2、50 倍全球产能有多难？ Terafab 计划落地奥斯汀，号称要在一个建筑里整合光刻、芯片生产、封装测试全流程，实现 AI 计 算全环节自主制造，还直奔 2 纳米先进制程，主打极致闭环、快速迭代。但这份野心，在产业数据 面前显得无比沉重。 周末马斯克宣布特斯拉、SpaceX、xAI 联手 Terafab 超级芯片工厂计划，目标年产能 1 太瓦算力芯 片，直接干到当前全球总产能的 50 倍，80% 产能送上太空，20% 供给特斯拉机器人与汽车。 这篇文章聊几个关键的问题。 1、马斯克为何非要自己造芯片？ 首先，马斯克为何非要自己造芯片？--还是算力太缺了。 当前全球芯片年产能约 20 吉瓦，仅能满足他未来需求的 2%，光是 Optimus 人形机器人，远期年产 10 亿台的目标就需要 100-200 GW算力芯片，更别说 xAI 的超大规模 AI 训练、SpaceX 的太空数据中 心，每一个都是算力吞金兽。 在马斯克看来，地球的算力天花板早已触顶：美国全国电网总容量才 0.5 太瓦，根本撑不起海量 AI、机器人同时运行；而且地面数据中心选址难、散热成本高，反观太空，太阳能效率是地面 5 倍，真空环境零散热成 ...

这段时间，由于各种因素，市场确实很难做，但Memory（主要是海外）和光却都表现的不错，OFC 期间，各家公司都展示对未来乐观的预期。这篇文章我们聊一下本周讨论度非常高的XPO，解析一 下Arista XPO的White Paper，看下有哪些市场可能忽略的一些细节。 "光"话题的讨论向来是比较激烈，海外某个大V因为写了CPO进展超预期的文章，被骂的已经删除了 自己的专栏。笔者之前因为转了一篇citi对于CPO超预期看法的会议内容，也引来了很多骂声。 笔者无意参战XPO/CPO什么时候到来的讨论，下面的内容大部分都是Arista XPO的内容解析，有些 是笔者结合自己的行业调研加入的，但不包含XPO/CPO等的时间进度。 一、AI 浪潮下的基础设施困局 当 AI 数据中心的规模突破 10 万 GPU，当单个训练任务的成本攀升至数百万美元，当传统网络架构 在极限带宽需求面前捉襟见肘，我们意识到，是时候重新定义数据中心光学互连了。这个白皮书是今 年3月份，Arista Networks 联合 45+产业伙伴发布的。 白皮书开篇就直指 AI 网络面临的五个关键痛点： OSFP 的极限 说到这里，就不得不提业界现在普 ...

"一个 130 亿参数的模型微调， 8 卡机器要跑一周，稍微把 batch size 调大一点就直接 OOM 。"最近，某 AI 创业公司算法负责人陷入两难，往上走， 切换大规模算力集群预算吓人，公司账上现金流撑不住；往下将就， 8 卡服务器勉强能跑，但永远在超载的边缘试探。 这不是个例。在国产超节点市场，一个尴尬的断层正在上演：一端是 8 卡算力的"入门级玩具"，另一端是数百卡集群的"天价巨兽"，最广阔的"中产阶 层"算力需求，却长期处于真空地带。 算力断层带来的直接后果就是，大量真正有业务场景、有落地需求的企业，要么被迫降级妥协，牺牲效率；要么被迫超前消费，背上沉重的算力成本包 袱。算力市场的供需错配，正在拖慢 AI 产业化的脚步。 8 卡的局限：当"入门级"成为"瓶颈级" 过去几年里， 8 卡服务器确实为 AI 普及立下汗马功劳。它门槛低、部署快，是无数算法工程师的"启蒙机器"。直到今天，对于小规模、轻量级推理场 景， 8 卡依然是够用的选择。 但问题在于， AI 产业的演进速度远超硬件迭代的预期。 当千亿参数大模型成为行业标配，更复杂的 MoE 架构成为主流，企业开始真正跑商业化的微调和推理任务时， ...

