Summary of Conference Call Notes Industry and Company Involvement - The discussion primarily revolves around the **Artificial Intelligence (AI)** sector and its implications for the **Chinese market** over the next decade, as well as the performance of **U.S. tech giants** like **Apple**, **NVIDIA**, and **Google** [1][2][3][6][7]. Core Insights and Arguments - **AI as a Market Driver**: AI is expected to be a significant driver of a structural bull market, with substantial advancements anticipated in China over the next ten years, supported by national efforts [1][2][3]. - **Quantum Computing and Autonomous Driving**: Breakthroughs in quantum computing are expected by 2027, with companies like NVIDIA already launching quantum chips. Autonomous driving technology is projected to see revolutionary changes by the end of 2026, presenting opportunities for domestic industries [1][4]. - **AI in Healthcare**: The application of AI in healthcare shows vast potential, as demonstrated by companies like Ant Financial. High dividend strategies and cyclical industries related to heavy metals are also worth monitoring, despite geopolitical risks [1][5]. - **Market Outlook for 2027**: Optimism is expressed regarding the A-share market in 2027, although caution is advised for the consumer and liquor sectors, which may face challenges without policy support. The S&P 500 is expected to see earnings growth surpassing that of the current year [1][6]. - **Performance of U.S. Tech Giants**: Companies like Apple, NVIDIA, and Google are noted for their strong profitability, with annual profits exceeding $100 billion. Their ability to continue investing and maintaining competitiveness is emphasized, with no bubble concerns identified [1][7]. - **NVIDIA's CUDA Platform**: NVIDIA's CUDA platform has established a strong competitive moat, capturing over 90% of the market share. Other companies face challenges in creating similar platforms due to compatibility issues [1][8]. Other Important but Potentially Overlooked Content - **Investment Strategies**: Long-term investors are encouraged to maintain their investment philosophies and not be swayed by short-term market fluctuations. The importance of finding a capable team for asset management is highlighted, as team performance is crucial for success [3][9][10][12]. - **Noise in the Market**: Various market noise factors, such as trade wars and interest rate changes, are deemed secondary to the primary trend of AI development. Investors are advised to focus on major trends rather than being distracted by these noises [14]. - **Assumptions in Investment Models**: The assumptions made in investment models significantly impact decision-making. It is essential to consider both optimistic and pessimistic scenarios when forecasting a company's growth [15][16]. This summary encapsulates the key points discussed in the conference call, providing insights into the AI sector, market outlook, and investment strategies.

Summary of Key Points from the Conference Call Industry Overview - The report focuses on the semiconductor industry, particularly the competitive dynamics between Google's Tensor Processing Units (TPUs) and Nvidia's Graphics Processing Units (GPUs) in cloud-based AI computing [2][7][8]. Core Insights - **TPU Growth**: Google's TPUs are expected to grow more rapidly than Nvidia's GPUs from a smaller base, with TPUs offering superior efficiency in performance per watt and per dollar for large-scale AI workloads [2][8]. - **Internal Usage**: Google relies heavily on TPUs for its internal AI training and inference, indicating the platform's maturity and reliability [2][28]. - **Market Forecast**: TPU shipments are projected to reach 4 million units in 2026 and grow to 7.2 million units in 2027, with MediaTek's share expanding from 8% in 2026 to 28% in 2027 [3][44]. Competitive Landscape - **Dual TPU Tracks**: Google is adopting a dual-track strategy for TPU development, collaborating with both Broadcom and MediaTek. This approach allows Google to diversify its supply chain and manage costs effectively [3][36][44]. - **Cost Efficiency**: MediaTek's service fees for TPUs are over 50% lower than Broadcom's, making it a significant player in the TPU supply chain [3][37]. Key Beneficiaries - **TSMC**: As the leading-edge foundry, TSMC is expected to benefit significantly from the demand for TPUs [4]. - **Other Suppliers**: Companies like ASE, KYEC, Advantest, and Celestica are also positioned to gain from the growing TPU market [4]. Technical Advantages of TPUs - **Design Efficiency**: TPUs are specifically designed for neural network computing, offering competitive performance-per-watt and performance-per-dollar compared to general-purpose GPUs [11][14]. - **Architecture**: The TPU architecture allows for higher compute utilization and efficiency, minimizing runtime loss compared to GPUs [16]. Software Integration - **OpenXLA**: Google's development of the OpenXLA software standard facilitates easier migration for developers transitioning from Nvidia GPUs to TPUs, enhancing the appeal of TPUs for external users [20][29]. Future Outlook - **Market Position**: Google is positioned as a key player in the frontier AI model development alongside OpenAI and Anthropic, driving substantial demand for TPUs [31][35]. - **Cloud Revenue Growth**: The cloud revenue for major hyperscalers, including Google, is expected to grow at a robust 29% CAGR from 2026 to 2028, driven by the shift towards AI-centric workloads [32][33]. MediaTek's Role - **Strategic Partnership**: MediaTek's collaboration with Google is expected to significantly enhance its market position, with potential sales from TPU v8X projected between $8 billion to $17 billion in 2027 [58]. - **Technology Development**: MediaTek is also advancing its SerDes IP technology, which is crucial for the TPU v8X project, potentially positioning it for future growth in the cloud and edge AI markets [56][58]. Conclusion - The competitive dynamics between TPUs and GPUs are evolving, with Google's strategic partnerships and technological advancements positioning it favorably in the semiconductor landscape. The expected growth in TPU shipments and the increasing reliance on AI workloads underscore the significant opportunities within this sector [2][3][8][31].

