本文是由投资/产品专员而非分析师撰写的文章汇编。 它不构成研究报告，也不应被解释为研究报告，也不旨在提供 专业、投资或任何其他类型的建议或推荐。 CSIWM 个股点评 2026 年 1 月 9 日 摘要 中信证券财富管理与中信里昂研究观点一致。根据中信里昂研究在 2026 年 1 月 8 日发布的题为《Supersize profits》 的报告，过去一个月海力士 2026 年营业利润的市场一致预期已上调 30%，自 9 月以来累计上调 120%。分析指出， 由于数据中心客户需求保持强劲，DRAM 供应商积极提价，考虑到供应商库存处于极低水平，这一趋势预计将在未 来几个季度持续。 SK 海力士 000660 KS 本文内容由 Ryan Lee (李昊谦) 提供 中信证券财富管理(香港) 产品及投资方案部 韩国科技行业 电话：(852) 2237 9250 / 电邮：wminvestmentsolutions@citics.com.hk 利润超预期增长 目前海力士 DRAM 和 NAND 库存分别仅为 2 周和 6 周，在旺盛需求与供给受限的双重推动下，内存市场已完全转向 卖方市场。尽管大客户要求供应商签订长期 ...

2026年初，全球内存行业再次迎来高爆发，"超级存储周期"的提法再次被摆放到前台。 美银发布一份研究报告指出，DRAM将在2026-27年开启史上最强超级周期，全球DRAM销售额预计从 2025年的1303亿美元飙升至2026年的2098亿美元，同比暴涨61%。 美银预计，2027年这个数字将进一步增至2314亿美元。趋势背后，AI需求爆发、供应结构性短缺与技 术迭代形成三重共振。 在报告中，美银大幅调高了SK海力士、三星电子、南亚科技等核心标的目标价，认为内存行业正从传 统周期性产业向AI驱动的成长性赛道蜕变。 01 超级周期的五大核心支撑 美银报告认为，内存行业此前深陷"扩产-过剩-降价-减产"的周期循环，但2026年启动的超级周期呈现本 质不同——AI带来的刚性需求与供应端的约束形成长期错配，推动行业进入高质量增长阶段。 首要驱动是AI内存需求的爆发式增长 以英伟达Rubin Ultra GPU、谷歌TPU v7为代表的AI加速器，对HBM(高带宽内存)、GDDR7等高端内存 的搭载量呈数倍提升，单台AI服务器的HBM搭载量已从早年的340GB增至1.17TB，英伟达下一代Rubin Ultra更是有望 ...

公众号记得加星标⭐️，第一时间看推送不会错过。 据传华硕将于 2026 年进军 DRAM 制造领域，这将确保其 PC 产品线拥有稳定的内存供应。 当前的内存危机已经影响到个人电脑行业的方方面面，而个人电脑制造商对此却束手无策。大多数制 造商已经提高了产品价格，而内存短缺将导致未来几年产品交付延迟。 但 看 来 ， 一 家 大 型 PC 制 造 商 正 计 划 进 军 DRAM 市 场 ， 以 应 对 内 存 短 缺 问 题 。 据 波 斯 科 技 媒 体 Sakhtafzarmag爆料，华硕计划大规模进军DRAM市场，最早可能在2026年实现。这家媒体此前曾准 确爆料过AMD和英特尔的CPU相关信息。不过，我们建议读者对此消息持保留态度。 回到之前的传闻，据称如果内存价格和供应无法恢复正常，华硕计划在2026年第二季度末之前建立专 门的DRAM生产线。目前的报告预测，内存短缺将持续到2027年底，甚至可能延续到2028年。华硕 作 为 全 球 最 大 的 PC 厂 商 之 一 ， 确 实 有 能 力 进 入 DRAM 市 场 ， 但 即 便 如 此 ， 为 他 们 专 门 建 立 一 座 DRAM生产工厂对他 ...

"这是我从事存储行业以来见过的最猛一次涨价。"一名资深的存储产业高管感慨道。 10月以来，受AI需求推动，三星、海力士、美光、闪存等巨头不断上调产品价格，存储行业开启了"超 级周期"。期间，三星电子等一度暂停DDR5 DRAM合约报价，引发供应链"断粮"危机。 瑞银预计，今年第四季度，DDR（双倍数据率内存）合约定价预计季增35%、NAND（闪存）价格上涨 20%，这一轮涨价幅度远超此前预期。不仅如此，2026年第一季的DDR合约价，更将进一步上涨30%， NAND价格上涨20%。 相关影响很快传导至A股市场。Wind数据显示，今年下半年以来，香农芯创、东芯股份、江波龙、佰维 存储股价分别上涨了294.51%、291.19%、195.51%、67.17%。 "存储是个强周期行业，大概三年一个周期，但AI需求至少把周期拉长了一年，现在行业一致判断行情 至少会持续到2026年底。这里有个关键变量，全球存储龙头在之前的下跌周期里，持续降价，把存储价 格压到了历史极低值，现在存储价格反弹上来也不扩产，这就导致市场焦虑性扫货，进一步助推了涨 价。"深圳一家存储行业上市公司高管李森（化名）在接受21世纪经济报道记者采访时 ...