Core Insights - The article discusses the transformative impact of artificial intelligence (AI) on the chip design industry, driven by the surge in demand for edge devices such as smart cars, smart homes, industrial robots, and wearables. This growth presents challenges including increased design complexity, high upfront costs, and intricate business processes [1] - The introduction of Arm's technology licensing subscription model offers a new solution to these challenges, enabling companies of various sizes and stages to participate in chip innovation during the AI era [1][3] Summary by Sections Arm Technology Licensing Subscription Model - The model is not merely an upgrade of traditional IP licensing but a comprehensive system designed to meet the diverse needs of the chip design industry, allowing for flexible access to technology [2][3] Arm Total Access - Arm Total Access provides a comprehensive suite of Arm IP products for large enterprises that need to develop complex systems and multiple projects. It includes three tiers of packages based on performance requirements: - Package A (Low Power) includes IPs like Cortex-A53, Cortex-M85, Ethos-U65, Mali-G310, CoreLink, and CoreSight [4] - Package B (High Efficiency) includes IPs such as Cortex-A520, Mali-G625, and Cortex-A55, along with Package A [5] - Package C (High Performance) includes IPs like Cortex-A7x, Neoverse-N, and Mali-G7x, along with Packages A and B [6] Arm Academic Access - Arm Academic Access is a free program for academic institutions, allowing them to use Arm technology for education, training, and non-commercial research. It includes access to over 70 CPU, GPU, and NPU IPs, as well as unlimited free tape-out opportunities (up to 1000 chips per tape-out) [7] Industry Significance - The subscription model represents an innovative business strategy and an upgrade in the chip design industry paradigm. It aligns with the trends of diversification in edge devices, rapid technological iteration, and personalized market demands. This model serves as an "accelerator" for chip innovation, enabling companies to better address design challenges and focus on core differentiation [8]

以下文章来源于More Than Semi ，作者mofan More Than Semi . More Than SEMI 半导体行业研究 今年三月的洛杉矶，OFC 大会再次刷新了规模纪录。走在展厅里，你能明显感受到一种兴奋的氛围 ——光通信技术正在从幕后走向台前，成为 AI 数据中心架构中不可或缺的核心要素。 Coherent 的 CEO Jim Anderson 在会上开门见山地说了一句话："我觉得现在可能是光通信行业有史 以来最好的时候。"这不是客套话，从他们展示的技术路线图和产品时间表来看，这家公司确实站在了 一个关键的转折点上。 一个 500 亿美元的基本盘，加上 200 亿美元的新故事 Coherent 现在的生意可以分成两部分来看。第一部分是他们已经在做的事情——插拔式光模块、DCI 收发器、传输设备，这些加起来大概是个 500 亿美元的市场。这块业务很稳，也在持续增长，但真正 让人兴奋的是第二部分。 他们押注了四个新方向，每一个都不是 PPT 上的概念，而是今年下半年到明年就能看到真金白银的产 品。这四个方向加起来，能再带来 200 多亿美元的增量市场。更关键的是，这些新业务的毛利率都比 传 ...

老黄演讲后，网上很多博主都发了关于GTC的内容，但是绝大多数都是新闻性质的，他们只讲了黄仁勋都说了啥。这篇文章我们结合 NVIDIA 的技术规 划，来聊一聊网上可能没有的分析和GTC的核心看点。对于万亿营收和CPO的信息，都是大家知道的了，这篇文章就不再赘述了。 1、CPX的黯然退场 在上个月下旬的时候，网上还依然传着很多关于 CPX 要使用 HBM 的传言。我当时就听到消息说 CPX 要取消，我当时还不太信，但事实证明确实是被 LPU 取代了。 那为什么会取代呢？就是因为英伟达的路线转换，他们要从 prefill 加速切换到 推理加速。 这些内容和观点，我们之前在星球中都讲过。 2、谁是这次 GTC 大会之后最大的受益方？ 那显然是三星。因为 LPU 是找三星独家代工，采用的是三星的 N4 工艺。这意味着三星不仅是英伟达全层级存储的核心供应商，现在更独揽了 LPU 的 代工大单。 在Rubin上，三星的综合价值量已经超过台积电了，因为台积电只负责代工加封装。 原文链接：https://globalsemiresearch.substack.com/p/nvidia-gtc-2026-is-samsung-t ...

