2026 年，AI 数据中心发展进入了一个新的瓶颈阶段。 单纯的算力堆砌似乎已经走到了尽头，随着代理式 AI 的兴起、混合专家模型（MoE）的普及，以及百万 token 级长上下文逐渐成为常态，大型云 厂商开始意识到，竞争的关键不再是谁拥有更多 GPU，而是谁能更高效地使用，谁能把系统整体效率真正拉起来。 但传统数据中心的短板已经暴露无遗：当代理式 AI 需要处理多轮复杂对话时，它必须频繁调用不同的专家模型，同时保持对历史上下文的准确记 忆。在这个过程中，GPU 可能大部分时间处于等待状态——等待数据从内存传输过来，等待网络将其他节点的计算结果送达，等待存储系统调出历 史记忆。 正是在这样的背景下，英伟达在 CES 2026 上正式发布了 Rubin 平台和 BlueField-4。它们并非是一次硬件的常规升级，而是对传统数据中心低效 这一症结开出的新处方，从某种意义上说，英伟达"重做了一次数据中心"，构建起了真正面向 AI 原生时代的硬件底座。 我们先从一切计算工作的核心——Rubin 平台说起。 与以往围绕单一核心器件不断迭代的路径不同，Rubin 最显著的变化在于，它从一开始就放弃了"单点升级"的思路， ...

公众号记得加星标⭐️，第一时间看推送不会错过。 半导体市场研究人员对人工智能在2026年对芯片供应的影响存在分歧。虽然所有人都认为市场将强劲 增长，但一些人预测市场已经发生了根本性的变化，并且到今年年底市场规模将超过1万亿美元。 据 Future Horizons 的首席研究员 Malcolm Penn 称，这是不可能的，因为产能不足，而且存在一些 根本性的弱点。他刚刚发布了今年的预测。 然而，人工智能数据中心建设方面的支出也推动了其他市场分析。这推高了对英伟达（Nvidia）以及 AMD等厂商的图形处理器（GPU）和高带宽内存（HBM）的需求，同时也推高了价格。 美国市场研究公司Creative Strategies董事长本·巴贾林表示："现在的情况截然不同。半导体行业将 永远改变。该行业的格局已经彻底改变。" "人工智能基础设施建设代表着半导体行业有史以来规模最大的潜在市场扩张——这是一个千兆级的 增长周期，无论从绝对美元数额还是从其对价值链各个环节的影响范围来看，都远远超过了以往的增 长时期。 "这些数字讲述了一个前所未有的规模的故事。全球半导体收入将从2024年的约6500亿美元攀升至本 十年末的1万 ...

Core Viewpoint - Vietnam is making a significant move in the semiconductor industry with the groundbreaking of its first semiconductor chip manufacturing plant, led by Viettel, signaling its commitment to becoming a key player in the global semiconductor landscape [1][18]. Group 1: Vietnam's Semiconductor Strategy - The Vietnamese government aims to establish 100 design companies, one wafer manufacturing plant, and about 10 packaging and testing factories by 2030, targeting an industry revenue of $25 billion and increasing local value addition to 10-15% [1][4]. - The semiconductor strategy is part of a broader national development plan, with a phased approach to attract foreign direct investment and build a self-sufficient ecosystem [4][6]. Group 2: Market Dynamics and Opportunities - Vietnam is positioned as a low-cost, low-friction location for multinational companies to diversify their supply chains amid global geopolitical tensions, with over 170 foreign investment projects in the semiconductor sector totaling nearly $11.6 billion [5]. - The semiconductor market in Vietnam is projected to grow from approximately $17 billion in 2023 to $31.28 billion by 2027, with a compound annual growth rate (CAGR) of about 11.6% [13]. Group 3: Local Industry Development - The semiconductor ecosystem in Vietnam is primarily driven by foreign companies, with local firms like FPT and Viettel emerging as key players in chip design and manufacturing [8][19]. - Viettel's wafer plant is expected to fill the gap in Vietnam's semiconductor production capabilities, allowing the country to handle five out of six major production stages [19]. Group 4: Challenges and Infrastructure - Vietnam faces challenges in building a robust local ecosystem, with a heavy reliance on foreign investment and a limited number of local suppliers and research institutions [22]. - There is a significant talent gap in the semiconductor sector, with current engineering personnel around 6,000, while future demand is estimated to exceed 20,000 [23]. - Infrastructure issues, particularly in power supply, pose risks for semiconductor manufacturing, necessitating improvements in energy security and reliability [24].

