但都很贵、很烦、很脏。你填完一个坑，前面又冒出一个坑。你一边填坑一边骂 街，一边还得保证车继续跑。最后这条路就成了高速公路——别人也想走，但他们要先把二十年的坑补 齐。 这类东西有两个特点：数量巨大，无法预测。 英伟达的护城河里，最硬的两块石头，大概是：CUDA和NVLink（以及它背后的整套集群网络能 力）。 先说CUDA。 CUDA最大的优势不是"先进"，而是"习惯"。习惯这东西很要命。你要让全世界工程师换生态，就像你 要让全世界改用 ...

Both companies dominate their specific niches in the AI value chain.Nvidia (NVDA 0.26%) and Taiwan Semiconductor Manufacturing (TSM 0.45%) are key players at different points along the artificial intelligence (AI) value chain. While Nvidia is the leading designer of the graphics processing units (GPUs) used to handle most AI workloads, TSMC is the global leader in fabricating and packaging advanced logic and memory chips.Both have benefited from the AI megatrend -- but which of these companies looks like a ...

公众号记得加星标⭐️，第一时间看推送不会错过。 9 月份，英伟达首席执行官黄仁勋和英特尔首席执行官陈立武达成协议，英伟达锁定了每股 23.28 美 元的英特尔收购价格。 该交易一直受到美国联邦贸易委员会的审查，该委员会正在调查英伟达可能持有的 4% 股份是否违反 了反垄断法。 然而，联邦贸易委员会于 12 月 18 日批准了这项交易。 根据英特尔提交的监管文件，此次2.14亿股股票的收购已于12月26日完成。英特尔股票周一收盘价为 每股36.68美元。 根据协议条款，英伟达和英特尔将联合开发"多代"用于数据中心和个人电脑的芯片，以期从消费者到 超大规模客户，在整个芯片客户群体中抢占市场份额。 这两家公司将致力于通过速度极快的 NVLink 连接他们的芯片，NVLink 每个 GPU 的带宽可达 1.8 TB/s（每个方向 900 GB/s），约为 PCIe 5.0 x16 插槽带宽的 14 倍。 在个人电脑领域，英特尔将为英伟达定制 x86 CPU，英伟达将把这些 CPU 集成到其 AI 基础设施平 台中，并推向市场。 英特尔还将能够制造集成英伟达RTX GPU芯片的x86系统级芯片（SoC）。这些新型x86 ...

Nvidia's operational dominance has translated into a stock story marked by sharp runs, healthy pauses, and strong long-term structure. NVDA has rallied aggressively through 2025, only stopping periodically to consolidate gains. After peaking near $212.19 in late October, the stock is down by 11.67%. Even after that dip, shares of the chip giant remain firmly in an uptrend, up over 41.9% year-to-date (YTD), and 22.9% over the past six months.Its CUDA platform entrenched Nvidia deeply within developer workflo ...

公众号记得加星标⭐️，第一时间看推送不会错过。 历经大约1年的规格制定工作后，由多家芯片、网通与云端服务大厂组成的UALink联盟，于2025年4 月发表UALink（Ultra Accelerator Link）1.0版规范，让这项讨论已久的开放式加速器芯片互连I/O 架构，终于正式问世。 UALink联盟成立于去年5月，成员包括AMD与Intel两大芯片厂商，加上Broadcom、Cisco两大网通 厂商，还有Google、HPE、Meta与微软等云端大厂，阿里云、苹果与Synopsys也于2025年1月加 入，理事会加上贡献者成员超过65家（但Broadcom已退出UALink理事会，降为一般贡献者成员）。 联盟的宗旨是建立一个开放的GPU加速器互连I/O规范，为AI伺服器与丛集中的加速器，提供高速、 低延迟的I/O连结架构，以对抗Nvidia的NVLink技术。 不过，在GPU互连规模方面，则是UALink 1.0占上风，允许最多1,024个GPU互连。而NVLink搭配 NVLink Switch，允许直连的GPU数量为576个。 Nvidia当前在GPU加速器市场上的「霸业」，不仅是建立在GPU本 ...