[Table_Reportdate] 2026年03月19日 [证券分析师： Table_Authors] 刘思佳 S0630516080002 liusj@longone.com.cn 证券分析师： 王洋 S0630513040002 wangyang@longone.com.cn 联系人： 邓尧天 dytian@longone.com.cn [晨会纪要 Table_NewTitle] 20260319 财经要闻 晨 会 纪 要 证券研究报告 HTTP://WWW.LONGONE.COM.CN 请务必仔细阅读正文后的所有说明和声明 [table_summary] ➢ 1.寒武纪（688256）：全年业绩高增，关注下一代思元芯片商业化落地——公司简评报告 ➢ 2.上证指数短线有调整风险，但长线向好未变——技术分析上证指数简评 ➢ 1.能源基础设施遭袭，伊朗议长称对抗已上升到全新阶段 ➢ 2.美联储按兵不动，仍维持2026年降息一次的预期 重点推荐 | 1. 重点推荐 | 3 | | --- | --- | | 1.1. 寒武纪（688256）：全年业绩高增，关注下一代思元芯片商业化落地——公 | | | 司简 ...

Images of mobile devices at the Taiwan Semiconductor Manufacturing Co. (TSMC) Museum of Innovation in Hsinchu, on Tuesday, Jan. 11, 2022.Asian technology stocks fell on Thursday as Iran's latest attacks on Qatar's Ras Laffan Industrial City and a surge in oil prices rattled investor sentiment, amplifying concerns over supply chain disruptions across the semiconductor industry. Materials derived from Middle Eastern energy markets are used extensively in electronics manufacturing, from printed circuit boards ...

在半导体产业的星辰大海中， 珠海市杰理科技股份有限公司 犹如一匹深藏不露的行业黑 马。这家扎根珠海十五载的芯片设计企业，在2024年，其自主研发的TWS耳机主控芯片年 度出货量以破竹之势突破20亿颗大关，问鼎全球市场占有率榜首；更以31.20亿元的年营收 和7.91亿元的净利润，在半导体寒冬中划出一道逆势上扬的亮眼曲线。 从"零"到"百亿"：技术破局，定义行业标准 2010年，中国芯片产业尚在襁褓，杰理科技以29人团队起步，锚定系统级芯片(SoC)设计赛 道。从首款蓝牙音频芯片AC690N的突围，到 AC697N 系列横扫市场，其产品以"低功耗 +高集成度"的黄金组合，直击TWS耳机爆发式增长的需求痛点。截至2024年，公司芯片累 计销量突破120亿颗，相当于全球平均每人持有1.5颗杰理芯片。 "自研IP体系"是杰理科技的核心武器。通过15年迭代，公司掌握射频、音频、视频等模块的 全栈技术，截至2025年6月30日，杰理科技拥有授权发明专利370项(含7项境外发明专利)、 集成电路布图设计64项以及软件著作权188项。相较依赖外购IP的同行，杰理可针对ANC降 噪、主从切换等场景需求，快速优化芯片架构——这正 ...

Core Viewpoint - Nvidia's stock has been stagnant since August 2025, but the underlying business continues to grow rapidly, presenting a potential buying opportunity for investors [1][2]. Company Growth - Nvidia manufactures GPUs and related software/hardware, maintaining a leading position in the market, which allows it to command premium prices over cheaper alternatives [4]. - The introduction of Rubin GPUs, which reduce inference token costs by 10 times and require four times fewer units for AI training compared to Blackwell GPUs, is expected to enhance Nvidia's revenue [5][6]. - Nvidia's revenue grew by 73% in Q1 FY 2026, with expectations of 77% growth in the following quarter, driven by increased spending from hyperscalers and strong demand for its products [10]. Market Dynamics - AI hyperscalers are projected to spend around $650 billion on AI data centers, with Nvidia estimating that global data center capital expenditures could reach $3 trillion to $4 trillion by 2030 [8]. - The potential return of sales to Chinese companies could provide a significant boost to Nvidia's growth, with previous expectations of $8 billion in export sales before restrictions were imposed [9]. Stock Valuation - Despite positive growth indicators, Nvidia's stock trades at 21.8 times forward earnings, slightly above the S&P 500's 21.2 times, suggesting it may be undervalued given its growth prospects [11]. - The market currently reflects a pessimistic outlook for Nvidia, but expectations are that this sentiment will shift as 2026 progresses, making it a compelling investment opportunity [14].