Core Insights - TSMC's 2nm process is in high demand, with estimated tape-out volumes 1.5 times that of the 3nm process, attracting companies like Apple, Qualcomm, and MediaTek to secure supply for competitive advantage this year [1] - Despite the advancements in chip technology, consumer interest in smaller process nodes is waning, leading companies to pivot towards architecture improvements and increased memory cache as key strategies [1][2] - Apple has reportedly secured over half of TSMC's initial 2nm capacity, while Qualcomm and MediaTek are also vying for the enhanced 2nm "N2P" process to gain an edge in wafer shipments and CPU frequency [1][4] Group 1 - TSMC is expected to cancel Apple's priority status as its largest customer due to the rise of AI, with TSMC's revenue increasingly coming from different sectors [3][4] - Analysts note that while flagship devices drive industry growth, consumers are now more focused on actual user experience rather than just annual specification upgrades [2] - TSMC is facing supply constraints for 2nm wafers, leading to price increases for advanced process technologies starting in 2026, with the estimated price for Apple's A20 SoC being $280 [6] Group 2 - Apple's A19 Pro chip showcased significant architecture improvements, achieving a performance increase of 29% with minimal power consumption [2] - The competitive landscape is shifting, with companies like MediaTek's Dimensity 9500s surpassing Snapdragon 8 Gen 5 by utilizing a 19MB CPU cache [2] - TSMC's CEO has reportedly informed Apple of a significant price increase, marking the largest in recent years, indicating a shift in the dynamics of their partnership [4]

Core Viewpoint - Intel's investment in Ohio for two advanced chip factories is part of its strategy to build a world-class foundry service, but the project has faced delays and skepticism regarding its viability and customer acquisition [1][2]. Group 1: Intel's Chip Manufacturing Plans - Intel announced a $28 billion investment to build two advanced chip factories in Ohio, with initial production planned for 2025, but has faced multiple delays due to challenges in securing external customers [1]. - The first Ohio factory is now expected to start production in 2030, and the success of Intel's 14A process node is contingent on acquiring a significant number of external clients [1][2]. - Intel's 18A process node is currently in production in Arizona, but initial yields have been problematic, although recent reports indicate yields have improved to over 60% [1][2]. Group 2: Progress and Future Prospects - Recent hiring announcements related to the Ohio factory construction suggest that progress may be accelerating, which could indicate a renewed commitment to the 14A process [2]. - Intel's CEO expressed confidence in the 14A process, stating that significant advancements in yield and intellectual property are expected, which may help attract external customers [2]. - The potential for Intel's 14A chips to be used by clients like Apple in 2029 could enhance the likelihood of success for this process node [3]. Group 3: Competitive Landscape - The ongoing shortage of advanced manufacturing capacity at TSMC may provide Intel with an opportunity to attract customers seeking alternative suppliers [3]. - Analysts suggest that while Intel's 18A process may not directly challenge TSMC's leadership, it could position Intel to surpass Samsung and become the second-largest foundry [6][7]. - Companies like Qualcomm and AMD are reportedly considering Samsung as a primary alternative for foundry services, indicating a shift in the competitive landscape [5][6].

