由 于 市 场 对 谷 歌 人 工 智 能 芯 片 —— 张 量 处 理 单 元 （ TPU ） 的 需 求 预 计 将 十 分 强 劲 ， 高 带 宽 内 存 （HBM）供应链正步入一个新阶段。有预测显示，谷歌将成为HBM市场仅次于英伟达的关键需求来 源，三星电子和SK海力士之间争夺供应链主导地位的竞争也日趋激烈。 自去年下半年以来，三星电子一直向AMD旗舰级AI加速器MI350系列供应12层HBM3E产品。随着加 速器供应来源日益多元化，HBM的需求来源也将更加广泛，韩国存储器制造商的议价能力也可能会 据台湾主流媒体2月26日报道，投资银行美国银行将谷歌今年的TPU出货量预测从400万颗上调至460 万颗。这大约是此前预测的2025年230万颗的两倍。半导体行业内外也有人指出，谷歌今年在HBM市 场的份额可能超过30%。 此次预测的上调表明专用集成电路（ASIC）市场正在实现结构性增长。尽管人工智能基础设施迄今 为止主要围绕图形处理器（GPU）构建，但针对特定任务优化的ASIC很可能在未来占据相当可观的 市场份额。诸如TPU之类的ASIC针对训练和推理等特定用途进行了优化，正在成为不断发展的人工 智能时代 ...

Core Insights - Microsoft has officially launched the Maia 200 chip, marking a new phase in the competition among cloud service providers and AI companies for Application-Specific Integrated Circuits (ASICs) [1] - The year 2027 is predicted to be a critical turning point, with several major manufacturers expected to significantly increase ASIC production [1] - Broadcom maintains a dominant market share of approximately 60%, which is unlikely to be challenged in the short term [1] Group 1: Market Dynamics - Major brands have established partnerships with Broadcom, while other competitors are primarily vying for the second position in the market [1] - Companies like Maywell, World Semiconductor, and MediaTek are actively positioning themselves, with any significant project win potentially solidifying their market standings [1] - The outcome of this competition is expected to be revealed in 2027, as companies like Google, AWS, Meta, Microsoft, OpenAI, Apple, and ByteDance plan to substantially increase their ASIC procurement [1] Group 2: Competitive Landscape - Maywell faces challenges as Broadcom has taken a dominant position in the market, previously shared between the two [2] - Maywell is focusing on AWS's Trainium product line but is encountering fierce competition from World Semiconductor [2] - MediaTek is gaining momentum, benefiting from Google's dual-spec ASIC strategy, and is seen as a strong contender for the second position in the market [2] Group 3: Future Prospects - World Semiconductor's future largely depends on the production progress of AWS's next-generation Trainium chips in 2026 and 2027, as well as its collaboration with Intel [2] - The company has yet to achieve significant success with other major U.S. cloud service customers and plans to expand into more mid-sized client projects [2] - Without securing larger-scale cloud service contracts, World Semiconductor may fall behind in the upcoming market share competition [2]