Summary of Key Points from the Conference Call Industry Overview - The conference call focuses on the **Global I/O Hardware** industry, particularly the **memory market** and its implications for **hyperscale capital expenditures (capex)** and **hardware margins** due to rising demand from AI applications and server requirements [2][8]. Core Insights and Arguments 1. **Memory Pricing Forecasts**: - UBS forecasts a **289% increase** in DRAM pricing per Gb from 2025 to 2027, surpassing the previous cycle's 89% rise [2][13]. - NAND pricing is expected to rise **144%**, exceeding the 92% rebound seen in 2024 [2][13]. 2. **Hyperscale Capex Growth**: - Hyperscale capex is projected to reach **US$827 billion** in 2026 and **US$915 billion** in 2027, reflecting increases of **43%** and **28%** respectively [3][27]. - Memory costs are estimated to add approximately **US$100 billion** annually to hyperscale capex, increasing from **US$53 billion** in 2025 to **US$155 billion** in 2026 and **US$252 billion** in 2027 [3][25]. 3. **Impact on PC and Smartphone Markets**: - PC unit forecasts for 2026 have been revised down from **267 million** to **255 million**, indicating a **4% decline** year-over-year [4]. - Smartphone unit sell-in estimates have also been reduced from **1.28 billion** to **1.20 billion**, reflecting a **5% decline** in 2026 [4]. 4. **Brand and ODM Margin Pressures**: - The rising memory costs are squeezing margins for brands and ODMs, with DRAM now accounting for **18%** of PC BOM costs and potentially **24%** for high-end smartphones by H226 [4][9]. - ODMs are forced to pass through memory costs, which boosts sales but does not enhance gross or operating profits, leading to lower margins [4][9]. 5. **Investment Preferences**: - Analysts recommend favoring AI hardware and components over traditional PC and branded companies due to the higher memory costs impacting margins [5]. Additional Important Insights 1. **Server Demand and Memory Costs**: - The demand for AI and traditional servers is strong, with server unit forecasts for 2025-26 revised up from **+6%** to **+13%** year-over-year [2][14]. - A typical data center server's memory cost is expected to rise from **46%** to **67%** of total server costs due to increased memory pricing [15]. 2. **Long-term Memory Cycle**: - The memory pricing cycle is described as the strongest in decades, with expectations of sustained under-supply into 2027 [10][13]. 3. **Capex and Cash Flow Dynamics**: - The capital intensity for internet companies is projected to rise from **10%** of capex/sales pre-AI to approximately **35%** in 2026, with a significant portion of capex financed through operating cash flow [28]. 4. **Future Outlook**: - The aggressive push for compute resources to meet growing cloud workloads and AI demands is expected to continue, although future growth rates may slow as spending becomes increasingly financed by debt and equity [28]. This summary encapsulates the critical insights and projections discussed during the conference call, highlighting the significant impact of memory pricing on the hardware industry and the strategic shifts in investment focus towards AI-related technologies.