Core Viewpoint - The AI industry is entering a "chip shortage era," which is expected to last until at least 2027, highlighting the importance of supply chain management alongside technological capabilities [37]. Group 1: AI Chip Demand and Supply - Anthropic generated an additional $6 billion in annual recurring revenue in just one month, primarily through its AI programming tool, Claude Code [4]. - The demand for AI chips, particularly those using TSMC's 3nm process, is expected to consume nearly 60% of TSMC's 3nm capacity this year, rising to 86% next year, squeezing out traditional mobile chip customers [11][12]. - TSMC's 3nm capacity is under pressure as major AI chip manufacturers like NVIDIA, AMD, Google, and AWS are all vying for this advanced process technology [8][9]. Group 2: Supply Chain Dynamics - NVIDIA has strategically locked in supplies of logic wafers and memory components, positioning itself as a major beneficiary in the ongoing supply chain competition [33][34]. - The shift in focus from power supply to silicon wafer availability indicates that while data centers and power supply have expanded, the chip supply has not kept pace [28][32]. - The production of high-bandwidth memory (HBM) is also facing challenges, as HBM consumes 3 to 4 times the wafer capacity compared to standard DDR memory, exacerbating the supply constraints [17][22]. Group 3: Market Implications - The competition for chip resources is leading to a "reallocation of bits," where AI applications are prioritized over consumer electronics, potentially resulting in higher prices and slower product cycles for smartphones and PCs [23][38]. - The pricing dynamics for HBM are shifting, with DDR memory prices rising, which may reduce the incentive for manufacturers to shift production capacity from DDR to HBM [22]. - The AI industry's rapid growth is outpacing hardware supply capabilities, leading to a scenario where access to chips becomes a critical factor for success in AI deployment [38]. Group 4: Future Outlook - TSMC's role has become increasingly pivotal, as its capacity allocation decisions directly impact the competitiveness of major players like NVIDIA, Google, and AMD [38]. - The ongoing competition for silicon resources may lead to a significant transformation in the AI landscape, where the ability to secure chips becomes more crucial than algorithmic advancements [38]. - The consumer electronics sector may face significant challenges as AI demand continues to dominate chip production, potentially leading to a decline in smartphone demand and increased costs for consumers [38].

Core Viewpoint - The article discusses the upcoming "5th Advanced Thermal Management Technology Summit and AI and AI Intelligent Agent Developer Forum" scheduled for March 23-24, 2026, in Shanghai, China, focusing on opportunities and challenges in thermal management technology across various sectors, including AI servers, automotive electronics, humanoid robots, and high-performance chips [3][4]. Group 1: Event Overview - The summit will feature over 70 speakers and 600 industry experts, covering key topics such as AI server data centers, automotive electronics, humanoid robots, and smart terminals [4]. - The event will include a main conference and four specialized sessions, focusing on high-performance and high-power chip thermal management technology [4]. Group 2: Key Topics and Technologies - Key technologies discussed will include AI and AI intelligent agents, AI server data centers, smart terminals, humanoid robots, high-performance and high-power chips, high thermal conductivity materials, and microchannel cooling technology [4]. - Specific sessions will address liquid cooling technologies, embedded microfluidic cooling, and thermal management solutions for AI computing infrastructure [5][6]. Group 3: Participating Companies and Presentations - Notable companies participating include ZTE Corporation, Lenovo, and Ansys, presenting on topics such as high-power liquid cooling technology and thermal simulation solutions [5][15]. - Various universities and research institutions, including Peking University and Xiamen University, will also present their findings on advanced thermal management technologies [6][17]. Group 4: Industry Challenges and Innovations - The summit will explore challenges in thermal management for electric vehicle motor controllers and the principles of automotive thermal management design [8][23]. - Innovations in thermal management for humanoid robots will be discussed, focusing on key difficulties such as high power density and dynamic coupling effects [7][22].