Core Viewpoint - The demand for custom chips is increasing among large-scale data center operators, with Broadcom positioned as a leader in this field, particularly highlighted by Google's successful use of Broadcom's custom-designed Tensor Processing Units (TPUs) for training its Gemini AI models [2][3]. Group 1: Custom Chips vs. General-Purpose GPUs - Custom chips are specifically designed for large-scale tasks, unlike NVIDIA's general-purpose GPUs, which are claimed to have broader market applicability [3]. - NVIDIA's CEO downplayed the threat posed by custom chips, asserting that NVIDIA's products are more versatile and can operate across various computing scenarios [3]. - Despite the rise of custom chips, NVIDIA remains a significant supplier for major clients like Google, which relies heavily on NVIDIA GPUs for its cloud infrastructure [3][4]. Group 2: Market Dynamics and Risks - Analysts believe NVIDIA faces limited market risks currently, but its dominance in the AI chip market is being tested as competitors enter the space [4]. - The market is diversifying, as evidenced by Broadcom's partnership with OpenAI for custom chip development, indicating a shift away from reliance on a single supplier [4][5]. - The high production barriers for Application-Specific Integrated Circuits (ASICs) favor NVIDIA, as smaller companies may struggle with the costs and time associated with developing custom chips [5]. Group 3: Financial Performance and Projections - Broadcom reported a 65% year-over-year increase in AI business revenue, reaching $20 billion, contributing to a record $37 billion in total semiconductor revenue [5]. - Analysts predict NVIDIA will maintain over 50% market share in the next five years, with a potential 70% share in the next three years [5]. - Morgan Stanley maintains a "buy" rating for both Broadcom and NVIDIA, while Wolfe Research is optimistic about Broadcom's future, raising its rating and projecting significant growth in TPU shipments [6]. Group 4: Investor Sentiment and Stock Performance - Investor sentiment towards Broadcom has been cautious despite its strong financial performance, with stock prices recently declining [7]. - NVIDIA's stock has seen slight increases but faces pressure from valuation concerns and geopolitical tensions [7]. - Analysts suggest that upcoming product announcements from NVIDIA could positively impact its stock price [8].

Core Insights - Intel's Jaguar Shores AI product line is in talks with ASIC design firm Wistron NeWeb Corporation to develop customized AI solutions, with completion expected in the first half of 2026 [1][2] - Intel has shifted its AI strategy to focus on annual product updates and has introduced the Xe3P architecture "Crescent Island" AI chip to enhance its competitive stance in the AI market [1] - Wistron NeWeb Corporation, a fabless ASIC design service provider, is collaborating with Amazon on the Trainium series AI chips, indicating its capability in the ASIC market [2] Group 1 - Intel's AI strategy has evolved from focusing solely on "inference" to a broader approach, responding to the growing AI market [1] - The collaboration with Wistron NeWeb is seen as a strategic move for Intel, leveraging Wistron's expertise in ASIC design and manufacturing to alleviate Intel's development burdens [2] - The Jaguar Shores product line is anticipated to launch in 2026, with a reasonable timeline for mass production suggested for the second half of that year [2] Group 2 - Wistron Neweb plays a dual role as a design and manufacturing partner, outsourcing actual wafer fabrication to companies like TSMC while focusing on critical processes such as chip integration and testing [2] - The partnership with Wistron Neweb is expected to provide Intel with a competitive edge in developing more powerful AI products as the industry matures [2] - The collaboration reflects a trend where established companies like Intel seek partnerships with specialized firms to enhance their product offerings in the rapidly evolving AI landscape [2]

Core Viewpoint - Samsung Electronics is set to supply 12-layer HBM3E to Broadcom, with plans for mass production starting as early as the second half of this year to next year, aiming to mitigate the impact of NVIDIA's HBM supply delays [1][3]. Group 1: Supply Agreements - Samsung has completed quality testing for HBM3E 12-layer with Broadcom and is negotiating supply volumes estimated between 12 billion to 14 billion Gb, with mass production expected soon [1]. - Samsung is also in active discussions to supply HBM3E 12-layer memory to Amazon Web Services (AWS), which plans to produce the next generation AI semiconductor "Trainium 3" using this memory [2]. Group 2: Market Dynamics - The surge in development of proprietary ASICs by major tech companies presents an opportunity for Samsung to offset the downturn in its HBM business [3]. - Samsung's initial plan to supply NVIDIA with HBM3E 12-layer was delayed due to performance issues, and the company is now adjusting its production rates for HBM3E lines [3]. Group 3: Production Goals - Samsung aims to double its total HBM supply to between 8 billion to 9 billion GB this year, compared to last year [1]. - The successful supply to NVIDIA and acquisition of more ASIC clients in the second half of this year is crucial for stabilizing Samsung's HBM business [3].