曾惠丹，华东理工大学教授、博士生导师，玛丽居里学者。上海电子化学品创新研究院首席研究员、中国硅酸盐学会理事，主要研究方向为特种玻璃和电子 浆料，研究应用于光电子器件、半导体和电子元件的特种钝化材料/封装基板和电极浆料。 出品｜搜狐科技 作者｜张雅婷 编辑｜杨锦 或许很少有人能料到，一种来自日本的布，竟会掐住全球AI芯片巨头的"咽喉"。 搜狐科技《思想大爆炸——对话科学家》栏目第148期，对话华东理工大学教授、博士生导师曾惠丹。 嘉宾简介： 高端电子玻璃纤维布，是芯片基板和印刷电路板不可或缺的核心骨架，由大概100-200 根单丝直径为4-7微米的玻璃纱股编织而成布后，再与树脂、铜箔等材 料复合，经过图形蚀刻、钻孔镀铜等精密工艺制成。 随着AI热潮带来的AI芯片需求爆发式增长，日东纺生产的高端电子玻璃纤维布供应日趋紧张，甚至陷入"一布难求"的局面。为确保供应稳定，英伟达CEO黄 仁勋近日亲自登门拜访日东纺，苹果公司也正积极向日本政府官员寻求支援。 在这场没有硝烟的供应链战争中，日东纺成为了最大赢家。这个成立于1923年的日本企业从棉纺织起家，至今已经有100多年的历史，目前在高端玻纤布市 场垄断了全球90%以上 ...

Core Viewpoint - Software stocks are experiencing significant declines due to concerns that artificial intelligence (AI) may disrupt their business models, leading to reduced reliance on software solutions [1][2]. Group 1: Market Trends - Investors are selling shares of software companies in anticipation of future struggles as AI changes business operations [2][3]. - The iShares Expanded Tech-Software Sector ETF, which includes major software stocks like Salesforce and Adobe, has dropped approximately 20% this year, despite a recent bounce following comments from Nvidia's CEO [8]. Group 2: Nvidia's Perspective - Nvidia CEO Jensen Huang argues that it is "illogical" to believe AI will completely replace software, suggesting instead that AI will enhance existing tools and solutions [5]. - Nvidia has significantly contributed to AI growth, reporting nearly $100 billion in profits over the past 12 months, a stark increase from under $5 billion a few years ago [6]. Group 3: Investment Opportunities - Current market conditions may present attractive buying opportunities for software stocks, with Salesforce and Adobe trading near multi-year lows [9]. - Investors are advised to evaluate each software company individually, focusing on upcoming earnings reports and growth amidst AI developments to determine the viability of investments [11].

Market Performance - The Hang Seng Index closed at 27,027, up 1.76% with a year-to-date increase of 5.45% [1] - The Hang Seng Technology Index rose by 1.34% to 5,417, but is down 1.78% year-to-date [1] - The Shanghai Composite Index increased by 1.41% to 4,123, with a year-to-date rise of 3.89% [4] - The Dow Jones Index reached a new high, closing at 50,135, up 0.04% year-to-date [4] Currency Trends - The Renminbi Index is at 98.340, up 0.36% over the last month and 1.44% over the last six months [2] - The Euro to USD exchange rate is at 1.187, increasing by 1.98% in one month and 2.17% in six months [2] Commodity Prices - Brent crude oil is priced at $67.63 per barrel, up 7.71% in one month and 4.01% in six months [3] - Gold is priced at $5,022.91 per ounce, with an increase of 11.38% in one month and 50.27% in six months [3] - Silver has seen a significant rise of 117.60% over the last six months, currently priced at $81.85 per ounce [3] Company Financials - Hong Kong Telecom (06823) reported a 4% increase in EBITDA for the year ending December 31, 2025, with total revenue rising by 5% to HKD 36.553 billion [9] - The number of 5G plan users in Hong Kong Telecom's postpaid customer base exceeded 2 million, representing 60% of the total postpaid customers [10] - Local data service revenue grew by 6% to HKD 14.31 billion, contributing to a 3% increase in overall local telecom service revenue [10] Mortgage Insurance Trends - In January, new mortgage insurance applications in Hong Kong fell by 8.7% to 460 cases, the lowest since November 2018 [11] - The total amount of new mortgage insurance also decreased by 7.1% to HKD 2.36 billion [11] - The usage ratio of mortgage insurance has dropped to 12.1%, the lowest level since November 2019 [12] Economic Indicators - The New York Federal Reserve reported a decrease in one-year inflation expectations to 3.1% in January, down from 3.4% in December [13][14] - The Federal Reserve's outlook on the U.S. fiscal situation and economic growth remains optimistic, with expectations of a significant GDP growth boost due to tax policy changes [15]

