3 月 GTC 大会热点展望 - 掘金 AIDC 系列电话会议 20260301 摘要 GPU 功率提升推动三次电源架构变革，模块化和 IVR 方案成潜在趋势。 传统分立式方案面临器件堆叠、布线损耗等瓶颈，模块化供电通过立体 PCB 集成电感、电容等，节省空间并提升耐流，IVR 方案则通过芯片内 部集成电源，降低电流需求，GTC 大会或将催化相关技术落地。 三次电源模块化升级对 PCB 工艺提出更高要求，需在设计阶段协同电源 设计，并引入新材料和压合工艺。中富电路已进入 MPS、伟创力等三次 电源供应链，威尔高也在推进。电源 PCB 市场规模相对较小，但与客户 配合迭代设计至关重要，较早布局者有望脱颖而出。 英伟达 Rubin Ultra 代际基本确定导入 HVC（800V）方案，以降低电 流和损耗。麦格米特已在英伟达供应链内，配合进行产品开发与预研， 进度领先大陆厂商。欧陆通等谷歌链电源厂商以及铂科新材等电感材料 企业也有望受益。 GTC 大会重点展示 Quantum 3,400 以及以太网 6,800、6,810 三款 CPO 交换机，量产时间预计在 2026 年底至 2027 年初。台积电良率提 升至 ...

Investment Rating - The industry investment rating is "Positive" (maintained) [1] Core Insights - Nvidia's performance exceeded expectations, with Q4 FY2026 revenue reaching $68.1 billion, a year-on-year increase of 73% and a quarter-on-quarter increase of approximately 19%. The data center business revenue was $62.3 billion, up 75% year-on-year and 22% quarter-on-quarter, driven by the continuous release of Blackwell and Blackwell Ultra products [5][13] - DeepSeek, in collaboration with Tsinghua University and Peking University, introduced the DualPath system to address storage bandwidth bottlenecks in LLM inference, achieving an average throughput increase of 1.96 times [6][16] - The global AI industry is expected to resonate, with significant capital expenditure increases from major players like Google and Meta, indicating a robust growth trajectory for AI-related sectors [14][18] Summary by Sections Nvidia Performance - Nvidia's Q4 FY2026 revenue was $68.1 billion, with a 73% year-on-year growth and a 19% quarter-on-quarter growth. Data center revenue reached $62.3 billion, marking a 75% year-on-year increase and a 22% quarter-on-quarter increase. The growth was primarily due to the release of Blackwell and Blackwell Ultra products, with the Grace Blackwell system contributing approximately two-thirds of the data center revenue [5][13][14] DeepSeek Innovation - DeepSeek's DualPath system addresses the KV-Cache storage bottleneck in LLM inference, significantly improving online inference throughput by 1.96 times. This innovation is seen as a milestone in AI computing, enhancing hardware utilization and facilitating the commercial application of large models [6][16] AI Industry Outlook - The report highlights a strong outlook for the AI industry, with major companies like Google and Meta increasing their capital expenditures significantly. This trend is expected to support the growth of AI computing and related technologies [14][18] Communication Data Tracking - As of December 2025, China had 4.84 million 5G base stations, with 1.204 billion 5G mobile users, reflecting an 18.74% year-on-year growth. However, 5G mobile phone shipments decreased by 27.3% year-on-year [28][30][41]

[Table_Page] 行业专题研究|国防军工 2026 年 2 月 28 日 证券研究报告 [Table_Title] 国防军工行业 | | AIDC | 电源的"最后一公里"，板载电源的高密高集成化革命 | | | | --- | --- | --- | --- | --- | | 分析师： [Tabl | 孟祥杰 | 分析师： 陈昕 | 分析师： | 吴坤其 | | e_Author] | SAC 执证号：S0260521040002 | SAC 执证号：S0260522080008 | | SAC 执证号：S0260522120001 | | | SFC CE.no: BRF275 | SFC CE.no: BWV823 | | SFC CE.no: BRT139 | | | 010-59136693 | 010-59136699 | | 010-59133689 | | | mengxiangjie@gf.com.cn | gfchenxin@gf.com.cn | | wukunqi@gf.com.cn | [Table_Summary] 核心观点： 识别风险，发现价值 请务必阅读末页的免责声明 ...

