马斯克的 TeraFab 计划，凸显出人工智能及电动汽车企业在保障芯片长期供应方面面临的压力正 不断加大。但从黄仁勋的表态来看，要复制全球顶尖的半导体制造能力，绝非短期建设就能完 成，而是一项需要跨越数代人努力的艰巨挑战。 （来源 ：半导体芯闻综合 ） 2026 年 1 月 30 日，黄仁勋发表讲话时，针对马斯克提出的建造并运营一座名为 "TeraFab" 的 超大规模晶圆厂计划作出了上述评论。该晶圆厂旨在为特斯拉的人工智能芯片与车载芯片提供长 期稳定的供应保障。 "建造晶圆厂绝非易事。" 黄仁勋指出，尽管生产设备可以通过采购获得，但真正的制造能力，取 决于数十年积累的工艺研发经验、复杂的设备整合技术，以及高度协同的供应链体系。他以台积 电等行业龙头企业为例，佐证先进芯片制造领域的成功关键在于持续稳定的技术落地能力，而非 单纯的资金投入。 供应链安全驱动马斯克的晶圆厂雄心 随着特斯拉对定制化人工智能处理器、自动驾驶硬件及车载计算系统的需求持续攀升，马斯克大 力推进自建大型芯片工厂的计划，希望借此降低未来遭遇半导体供应短缺的风险。 科技媒体 Wccftech 去年 11 月曾援引黄仁勋的观点称，即便是成熟的科技 ...

如果您希望可以时常见面，欢迎标星收藏哦~ 当前科技行业供应链，尤其是消费电子领域，正面临存储芯片及各类原材料供应短缺的压力。日 经亚洲此前报道，小米、OPPO、vivo 和传音等多家中国智能手机厂商已下调了今年的出货预 期。 尽管苹果是全球最大的采购商之一，但也在应对潜在的供应限制问题。其部分现有供应商已将资 源和重心转向服务英伟达、谷歌、亚马逊等快速扩张的人工智能领域头部企业。在本周四的财报 电话会议上，苹果公司警告称，iPhone 正面临供应限制，这一问题已对今年 1 至 3 月的季度业 绩造成影响。日经亚洲此前还报道，由于人工智能服务器系统消耗了全球大部分高端芯片基板用 玻纤布，苹果也在为这类材料的供应问题担忧。 这家 iPhone 制造商即将在其位于加州库比蒂诺的总部召开年度供应商大会。一位所在企业获邀 参会的消息人士透露，本次大会规模有所扩大，纳入了更多零部件厂商和原材料供应商，目的是 确保今年供应链的稳定性。 据日经亚洲报道，受营销策略调整及供应链紧张影响，苹果公司正优先保障 2026 年三款高端新 iPhone 机型的生产与出货，同时推迟标准版机型的上市时间。 四位知情人士透露，这家美国科技巨头 ...

Core Viewpoint - The semiconductor industry is entering a new phase with the mass production of 2nm processes, where advanced packaging technology is becoming a focal point due to its critical role in chip performance enhancement beyond mere transistor scaling [1]. Group 1: Advanced Packaging - Advanced packaging is not a single technology but a series of solutions aimed at enhancing chip integration, connectivity, and system performance, evolving from traditional packaging to more complex structures like 2.5D and 3D stacking [2]. - This technology does not directly increase chip computation speed but allows for more efficient utilization of computational power, akin to equipping characters with suitable gear to maximize their potential [2]. Group 2: Performance Impact of Connectivity - The layout of connections within advanced chips significantly affects performance, as data movement can consume more power than computation itself; inefficient routing leads to delays and energy waste [3]. - Advanced packaging improves thermal management, which is crucial as chip stacking density increases; effective heat dissipation is essential for maintaining performance levels [3]. Group 3: Differentiated Needs in Applications - Different application scenarios have distinct packaging requirements; AI and data center chips prioritize maximum output and bandwidth, while mobile device chips focus on compactness and power efficiency [4]. - AI chips are designed for high performance, while mobile chips must balance integration and power consumption, reflecting a divergence in packaging strategies [4]. Group 4: Innovations in Materials and Techniques - The industry is exploring glass substrates to replace traditional plastic materials, offering benefits such as finer signal lines and better thermal stability, which can lower production costs by allowing more chips to be packaged simultaneously [5]. - Panel-level packaging (FOPLP) represents a shift towards efficiency by utilizing square packaging instead of circular wafers, maximizing space usage and reducing costs [5].