来源：格隆汇APP Hawtin还指出，人工智能竞赛中一个关键因素可能决定胜负：专有数据。"数据的所有权将绝对决定谁 能最终胜出。"并以谷歌母公司Alphabet去年的强劲表现为例。Hawtin表示，尽管Alphabet去年起初表现 不佳，但最终仍成为"科技七巨头"中表现最好的公司，全年涨幅达65%，而"这主要归功于其庞大的专 有数据集"。 格隆汇2月11日｜"科技七巨头"——苹果、微软、亚马逊、Alphabet、Meta、特斯拉和英伟达——不断 增长的资本支出，预示着投资者面临的风险增加，并可能成为这些主导公司最终跑输大盘的催化剂。这 是英国基金管理公司Liontrust全球股票主管Hawtin的最新观点。他指出，最近财报中宣布的激进支出计 划是那些寻求可靠回报的股东们最关心的问题。 ...

Summary of Conference Call Notes Industry Overview - The focus is on the AI networking sector, particularly the advancements in CPO (Coherent Passive Optical) technology and its implications for data centers and optical interconnects [1][2][3]. Key Points and Arguments 1. **CPO Technology Developments**: - Coherent has secured a significant order from a leading AI data center client for its CPO system, with initial revenue expected to start in Q4 2026 and become more substantial in 2027 [1]. - Lumentum has also received commitments worth hundreds of millions for Scale-out CPO CWLasers, with mass production contributions anticipated to begin in H1 2027 [1]. 2. **Market Projections**: - It is estimated that laser suppliers like LITE and COHR will have a combined order value of $400 million in 2027, leading to the shipment of approximately 80,000 Scale-out CPO switches that year [1]. - Nvidia is expected to accelerate the rollout of Scale-out CPO switches, with projected shipments of 2,000 units in 2025, 20,000 in 2026, and 80,000 in 2027 [2]. 3. **Nvidia's Product Launches**: - Nvidia plans to launch its Quantum-X Scale-out CPO switch in H2 2025 and Spectrum-X in H2 2026, with the first generation featuring a pluggable architecture [2]. - The new generation CPO switch is anticipated to have significant improvements in thermal performance and bandwidth, with production ramping up in Q2 2026 [2]. 4. **Long-term Opportunities**: - Lumentum highlighted that as copper interconnects approach their limits, optical Scale-up presents a long-term structural opportunity starting from late 2027 [3]. - Nvidia is expected to introduce CPO/NPO technology within its NVL576 architecture for Scale-up interconnects, indicating a competitive landscape between CPO and NPO technologies [3]. 5. **Beneficiary Companies**: - Key beneficiaries identified include FAU, CWLaser, ELSFP modules, and Shuffle components suppliers [4]. - Lumentum is viewed positively due to the upward demand for CWLasers, while the impact on Coherent is considered neutral [4]. - Browave is expected to benefit significantly due to its 50% market share in the CPO Shuttle Box sector, with production ramping up in Q2 2026 [4]. Additional Important Insights - The competition between CPO and NPO technologies is expected to continue, with both solutions sharing the same suppliers, indicating a complex supply chain dynamic [3]. - The overall market for Scale-up CPO is seen as a pure incremental opportunity for the optical interconnect supply chain, aimed at replacing copper interconnects without affecting Scale-out solutions [3].