Core Viewpoint - Micron Technology's stock has declined following a strong earnings report due to concerns over helium supply disruptions caused by an attack on Qatar's energy infrastructure, which is critical for semiconductor manufacturing [1][2][6]. Group 1: Financial Performance - Micron reported a nearly 200% year-on-year increase in revenue, reaching $23.86 billion for the quarter [1][6]. - Despite the strong earnings, Micron's stock fell after hours, indicating investor concerns about future costs and supply chain issues [1][6]. Group 2: Impact of Qatar's Energy Sector - The attack on Qatar's Ras Laffan industrial complex poses a direct threat to Micron, as Qatar produces approximately one-third of the world's helium, essential for semiconductor manufacturing [2][3]. - The semiconductor industry typically maintains a helium inventory of two to six weeks, and the current situation may lead to supply shortages, impacting production [4]. Group 3: Cost and Margin Pressures - Experts predict that helium prices could triple due to the attack, significantly affecting the semiconductor sector's cost structure [5]. - Rising global LNG prices are expected to increase electricity and logistics costs, further squeezing Micron's margins as it ramps up high-cost production for AI [7][8]. Group 4: Market Sentiment and Recommendations - Despite a strong earnings report, Wall Street analysts suggest that Micron's stock price may have risen too much, with a consensus rating of "strong buy" but a mean target indicating a potential downside of about 13% [9].

Group 1 - Nvidia showcased the Vera Rubin POD platform at GTC, focusing on enhancing competitiveness in cluster computing and inference capabilities for Agent applications [1] - The Vera Rubin POD consists of two types of racks: MGXNVL rack for core GPU computing tasks and MGXETL rack for collaborative processing through direct interconnects [1] - A single Vera Rubin 1152 SuperPOD is composed of 16 Vera Rubin NVL72 racks, 2 Vera CPU racks, 10 Groq 3 LPX racks, 2 BlueField-4 STX storage racks, and 10 Spectrum-6 SPX network racks, highlighting a heterogeneous collaborative system architecture [1] Group 2 - The Groq3 LPX rack accelerates decoding with 256 LPU processors, 128 GB on-chip SRAM, and a bandwidth of 640 TB/s, enhancing the performance of the Vera Rubin NVL72 and LPX combination [2] - Under conditions of 400 TPS per user, the combination of Vera Rubin NVL72 and LPX can achieve up to 35 times the TPS improvement per megawatt compared to NVIDIA GB200 NVL72, making it suitable for low-latency, interactive Agent applications [2] Group 3 - The Vera CPU rack integrates 256 Vera CPUs with a high-density liquid cooling design, supporting over 22,500 concurrent reinforcement learning or agent sandbox environments for testing and validating outputs from Vera Rubin NVL72 and LPX [3]

NVIDIA 展示了一个全面的 AI 超级计算机产品矩阵，旨在覆盖训练、高吞吐量 推理和低延迟推理等不同类型的 AI 负载。该矩阵在之前发布的六款芯片（包括 Ruby GPU、Viva CPU、Spectrum-X、BlueField 和 ConnectX 等）基础上， 重点整合了 2025 年 12 月底收购的 Groq 公司的技术。具体发布了一款支持 256 颗 Groq LPU 互联的机柜产品，每个托盘包含 8 颗芯片，共计 32 个托盘。 这一举措被视为对低延迟推理能力的显著增强，完善了其针对各类 AI 工作负载 的芯片解决方案。 Groq LPU 的整合与近期崛起的"龙虾（Agentic AI）"应用之间存在何种 关联，这预示着怎样的行业趋势？ GTC 2026 现场解读：AI 基础设施新范式 20260318 摘要 NVIDIA 战略从芯片销售转向 AI 工厂平台，2026 年重点聚焦极低延迟 的生成式 AI 推理场景。 整合 Groq 技术推出支持 256 颗 LPU 互联的机柜，旨在通过软硬件协同 支撑 Agentic AI 落地。 互联技术路径明确：中短期光铜并存，下一代 Fairwood ...

Summary of Key Points from the Conference Call Industry Overview - The conference call primarily discusses developments in the technology sector, particularly focusing on artificial intelligence (AI), optical interconnects, and memory pricing trends. Key Companies Mentioned - **Nvidia (NVDA)**: Received approval from Beijing to sell H200 AI chips in China and is adapting Groq chips for the Chinese market [6][8] - **Alibaba (BABA)**: Increasing prices on AI computing chips and cloud storage products by up to 34% to capitalize on surging demand [9][10] - **Micron Technology (MU)**: Scheduled to release earnings, indicating potential market movements [3] - **LITE**: Identified as a core player in AI optical interconnects with long-term upside potential [14] - **Trade Desk (TTD)**: Downgraded to Neutral from Buy by Rosenblatt [46] - **Circle (CRCL)**: Price target raised due to confidence in blockchain infrastructure [58] Core Insights and Arguments - **Nvidia's Market Strategy**: Nvidia is preparing to leverage its AI chips for inference tasks in China, indicating a strategic move to capture market share in a growing sector [8] - **Alibaba's Price Adjustment**: The price hike reflects a broader trend among tech giants to monetize AI investments amid increasing demand [10] - **Optical Interconnects**: LITE is positioned across key segments of AI optical interconnects, with significant growth potential driven by the transition from copper to optical technologies [15][17] - **Memory Pricing Trends**: Bernstein reports that memory pricing is tracking above expectations, with some SKUs experiencing near-100% quarter-over-quarter increases due to tight supply [40][41] Additional Important Points - **Market Sentiment**: JPMorgan notes strong demand for optical networking despite supply constraints, highlighting the importance of lasers and key components [26][27] - **CPO Adoption**: UBS indicates that co-packaged optics (CPO) adoption is delayed, with a hybrid approach being more prevalent in the near term [22] - **Microsoft's Legal Concerns**: Microsoft is considering legal action against Amazon and OpenAI over a $50 billion cloud deal, which may impact its exclusive agreements [31][32] - **Financial Performance**: DOCU reported strong revenue and billings, exceeding market expectations, with a focus on identity and access management (IAM) as a growth lever [42][44] Conclusion The conference call highlights significant developments in the technology sector, particularly in AI and optical interconnects, with companies like Nvidia and Alibaba taking strategic steps to capitalize on market demand. The memory pricing landscape is also showing promising trends, while legal and competitive dynamics are evolving in the cloud services space.