汽车制造商一直面临着电动汽车推广缓慢和来自中国竞争对手的激烈挑战。与此同时，汽车制造商 仍在消化疫情高峰期积累的芯片库存，这意味着近年来该行业对半导体的需求一直低迷。 汽车行业的疲软与人工智能芯片形成了鲜明对比，尽管人们担心人工智能泡沫，但随着企业寻求更 先进的半导体来为数据中心提供动力，人工智能芯片的需求依然强劲。 人工智能芯片与传统半导体之间的需求分化，迫使意法半导体近年来多次下调业绩预期并裁员。奇 瑞在6月份表示，预计到2027年将有5000名员工离开公司，其中包括意法半导体4月份宣布的 2800个裁员岗位，这是其重组生产布局和削减成本计划的一部分。 如果您希望可以时常见面，欢迎标星收藏哦~ 意法半导体第四季度销售额有所增长，因为客户对用于个人电子产品、通信设备、计算机外围设备 和工业机械的芯片的需求增加，尽管汽车行业对半导体的需求难以反弹。 这家欧洲芯片制造商公布的销售额为33.3亿美元，在经历了几个季度的下滑后，实现了同比增长 0.2%，重回正轨。销售额高于公司此前预期的中值，也超过了Visible Alpha分析师预测的32.8 亿美元。 尽管有所改善，但首席执行官Jean-Marc Chery表 ...

Core Insights - The rise of artificial intelligence has led to a consensus that Nvidia's chips are essential for large AI projects, but Amazon and Google have begun to challenge Nvidia's dominance by developing their own AI chips [1][2]. Group 1: Market Dynamics - Nvidia holds a commanding 92% market share in the AI chip sector, with projected revenues nearing $200 billion by 2025 [2]. - Amazon's self-developed AI chip, Trainium, is expected to generate "tens of billions" in revenue by 2025, while Google's Tensor Processing Unit (TPU) has already reached hundreds of billions in revenue [2]. - The competition from Amazon and Google is particularly significant as they are both expanding their chip businesses while still relying on Nvidia as a core supplier [2][3]. Group 2: Strategic Partnerships - Anthropic, a leading AI company, is reducing its reliance on Nvidia chips and has secured significant chip sales for Amazon and Google, totaling hundreds of billions [3][4]. - Google has allowed Anthropic to install its chips in non-Google data centers, marking a new collaborative model in the industry [3][4]. - Amazon's investment of $4 billion in Anthropic was motivated by the promise of using Amazon's chips for AI system development, aiming to create a competitive alternative to Nvidia [5][6]. Group 3: Industry Implications - The rapid growth of Amazon's chip business, driven by partnerships with companies like Anthropic, is expected to lead to significant industry changes, signaling that Nvidia chips are not the only option available [6][7]. - Other chip manufacturers, such as AMD and Cerebras, are also seeking partnerships with AI companies to expand their market presence [6][7]. - The increasing acceptance of non-Nvidia chips by companies like Anthropic and OpenAI is likely to encourage broader industry adoption of alternative chip solutions [7].

Core Insights - Microsoft has deployed its first self-developed AI chip, Maia 200, in a data center and plans to expand its deployment in the coming months, positioning it as a core for AI inference computing power [1] - The Maia 200 chip is optimized for high computational loads during the mass production phase of AI models, boasting performance that surpasses Amazon's latest Trainium chip and Google's Tensor Processing Unit (TPU) [1] - Despite the introduction of its own chip, Microsoft CEO Satya Nadella emphasized the company's ongoing partnerships with Nvidia and AMD, indicating a strategy of not solely relying on vertical integration [1] Summary by Sections Chip Development and Deployment - Microsoft has announced the deployment of the Maia 200 chip, which is designed for AI inference and optimized for high-load scenarios [1] - The chip's performance is reported to exceed that of competitors like Amazon and Google [1] Strategic Partnerships - Nadella highlighted the importance of maintaining relationships with other chip manufacturers, stating that innovation from partners is crucial for future competitiveness [1] - The company will continue to purchase chips from Nvidia and AMD, despite its own chip development [1] Internal Usage and Future Plans - The Maia 200 chip will first be utilized by Microsoft's "Super Intelligence" team, led by Mustafa Suleyman, a co-founder of Google's DeepMind [2] - This initiative aims to reduce reliance on external AI model providers like OpenAI and Anthropic [2] - The chip will also support OpenAI models running on Microsoft's Azure cloud platform, although access to advanced AI hardware remains limited [2]

如果您希望可以时常见面，欢迎标星收藏哦~ 铠侠控股株式会社的子公司铠侠株式会社（Kioxia Corporation）与闪迪公司今日宣布，将双方 在铠侠四日市工厂的合资协议延长五年。该协议原定于 2029 年 12 月 31 日到期，现有效期将延 长至 2034 年 12 月 31 日，为两家公司长达 25 年以上的深厚合作关系再添重要里程碑。 此次协议延期凸显了铠侠与闪迪的深度协作 —— 双方将继续借助人工智能赋能的智能制造技术 及规模经济优势，保障先进 3D 闪存的稳定生产。其高端 3D 存储产品对满足生成式人工智能应 用催生的海量需求至关重要。此外，铠侠北上工厂的合资协议也将同步调整，与四日市工厂协议 保持一致，有效期延长至 2034 年 12 月 31 日。 根据这份续约协议，闪迪将向铠侠支付 11.65 亿美元，用于支付制造服务费用及确保持续供货。 该笔款项将在 2026 年至 2029 年期间分期支付。 "我们很高兴能深化与闪迪的战略合作伙伴关系。" 铠侠总裁兼首席执行官早坂伸夫（Nobuo Hayasaka）表示，"这份协议不仅认可了铠侠制造业务的价值，更通过全球最大闪存制造基地的 规模经济效应 ...