虽然MB.Drive Assist Pro无法实现Drive Pilot曾经承诺的完全自动驾驶功能，但梅赛德斯-奔驰将其 视为通往未来可全球推广系统的桥梁。"我们不想提供一个对客户而言益处甚微的系统，而且我们知 道，未来两三年内将会推出一个能为客户带来更多益处的系统，"公司发言人托比亚斯·穆勒告诉The Verge。 梅赛德斯-奔驰暂停了其Drive Pilot自动驾驶辅助系统项目，由于成本不断攀升、可用性有限以及供 应商格局变化，该公司正在调整其自动驾驶发展路线。该系统是美国首款也是目前唯一一款获得L3 级认证的自动驾驶产品，能够实现真正的"手不离手、眼不看"驾驶，但仅限于特定条件。Drive Pilot 系统原计划于2023年搭载于EQS电动轿车和S级旗舰车型上，但预计本月发布的新款S级轿车将不会 配备该系统。 （来源 ： 编译自 techspot ） 公众号记得加星标⭐️，第一时间看推送不会错过。 自动驾驶仍然是梅赛德斯-奔驰的长期目标，但监管框架不断限制着"无人值守"系统的运行方式和范 围。目前，该公司似乎专注于完善能够在这些限制条件下可靠运行的解决方案，而不是过早地推进完 全自动化。 该系统在汽车行 ...

Core Viewpoint - The article discusses the rapid development and commercialization of High Bandwidth Flash (HBF) technology, which is expected to surpass High Bandwidth Memory (HBM) by 2038 due to the increasing demand for high-capacity storage solutions driven by AI data needs [1][5]. Group 1: HBF Technology Development - HBF technology is anticipated to be implemented in products by major companies like Samsung and SanDisk by the end of 2028, leveraging the design and process technologies developed during HBM research [2][3]. - HBF is designed to stack multiple 3D NAND flash memory vertically, similar to HBM, to enhance bandwidth, with predictions of exceeding 1638 GB/s in bandwidth [3]. Group 2: Market and Capacity Insights - HBF's capacity is projected to be 512 GB, significantly higher than HBM4's 64 GB, addressing the limitations of HBM in meeting the explosive growth of AI data requirements [3]. - The collaboration between Samsung, SK Hynix, and SanDisk aims to standardize HBF technology, with product development targeted for release by 2027 [3]. Group 3: HBF vs HBM - HBF is expected to complement HBM in AI accelerators, with a larger capacity but slower speed, making it suitable for different types of data processing tasks [4]. - The transition to HBM6 will involve a network of HBM units, while HBF will provide a more extensive data storage solution, potentially changing data processing pathways in GPUs [5]. Group 4: Future Predictions - By 2038, the market for HBF is predicted to exceed that of HBM, indicating a significant shift in memory technology preferences within the industry [5].

Core Viewpoint - NVIDIA is strengthening its data center layout and advancing personal and edge AI computing platforms, with a roadmap from N1 and N1X to the next-generation N2 and N2X series for the AI era [1] Group 1: Product Development and Market Strategy - The N1X architecture-based Windows on Arm (WoA) platform is expected to launch in Q1 2026, targeting the consumer market, with three additional versions set for Q2 2026 [1] - NVIDIA's DGX Spark, the world's smallest AI desktop computer, has been released, while the launch of laptops with N1X/N1 platforms has been delayed to 2026 due to various factors [2][3] - The N1 and N1X platforms are positioned as high-end PC platforms, with N1X offering superior computing capabilities aimed at high-performance AI PCs and professional markets [3] Group 2: Ecosystem and Collaboration - NVIDIA is not replicating Intel's long-standing PCL ecosystem but is instead using its FAE team to create reference design documents for OEMs and ODMs [3] - NVIDIA plans to introduce an AVL (Approved Vendor List) and RVL (Recommended Vendor List) system to accelerate ecosystem development [3] - NVIDIA has invested $5 billion in Intel, with both companies collaborating on server and PC platforms, where Intel will customize x86 processors for NVIDIA's AI infrastructure [4][5]