Core Viewpoint - The article discusses the significant power shortage in the U.S. driven by the rapid expansion of data centers and the slow pace of infrastructure development, predicting a power gap of approximately 45 GW by 2028 and potentially 68 GW by 2029 [2][4]. Group 1: Power Shortage Dynamics - The power shortage in the U.S. is not just a challenge of energy quantity but also a competition for "Time to Power," with new data centers facing delays of up to 5 years due to labor shortages and regulatory processes [4]. - The ongoing U.S.-Iran conflict has disrupted LNG exports, affecting global supply and leading to increased competition for North American resources, which may raise wholesale electricity prices in gas-dominated regions [5]. Group 2: Solutions to Power Shortage - The market is shifting towards "stock game" strategies to address the power gap, as new grid construction is lagging behind demand [6]. - The self-generation (BYOG) model is emerging as a key solution for AI data centers, allowing for faster deployment and reduced reliance on public grids, cutting the setup time from nearly 5 years to just a few months [7]. Group 3: Bloom Energy's Role - Bloom Energy's solid oxide fuel cell (SOFC) technology aligns well with the BYOG model, offering rapid deployment and compliance advantages, with delivery times of 3-4 months [8]. - Bloom Energy holds a dominant market share in the SOFC sector, with 75%-85% of the global market and 65% in the data center backup power segment, positioning it as a critical provider for addressing power shortages [10]. Group 4: Orders and Capacity - Bloom Energy's total backlog reached $20 billion by the end of 2025, a 65% increase from 2024, indicating strong future performance [10]. - The company has secured significant contracts, including a $5 billion partnership with Brookfield for AI infrastructure and a $2.65 billion order from AEP for Amazon's power needs [10][11]. Group 5: Supply Chain Insights - Key suppliers for Bloom Energy include Chunhui Instrument, which provides temperature sensors, and Sanhuan Group, which supplies SOFC membrane materials, both of which are expected to benefit from the growing demand for SOFC systems [14][15].

Core Viewpoint - The AI optical network market is projected to reach $73 billion by 2030, accounting for nearly 40% of the entire AI network market [1] Group 1: Importance of Optical Technology - The traditional data transmission methods in data centers are copper cables and optical fibers, with copper being preferred for its cost-effectiveness and reliability [1] - AI has changed the requirements for data transmission, necessitating the use of optical technology due to distance, bandwidth, and power consumption challenges that copper cables cannot meet [2][3] - The trend indicates that optical technology will not only maintain its existing market share but will also expand into areas previously dominated by copper cables [3] Group 2: Optical Transceivers - Optical transceivers, the most common optical interconnect technology, convert optical signals to electrical signals and vice versa [6] - Despite concerns about high power consumption, analysts predict that the optical transceiver market will grow from $12.6 billion in 2025 to $45.4 billion by 2030, with a compound annual growth rate of 29% [7][8] - Factors contributing to this optimism include a changing demand cycle, tight production capacity, and significant technical advantages of optical modules [8][9] Group 3: Laser Technologies - The core components of optical modules are lasers, with three main technology routes: VCSEL, EML, and CW lasers, each serving different market needs [12][13] - VCSEL is the main player for 800G applications, while EML is crucial for high-end markets, and CW lasers are gaining traction due to their cost-effectiveness and scalability [12][13] Group 4: LPO/LRO Technologies - LPO and LRO are transitional technologies that reduce power consumption by eliminating or partially removing the DSP from optical modules [14][15] - These technologies are expected to capture a niche market, particularly in the 1.6T segment, with a projected market size of $5 billion by 2030 [15] Group 5: Co-Packaged Optics (CPO) - CPO represents a significant shift by integrating optical devices directly into the switch chip packaging, reducing signal loss and power consumption [17][20] - The market for CPO optical devices is expected to reach $1.5 billion by 2030, with the overall CPO market, including switches, projected to be $24.4 billion [28][29] - CPO is anticipated to open a new market for Scale-up networks, which is expected to reach $9.4 billion by 2030, surpassing the Scale-out market [29] Group 6: Optical Circuit Switches (OCS) - OCS technology is projected to grow from $1 billion to $4 billion by 2030, addressing the limitations of traditional circuit-switched networks [33] - OCS offers lower latency, reduced power consumption, and greater flexibility, making it suitable for large AI clusters [36][40] Group 7: Investment Opportunities - Key companies recommended for investment include Marvell, Lumentum, Coherent, Nvidia, and Broadcom, each with strong positions in the optical technology market [44] - The report emphasizes that the market is expected to benefit all players due to strong demand and the time required for capacity expansion [45]