更大的模型、更高端的GPU、更密集的算力投入，构成了这一阶段最鲜明的特征。训练能力一度 等同于技术先进性，也决定了厂商在产业链中的话语权。但随着大模型规模趋于稳定、训练节奏 放缓，AI系统的主要负载正在发生转移。真正被频繁调用、持续消耗算力的，不再是一次性的模 型训练，而是无处不在、实时发生的推理请求。 在 AI 产 业 从 训 练 走 向 推 理 的 关 键 转 折 点 ， 曦 望 科 技 于 2026 年 1 月 27 日 举 办 首 届 Sunrise GPU Summit产品发布会，正式发布新一代推理GPU芯片启望S3，并同步推出寰望SC3超节点解决方 案及推理云计划。这是曦望在完成近30亿元战略融资后的首次系统性技术亮相。 如果您希望可以时常见面，欢迎标星收藏哦~ 过去两年，AI产业的重心高度集中在训练。 在发布会上，中国工程院院士、浙江大学信息学部主任吴汉明出席发表致辞并指出，当前国内正 处于集成电路与人工智能深度融合的关键时期，算力作为核心生产力，关系到科技自立自强战略 的落实。 "AI规模化应用对算力提出了前所未有的要求，传统芯片已难以满足多场景下的高效能需求。"吴 汉明强调，推理算力价值的实 ...

Core Viewpoint - Elon Musk announced that Tesla plans to build and operate a semiconductor manufacturing facility called "TeraFab," which will require billions of dollars in investment and signifies an expansion beyond its core electric vehicle business [1][5]. Group 1: TeraFab Project - The TeraFab will be a large-scale factory covering various aspects of semiconductor production, including logic circuits, memory, and packaging, with production based in the United States [1]. - Musk emphasized the necessity of TeraFab to avoid potential capacity bottlenecks in the next three to four years, as existing suppliers like TSMC and Samsung cannot meet Tesla's demand [1][2]. - The factory is expected to have a production capacity that exceeds 100,000 wafers per month, positioning Tesla among the largest semiconductor manufacturers globally [5]. Group 2: Financial Considerations - Tesla is projected to spend over $20 billion on capital expenditures this year, with funding sources for the TeraFab and other infrastructure projects still under consideration [3]. - The company has over $44 billion in cash and investments, which will be utilized alongside other financing options, including potential bank loans [3]. Group 3: Industry Challenges - Building a semiconductor factory involves significant economic costs, with advanced facilities requiring hundreds of billions in fixed costs and a lengthy timeline from construction to full operation [2][6]. - The complexity of semiconductor manufacturing is often underestimated, as it requires extensive engineering expertise and a deep understanding of various processes [6][7]. - The industry faces challenges in achieving high yield rates for new manufacturing processes, which are critical for market stability and profitability [8].

Core Viewpoint - The article provides a comprehensive overview of optical transceiver terminology and standards, particularly focusing on the IEEE 802.3 standards that define the electrical and optical specifications for physical layer (PHY) communications. It aims to equip readers with the knowledge to understand optical transceiver product specifications like an industry expert [3][39]. Group 1: Optical Transceiver Basics - The naming convention for optical transceivers is derived from the IEEE Ethernet Working Group, which defines various standards under IEEE 802.3, including the upcoming 802.3dj standard that will support data rates of 200 Gbps to 1.6 Tbps [3][4]. - The optical interconnect definition typically follows a format that includes connector size, base speed, transmission distance, channel count, modulation scheme, multiplexing method, fiber mode, and additional information [3][4]. Group 2: Form Factors and Data Rates - The first part of the product name indicates the size specification of the pluggable connector, with QSFP representing a quad-channel small form-factor pluggable connector widely used in 400G networks [4][6]. - The maximum data rates for various form factors are outlined, with QSFP-DD supporting up to 400 Gbps and OSFP supporting up to 800 Gbps, indicating a trend towards higher data rates in optical communications [8][9]. Group 3: Effective Distance and Parallel Channels - Optical communication technologies are categorized into nine distance levels, from very short range (VSR) to long-distance (ZR), with the effective distance influenced by data rate, modulation method, and fiber quality [12][14]. - The example product operates with four parallel optical connections to achieve a total bandwidth of 400 Gbps, meaning each channel runs at 100 Gbps [15]. Group 4: Multiplexing and Modulation Schemes - Multiplexing methods, such as CWDM4, allow data from independent parallel channels to be combined into a single aggregated connection, enhancing bandwidth efficiency [18][20]. - Modulation schemes, including PAM4, are essential for converting electrical signals into optical signals, especially at higher data rates where signal distortion becomes a concern [21][24]. Group 5: Fiber Modes and Additional Information - Two main types of fiber are used in optical networks: single-mode fiber (SMF) and multi-mode fiber (MMF), with SMF being more suitable for long-distance and high-speed applications [31][36]. - Additional information in optical transceiver specifications may include reach, connector types, and the need for digital signal processing (DSP) to maintain signal integrity over longer distances [37][39].