当AI 应用逐步从「训练」转向「推理」，对于高速资料储存与即时存取的需求大幅提升，必须仰 赖HBM 等记忆体，将资料持续供应给GPU 进行运算。 即便通用记忆体的效能不及HBM，市场对高效能通用记忆体的需求仍快速成长。在推理初期阶 段，工作负载多由GDDR7、LPDDR5X 等通用DRAM 处理，而HBM 则主要保留给更密集的推理 任务。 NVIDIA 在以推理为主的AI 加速器中采用GDDR7，便是一项代表性案例。 记忆体业者也计划透过开发AI 导向的高效能产品，延续「记忆体为核心」的产业趋势。举例来 说，记忆体内运算（Processing-In-Memory，PIM）技术，让记忆体可承担部分原本由GPU 执行 的运算工作。报导也指出，垂直通道电晶体（VCT）DRAM 与3D DRAM 等新技术，透过在更小 面积中储存更多资料来提升密度，预期将陆续导入市场。 如果您希望可以时常见面，欢迎标星收藏哦~ 随着人工智能需求快速攀升，记忆体价格大幅上涨，半导体产业的获利结构正出现变化。根据《韩 国经济新闻》报导，三星电子的记忆体部门，以及SK 海力士，预期在2025 年第四季的毛利率将 超越台积电，这将是自2018 ...

Core Viewpoint - The storage industry is experiencing a super cycle driven by the explosive demand for AI, leading to significant structural changes and price increases in storage products [2][3]. Group 1: Market Dynamics - The storage market is currently characterized by supply shortages and rising prices, with major players like Samsung, SK Hynix, and Micron adopting different strategies to adapt to these changes [2]. - HBM (High Bandwidth Memory) supply is tight, with SK Hynix's capacity largely locked in by AI clients until the end of 2026, leaving less than 10% of the market share for smaller manufacturers [2]. Group 2: Samsung's Strategy - Samsung plans to reduce HBM production capacity by 30%-40% to focus on general DRAM production, which is expected to yield higher profit margins, exceeding 60% for products like DDR5 and LPDDR5X [4][5]. - The company is also converting NAND flash production lines to DRAM to meet the growing demand for large-capacity products in data centers [5][6]. - Samsung aims to enhance its production capacity for 1cnm DRAM, targeting an increase to 200,000 wafers per month by the end of 2026 [6]. Group 3: SK Hynix's Approach - SK Hynix plans to double its DRAM capacity investment, focusing on HBM and data center applications, while also significantly increasing its 1c DRAM production capacity by 8 to 9 times [7][8]. - The company expects its HBM sales to generate approximately 25 trillion Korean won in operating profit, a nearly 50% increase from the previous year [7][8]. - SK Hynix is balancing its focus between HBM and general DRAM to meet the rising demand from AI applications, with a significant portion of its production dedicated to data center products [9][10]. Group 4: Micron's Shift - Micron is decisively shifting its focus away from consumer-grade products, terminating its mobile NAND development and reallocating resources to higher-margin HBM and enterprise DRAM products [11][12]. - The company has signed a three-year supply agreement with NVIDIA for HBM chips, indicating a strong commitment to the data center market [12]. - Micron's strategy includes building new factories to enhance HBM production capabilities, with plans to increase HBM capacity fourfold by 2026 [13][21]. Group 5: Underlying Factors Driving Change - The strategic shifts among storage giants are driven by profit margins, market demand, and technological adaptability, with each company responding differently based on its market position and technological capabilities [15][16]. - The demand for high-capacity, cost-effective storage solutions is surging due to AI applications, prompting companies to reassess their production strategies [16][17]. - The complexity of HBM production and the slow yield improvement are pushing companies to optimize their existing capacities and focus on more efficient production processes for general DRAM [17][21]. Group 6: Industry Impact - The transition in production strategies among storage giants is reshaping the entire industry, with rising memory prices affecting consumer electronics and leading to a re-evaluation of supply chains [22][23]. - The supply-demand imbalance is expected to persist until 2027, as new production lines come online, fundamentally altering the storage market landscape [23].

Core Viewpoint - JEDEC is nearing the completion of the SPHBM4 standard, which aims to provide HBM4-level bandwidth with a narrower 512-bit interface, higher capacity, and lower integration costs through compatibility with traditional organic substrates. However, it is unlikely to replace GDDR memory [2][4]. Group 1: SPHBM4 Overview - SPHBM4 proposes to reduce the HBM4 memory interface width from 2048 bits to 512 bits while maintaining the same total bandwidth through a 4:1 serialization method [4]. - The design of SPHBM4 prioritizes performance and capacity over power consumption and cost, making it less suitable for gaming GPUs compared to GDDR7 [6]. Group 2: Cost and Integration Challenges - Although SPHBM4 is cheaper than HBM4 or HBM4E, it still requires stacked HBM DRAM chips, which are larger and more expensive than commodity DRAM ICs, leading to potential cost increases when replacing multiple GDDR7 chips with a single SPHBM4 chip [6]. - The integration of SPHBM4 on traditional organic substrates allows for 2.5D integration without expensive silicon interposers, potentially lowering costs and increasing design flexibility [8]. Group 3: Technical Considerations - SPHBM4 aims to achieve four times the memory capacity of HBM4 under the same chip conditions, but practical design often involves trade-offs between memory capacity, computational power, and chip functionality [5]. - The complexity of a 512-bit memory bus remains a challenge, but JEDEC's support for traditional organic substrates may alleviate some layout constraints in large packages [8].