Group 1 - Intel's Ohio One project, initially touted as the largest wafer factory in the U.S., has faced delays and scale reductions but remains active, with recent contractor job postings indicating acceleration in progress [1] - The Ohio One facility is designed to host up to eight wafer fabs, with the first two expected to begin production by 2030-2031, significantly delayed from the original 2025 target [1] - The first wafer fab will be the 1st fab, followed by the 2nd fab expected to start production a year later, coinciding with Intel's development of its next-generation 14A process [1] Group 2 - Intel's latest Panther Lake mobile CPUs are produced in Arizona and Oregon using the new 18A process, marking a significant advancement for domestic chip manufacturing, although there are no large external customers for this process [2] - CEO Pat Gelsinger has shifted from a pessimistic view of the 18A process to a more optimistic stance recently, despite the 14A process now expected to reach mass production in 2027 [2] - The Ohio One site was initially predicted to be the birthplace of the 14A process, but this timeline has shifted, with derivative products already occupying a significant portion of Intel's lineup before the site's official opening [2] Group 3 - The U.S. government allocated $8.9 billion from the CHIPS Act to Intel in exchange for a 10% equity stake, alongside a $5 billion collaboration with NVIDIA, signaling a resurgence in Intel's competitiveness, particularly in the foundry business [3] - Intel is accelerating the construction of the Ohio One facility after years of stagnation, but the timeline for its completion remains uncertain [3] Group 4 - TSMC is facing a "happy dilemma" with its 3nm process capacity fully booked until 2027, prompting significant increases in capital expenditure plans [5] - This capacity constraint is reshaping the market dynamics, leading TSMC's top clients to consider alternatives like Samsung and Intel [5] - TSMC's capital expenditure for 2026 is projected to be between $52 billion and $56 billion, exceeding previous expectations due to the overwhelming demand for advanced nodes [5] Group 5 - The supply-demand imbalance is causing market oversupply, with major clients like Apple, NVIDIA, AMD, Broadcom, Qualcomm, and MediaTek seeking alternative capacities as TSMC's advanced process node market share is expected to drop from 95% to 90% [6] - TSMC's 3nm monthly production capacity is planned to expand to 190,000 wafers by the end of 2026, but this will still fall short of client demand [6] - TSMC is adopting a more aggressive strategy by delaying new 3nm process development and encouraging clients to shift products originally slated for 2027/2028 to the 2nm GAA process [6] Group 6 - Samsung's Taylor wafer fab is seen as a more likely alternative for clients seeking supply options compared to Intel, with Qualcomm and AMD being the most likely to consider Samsung [7] - Discussions have indicated that Apple and Broadcom are evaluating Intel, but significant work remains for Intel's 14A process to be competitive [7]

消息人士称，台积电此举也是为了消除外界对其可能变成"美国台积电（ASMC）"的担忧，因为该公 司近期在美国的工厂扩张削弱了其"硅盾"。 公众号记得加星标⭐️，第一时间看推送不会错过。 包括《自由时报》在内的台湾媒体19日报道称，全球最大的晶圆代工厂（半导体代工制造）台积电计 划再建四座最先进的封装（AP）工厂。 据消息人士透露，台积电高级副总裁兼副联席首席运营官侯永庆将于22日宣布在台南地区扩建四座更 先进的AP工厂，其中包括嘉义科学园和南方科学园。 消息人士解释说，台积电计划于今年上半年在紫怡科技园的AP工厂1（P2）开始量产，并将设备运至 工厂2（P2）。 他还补充说，采用先进封装技术"芯片封装在晶圆基板上（CoWos）"的生产已在AP8工厂开始，该工 厂是群创光电（Innolux）的翻新工厂，群创光电是台湾富士康集团旗下的面板制造商，于2024年被 收购。 台积电还宣布计划在翟氏科技园和南方科技园各扩建两座AP工厂，以解决CoWos产能不足的问题。 这一特殊制造环节目前是全球AI GPU供应的主要瓶颈，决定着科技巨头构建下一代"超级智能"集群 的速度。由于台积电的CoWoS生产线已基本售罄至年底，而"热 ...