通信&人工智能"周周谈" — GTC 大会总结 20260318 摘要 NVIDIA 定义数据中心为"Token 工厂"，核心逻辑从技术演进转向每 瓦电力 Token 产出最大化，驱动 AI 服务实现非线性定价弹性。 官方指引 2027 年底 Rubin 与 Blackwell 平台累计收入达 1 万亿美元， 且未计入 GraceCPU、GrokLPU 及 RubinUltra 等潜在增量。 AI 应用 Token 消耗量与年化收入呈指数级增长，Azure AI 业务及 OpenAI 收入在两个月内接近翻倍，NIM 被定位为驱动企业级应用的新 操作系统。 Rubin 平台采用极致协同设计，整合 7 种芯片与 5 种机柜，其中 CPU 机 架为今年纯新增量，专为 Agent 时代高单线程性能需求设计。 硬件规格大幅升级：计算托盘 PCB 层数需达 50 层以上，Rubin 平台 100%采用液冷散热，机架安装时间由 2 天缩短至 2 小时。 混合推理架构实现任务拆分：RubinGPU 负责高吞吐预填充，GrokLPU 负责低延迟逻辑任务，后者延迟较传统方案降低约 10 倍。 连接方案采取差异化策略：主流 72 ...

Summary of Key Points from the Conference Call Industry Overview - The memory cycle is increasingly influenced by the shift to agentic reasoning, with OpenClaw accelerating this transition. The intelligence layer is growing exponentially, making memory a critical bottleneck in AI chip development [1][2] Core Insights - **Heterogeneous Memory Architecture**: The AI computing bottleneck is shifting towards multi-tiered memory systems, with DRAM being a primary beneficiary as AI workloads become more data-intensive [2] - **DRAM Pricing Trends**: - 2Q26 DRAM DDR5 pricing is expected to start at +50% QoQ due to low inventory levels among hyperscale customers and producers. DDR4 prices are projected to increase by +30-40% [3] - Samsung is leading in HBM4 ramp-up, with HBM3E pricing increasing in 2026 due to renewed contracts [3] - **Earnings Estimates**: - SK hynix's EPS is raised by 24-32% for 2026-27, reflecting revised DRAM pricing. The price target for SK hynix shares is increased to W1,300,000 [4] - Samsung's smartphone shipments forecast for 2026 has been revised from -6% YoY to +5% YoY, driven by strong pre-orders and market share gains [4] Market Dynamics - **Stock Implications**: Companies involved in AI chip design, manufacturing, and semiconductor equipment are expected to benefit from the current market dynamics. There is significant upside potential as earnings growth is anticipated amid capacity constraints [5] - **Demand Evidence**: Intel and AMD report high demand for their server processors, indicating a shift towards specialized CPUs that manage extensive memory tiers [17][20] Technological Developments - **Agentic AI Workloads**: The bottleneck in AI infrastructure is shifting towards memory, with agentic AI requiring more DRAM compared to traditional models. This shift is driven by new solutions like OpenClaw, which perform complex tasks rather than just generating responses [10][12] - **Memory Requirements for OpenClaw**: The memory needs vary significantly based on usage mode, with local model mode requiring substantial DRAM (32GB recommended for larger models) [19][21] Pricing Forecasts - **DRAM and NAND Pricing**: - DRAM pricing is expected to rise significantly, with DDR4 and DDR5 prices projected to increase by 28-33% in 2Q26. - NAND eSSD pricing could double in 2Q26, while consumer products may see price hikes of 60-80% [3][26] - **Long-term Pricing Trends**: The pricing forecasts for various memory types indicate a strong upward trend, with total NAND flash expected to rise by 30-35% in 2Q26 [42] Financial Projections - **Earnings Revisions**: - Sales and operating profit estimates for SK hynix have been revised upwards, with significant increases in DRAM and HBM sales expected [46] - The net profit for SK hynix is projected to rise by 24% for FY26, with EPS estimates reflecting a similar growth trajectory [45] Conclusion - The memory industry is poised for significant growth driven by the increasing demand for AI applications and the transition to more complex memory architectures. Companies like SK hynix and Samsung are expected to benefit from these trends, with revised earnings estimates reflecting a positive outlook for the sector [4][5][44]