类似的局面并非首次上演。2018年比特币暴涨时，显卡工厂产能几乎全数被区块链矿场吞噬，最终造成零售缺货、价格飞涨。 今年则在AI算力战下重复发生。行业人士承认，如今各大代工厂即便满负荷运转，仍填不满数据中心和企业对DRAM、NAND 的胃口，比如十铨管理层就坦言，现在不只是台式机内存，连GDDR6、GDDR7、LPDDR5几乎都买不上，在东南亚的部分电 商平台，发货时间甚至被拉长至半年以上。 事实上，冷数据存储方式也在近期出现变化。由于硬盘制造商未扩产，越来越多的数据中心将冷数据转移至固态硬盘，这进一 步加剧了NAND闪存的紧俏状况。假如某地突然增加机械硬盘生产，或许对固态盘供应有正向影响，但目前看来距离这种转折 还很遥远。 也有极少数反向案例可供参考。去年日本市场的DDR4库存意外过剩，偶然导致本地用户短时间享受了平价内存。然而仅两个 月之后，全球行情一体化依旧将价格拉升到高位，证明区域性"小阳春"难以长久。从整体趋势看，只要AI赛道继续狂奔，大众 购买力就只能被动等待新一轮市场变革带来的再平衡。 未来几年，除非有技术突破降低DRAM/NAND生产难度，否则对于绝大多数玩家和消费级用户，攒机和升级的窗口期只 ...

Core Insights - AMD plans to increase the price of Radeon GPUs by 10% in early 2026 due to rising costs and high demand for computing, particularly driven by AI [1][2] - The semiconductor industry is experiencing a surge in memory prices, with DRAM and GDDR6 costs significantly impacting profitability and pricing strategies for companies like AMD and Nvidia [5][6] Group 1: AMD's Pricing Strategy - AMD's CEO Lisa Su highlighted that the demand for compute resources is unprecedented, with AI demand being described as "insatiable" [1] - The company has faced increased costs for components, particularly DRAM and VRAM, which have led to a necessary price hike for consumers [2][4] - By the end of 2025, DRAM contract prices had increased significantly, contributing to the overall rise in GPU prices [6] Group 2: Market Dynamics - The competition in the semiconductor market is intensifying, with both AMD and Nvidia facing supply chain challenges, particularly in memory production [3][4] - The shift in focus towards AI-class components has resulted in a "memory bull market," affecting the availability and pricing of consumer-grade memory [5] - The next-generation GDDR7 pipeline remains tight, further pressuring current-generation memory prices and impacting retail pricing strategies for 2026 [7]

Core Viewpoint - SK Hynix is significantly increasing its production capacity for sixth-generation 10nm DRAM, betting on the AI inference application market, with plans to raise monthly production from approximately 20,000 wafers to between 160,000 and 190,000 wafers, representing an increase of 8 to 9 times, which will account for over one-third of its total DRAM capacity [1][2]. Group 1: Production Capacity Expansion - SK Hynix plans to enhance its advanced memory chip production capacity, focusing on the 1c DRAM technology node, with expectations to add 140,000 wafers per month at its Icheon plant, which is seen as a minimum increase [3]. - The company has achieved a yield rate of over 80% for 1c DRAM, which is primarily used for manufacturing the latest general DRAM products like DDR5, LPDDR, and GDDR7 [3]. Group 2: Market Demand Shift - The strategic shift of SK Hynix is driven by the transition of AI applications from training to inference, leading to a surge in demand for cost-effective general DRAM, which is expected to grow at a rate comparable to HBM [5]. - Major tech companies, including NVIDIA, Google, and OpenAI, are developing custom AI accelerators that integrate a significant amount of general DRAM, indicating a shift in market demand structure [5][6]. Group 3: Financial Outlook - Analysts predict that SK Hynix's facility investment will exceed 30 trillion KRW next year, a substantial increase from the estimated 25 trillion KRW this year, with operating profit expected to surpass 70 trillion KRW, setting a historical record [2][9]. - The company has successfully negotiated a price increase of over 50% for HBM4 to above $500 per chip, which is expected to bolster its profitability [8][9]. Group 4: Competitive Landscape - Samsung is also expanding its 1c DRAM production capacity, aiming to reach 200,000 wafers per month by the end of 2026, reflecting the competitive nature of the next-generation memory market [10][13]. - The ongoing expansion by both SK Hynix and Samsung highlights the strong confidence in the demand driven by AI applications, as the supply of DRAM is currently unable to meet the rapid growth in demand [11][13].