Company and Industry Summary Company: Ta Liang Key Points - **Upgrade of Communication Protocol**: The upcoming NVIDIA Rubin platform will feature an upgraded communication protocol between GPU and NVSwitch, utilizing 224GSerDes on PCB with a significant signal frequency increase to a maximum of 56GHz, compared to 28GHz on the Blackwell platform. This upgrade indicates that even minor via stubs can lead to severe signal degradation, establishing a ±2mil stub height as the threshold for signal transmission quality. Accurate closed-loop measurement is deemed crucial for manufacturing quality control [1][1][1] - **StubMapper Device**: Ta Liang has developed a proprietary device called StubMapper, which enables ±2mil depth measurement to guide subsequent back-drilling processes executed by Ta Liang's 6-axis CCD drilling machine. StubMapper can provide profile measurement data back to the back-drilling machine, enhancing precision in manufacturing [1][1][1] - **Increased Demand for Back-Drilling**: There has been a surge in industry demand for back-drilling to meet the requirements of high-end AI PCBs, resulting in an extension of PCB equipment delivery cycles from 1.5 months to over 3 months. This demand has led Ta Liang to consider outsourcing more low-end drilling tool production to focus on high-end tool manufacturing [2][2][2] - **Production Capacity and Future Plans**: Ta Liang's current monthly production capacity is approximately 300 sets of drilling tools. The company plans to enhance the throughput of StubMapper by introducing a four-probe version, with mass production scheduled for March 2026, and initial customer reservations already in place [2][2][2] - **Semiconductor Equipment Division**: Ta Liang also operates a semiconductor equipment division focused on measurement and automated optical inspection (AOI), offering 7 types of inspection equipment used in processes such as HBM incoming material and flip-chip bonding. This division has received certification from a leading foundry for CoWoS and SoIC platforms, with CoWoS-related revenue accounting for approximately 75% of its semiconductor equipment business [2][2][2] - **Milestone Achievement**: The certification from the leading foundry is considered a key milestone for Ta Liang, facilitating smoother entry into the OSAT supply chain as OSAT manufacturers expand their 2.5D capacity [2][2][2]

Core Insights - The International Solid-State Circuits Conference (ISSCC) will take place from February 15 to 19, 2026, in San Francisco, showcasing significant advancements in semiconductor technology [2] Group 1: AI Chips - AMD's latest AI GPU, Instinct MI350, features a CDNA4 architecture with a theoretical peak performance increase of 1.9 times compared to its predecessor, and improvements in HBM input/output bandwidth and memory capacity by 1.5 times [2] - Rebellions has developed a large-scale AI inference subsystem using UCIe protocol, achieving a performance of 56.8 TPS on the Llama 3.3 model with 700 billion parameters [3] - IBM's AI accelerator, Spyre, is optimized for inference, boasting a throughput 32% higher than the latest GPUs and energy efficiency 2 to 3 times better [3] - MediaTek's MADiC, a generative diffusion accelerator, achieves performance of 7.4 TOPS/mm² and 17.4 TOPS/W, designed for generative image editing on edge devices [4] - NVIDIA's ALPhA-Vision real-time image processor has a face detection latency of 787 microseconds and an accuracy rate of 99.3% [5] Group 2: Memory Technologies - SanDisk and Kioxia have developed a 3D NAND flash memory with a density of 37.6 Gbit/mm², capable of reaching a storage capacity of 2 Tbit and a write speed of 85 MB/s [6] - Samsung is set to release a DRAM module with a capacity of 36GB and a data transfer rate of up to 3.3 TB/s, utilizing 12 chips stacked together [7] - SK Hynix has developed a 16Gbit LPDDR6 SDRAM with a data transfer rate of 14.4 Gbps per I/O pin [7] - Samsung will also introduce a 16Gbit LPDDR6 SDRAM with a data transfer rate of 12.8 Gbps [8] - SK Hynix's 24Gbit GDDR7 DRAM targets mid-range AI inference applications with a data transfer rate of 48 Gbps [8] Group 3: Image Sensors - STMicroelectronics will showcase a lidar receiver with a field of view of 54°×42° and a power consumption of 153 mW [9] - Sony Semiconductor Solutions has developed a Ge-on-Si SPAD sensor array designed for low-power AR/VR applications, with a power consumption of 26 mW at 30 fps [10] - SmartSens Technology's CMOS image sensor features 200 million pixels and supports 8K video recording at 60 fps [11] Group 4: AI Chip Presentations - NVIDIA will present its GB10 processor for desktop AI supercomputers, featuring 20 Armv9.2 cores and a performance of 31 TFLOPS in FP32 mode [12] - STMicroelectronics will discuss the STM32N6 microcontroller series, integrating an Arm Cortex-M55 CPU and a Neural-ART NPU with performance of 600 GOPS and 3 TOPS/W [13] - Microsoft will explain its AI accelerator architecture, MAIA, focusing on packaging technology and power management [13]