Investment Rating - The report maintains an "In-Line" industry view for Greater China Technology Semiconductors [6]. Core Insights - China's AI GPU self-sufficiency is projected to increase from 34% in 2024 to 82% by 2027, with the total addressable market (TAM) for cloud AI expected to grow at a CAGR of 28% to reach US$239 billion by 2027, with China accounting for approximately US$48 billion [2][17]. - The overall semiconductor self-sufficiency ratio in China is currently at 24%, up from 20% in 2023, driven by government subsidies, inventory digestion, and capacity ramp-up in memory and leading node products [3][11]. - Local GPU suppliers like Huawei and Cambricon are primarily supported by SMIC, which faces capacity expansion challenges [2][18]. Summary by Sections AI GPU Self-Sufficiency - The self-sufficiency ratio for AI GPUs in China is estimated at 34% for 2024 and expected to rise to 82% by 2027, with significant growth in the cloud AI market [2][17]. - The TAM for cloud AI is projected to grow to US$48 billion in 2027, with China expected to capture 20% of global demand [2][17]. Semiconductor Market Overview - China's semiconductor market was valued at approximately US$183 billion in 2023, with local companies generating US$43 billion in revenue, marking a 36% increase from US$32 billion in 2023 [3]. - The self-sufficiency ratio for semiconductors in China is currently at 24%, reflecting a 4 percentage point increase from the previous year [3][11]. Localization Progress - Significant advancements have been made in memory, image sensors, and power semiconductors, while equipment and EDA progress has been slower than expected [8]. - Local vendors in memory and power semiconductors are benefiting from the growth in electric vehicles and gaining market share over global competitors [8]. Stock Implications - The report is equal-weight on SMIC and Empyrean Technology, noting that while SMIC is crucial for local AI chip production, potential Nvidia acquisitions could impact domestic GPU market share [4]. - Positive outlook on China wafer fab equipment makers like Naura, AMEC, and ACM Research due to local foundry and memory capacity expansion [4].

GenAI的内存解决方案 第 1 部分：能力的变化 所需能力 GenAI 应用需要高速、高带宽且低延迟的内存，以便实时处理海量数据。在需要实时决策和 预测的推理环节，数据的快速访问就显得尤为关键。 GenAI内存解决方案第 2 部分：HBM的竞争态势 内存设计的挑战与解决方案以及内存技术的最新趋势正在塑造高性能计算的未来及其竞争 格局。 竞争态势 技术革新： 具有传统接口的动态随机存取存储器（DRAM）在带宽和延迟方 面 存在局限，因此像高带宽内存（HBM）这类利用硅通孔（TSV）堆叠 DRAM 的 技术，就成为满足这些性能需求的关键解决方案。与内存设计相关的挑战与应 对办法，以及内存技术的新兴趋势，正塑造着高性能计算的未来与竞争格局。 优化策略： 未来，像3D-IC和（或）CoWoS等封装技术的进步，将在智能手 机、PC 等不同领域得到应用。智能手机受空间和成本限制，人们会尝试多种办 法，在不增加成本与空间占用的前提下，降低延迟、减少能耗。 应变准备 ： 目前仍不清楚到 2030 年哪些类型的GenAI模型和应用会流行，以 及具体数量是多少。因此，支持架构层面的进步并构建生态系统，以便能够应 对任何变化，将 ...

Group 1: Core Insights - GenAI applications require high-speed, high-bandwidth, and low-latency memory for real-time data processing, making fast data access critical for decision-making and predictions [2] - The competition landscape in memory technology is evolving, with challenges in memory design and emerging trends shaping the future of high-performance computing [4][6] - Custom HBM (High Bandwidth Memory) is becoming essential for AI systems, with significant growth expected as companies like NVIDIA, Amazon, and Microsoft drive its development [26][28] Group 2: Competitive Landscape - Samsung is lagging in HBM due to conservative actions in advanced testing and packaging, focusing too much on cost while failing to meet customer demands for heat dissipation and power consumption [7] - SK Hynix is leading the market by addressing NVIDIA's power and thermal constraints through innovative memory unit designs and backend processing [7][8] - Micron is entering the HBM market later than competitors, planning to launch HBM3e by 2025, while adopting proven bonding equipment and internal voltage conversion technologies [7][8] Group 3: China's Influence - China's influence in the storage market is expected to grow, depending on the new demand environment shaped by GenAI, technological changes, and U.S.-China policies [9] - By 2024, China's DRAM manufacturer CXMT is projected to account for 13% of global capacity, 6% of shipments, and 3.7% of revenue, with potential growth to match Micron's capacity by 2025 [12] - The Chinese semiconductor industry is working towards domestic supply chains for HBM and DRAM, but faces challenges due to U.S. regulations on advanced equipment [12][18] Group 4: Smartphone Integration - The integration of GenAI with smartphones is a significant development, requiring transformative memory solutions and altering the competitive landscape [11] - Samsung plans to improve HBM3e by Q1 2025, potentially increasing its HBM shipment from 8-9 billion gigabits to 11-12 billion gigabits [11] - China aims for domestic production of HBM3 by 2025, with a focus on self-sufficiency in GPU manufacturing and related technologies [11] Group 5: Future Trends and Innovations - The future of memory solutions will focus on balancing bandwidth, capacity, power consumption, and cost, with ARM's recent strategic adjustments influencing configurations [22] - Innovations like Processing In Memory (PIM) and Wide I/O are expected to enhance performance in high-end smartphones and other applications, although their adoption timelines vary [15][22] - The custom HBM market is projected to grow significantly, with estimates suggesting it could account for 30-40% of the overall HBM market by 2030 [29]

DeepSeek 的大语言模型（LLM）因其在性能上接近 ChatGPT ，但成本却大幅降低而受到关注。市 场的即时反应褒贬不一。虽然数据训练成本，比如数据标注和归类等方面的成本可能没有体现出 来，而这部分成本由政府支持，但 DeepSeek 在训练效率和低成本方面的优势依然十分明显。 DeepSeek 能否助力中国芯片制造？ 中国的存储芯片或已具备成本竞争力 ： 假设中国政府对构成总成本很大一部分的固定成本提供支持 ，那么与同行相比，中国已能实现有竞争力的成本。例如，2024 年第一季度 DRAM （动态随机存 取存储器）每 Gb （ 千兆字节 ）的价格为 $0.34 ，此时高价的 HBM （高带宽存储器）对平均售价 的影响较小。而韩国 DRAM 的总成本大约为售价的 67% ，约为 $0.23 ，在不计固定成本的情况 下，中国的成本可能低至 $0.20 。（不过，中国的固定成本远高于韩国。） 高效的软件为低端硬件打开市场： 中国的策略是通过规模优势弥补与竞争对手在性能上的差距。华 为最新的 GPU —— Ascend 920 支持 HBM2 和 HBM2e ，而这些对于行业同行来说已是两年前的标 准，并未 ...