徐强 国泰海通分析师: 好的，各位投资者大家下午好，欢迎参加我们国泰海通电新团队春节不打烊的系列电话会 议。我是国泰海通的环保电新首席徐翔，今天我和我们孙潇浩孙老师一起给各位领导做去 做一些 AIDC 方面的汇报。明天就是这个第一个交易日了，那在此，我们国泰海通电新团 队祝大家开市大吉，业绩长虹。这个假期期间，我们围绕了 AIDC、太空算力、太阳翼、 锂电供应链，然后开展了一系列的这个电话会议和专家的这个交流。那这今天下午的电话 会议主要是汇报 A I D C 硬件端的这些投资机会。 然后同时，今天晚上我们还安排了这个卫星载荷的专家，然后来交流一下太空算力以及太 阳翼的这一方面的一个，这个投资内容。今晚 8 点，也欢迎大家这个就是按时接入。首先 我们一直从这个去年开始吧，去去年，24 年底，前年开始，就一直在推荐 AIGC 的这些投 资机会。市场，其实对于 IPC 的这些投资的持续性，持，一直是存，一直在质疑当中，这 个不断的这个提升预期的。那包括像现在。大家对于这个北美地区，像 Oracle 他们的一 个投资的持续性，以及北美这些算力是否存在泡沫，一直有比较大的一个争议。 但是，近期我们在在使用，包括像 O ...

2026 年，AI 数据中心（AIDC）行业将继续保持高景气度，值得投资者重点关 注。回顾 2025 年，AIDC 板块经历了两大波行情：第一波发生在 2 月至 3 月 中旬，由海外头部 CSP 厂商和国内互联网巨头如阿里、字节、腾讯等发布的数 据中心资本支出预期超出市场预期所驱动；第二波发生在 7 月中旬至 9 月，由 于海外 CSP 厂商在中报财报中进一步上调资本开支预期，以及 H20 解禁事件 和国内互联网大厂三季报上调预期共同催化。 年底出现的小行情主要由谷歌推 出 TPU 及 NV 的业绩超预期、阿里上调资本开支预期以及 NV 举办电力设备短 缺峰会等事件驱动。从这些表现来看，股价变动的核心因素包括国内外厂商上 电力设备 2026 年年度策略：AI 大周期开启，海外扩展空 间广阔 20260223 摘要 数据中心资本支出受海外 CSP 厂商和国内互联网巨头驱动，尤其在 2 月 至 3 月中旬及 7 月中旬至 9 月，厂商上调 CAPEX 预期是股价变动的核 心因素，订单及业绩超预期和技术突破亦有影响。 液冷和 SST 板块分别在 7 月中旬至 9 月及 10 月中旬至 10 月底走出独 立行情，液冷 ...

Summary of Key Points from Conference Call Industry Overview - The conference call focuses on the domestic AI computing power industry, highlighting advancements in AI video generation and the increasing demand for computing resources during peak traffic events like the Spring Festival [1][2]. Core Insights and Arguments - The AI video generation model Seedance 2.0, developed by ByteDance, has evolved into a "director-level" tool, capable of producing multi-scene, movie-quality videos with original audio in just 60 seconds. This advancement transforms AI video into a significant productivity tool [1][2]. - During the Spring Festival, ByteDance leveraged high visibility from the New Year's Gala to achieve peak concurrent usage, emphasizing the growing demand for computing power [1][2]. - The "Doubao" AI cloud partnership with CCTV for the Spring Festival resulted in a total of 1.9 billion AI interactions, with peak activity reaching 633 billion tokens per minute. This indicates a substantial increase in consumer engagement and computing requests [3]. - Major companies like Tencent and Alibaba are competing for consumer entry points, with Tencent distributing 1 billion yuan in red envelopes and Alibaba launching a 3 billion yuan free order campaign, which enhances the consumption ecosystem and drives up computing demand [3]. - The domestic AI landscape is characterized by a strong demand-side pull, with the native adaptation of the Zhiyuan GLM-5 model to domestic GPU ecosystems, optimizing compatibility with seven major domestic chip platforms [3]. Important but Overlooked Content - Key companies to watch in the server segment include Inspur Information, ZTE, and Unisplendour. In the networking switch sector, focus on Huaqin Technology, Shengke Communication, Ruijie Networks, and StarNet Ruijie. For liquid cooling and power solutions, companies like Invec, Shenling Environment, and Kexin Innovation Source are noteworthy [4]. - A risk factor mentioned is the potential for industry progress to fall short of expectations, which could impact the overall growth and development of the AI computing power sector [5].

Summary of Key Points from the Conference Call Industry and Company Involved - The conference call focuses on the domestic AI video generation industry, specifically highlighting ByteDance's AI video generation model Seedance 2.0 and its impact on the market [1][2]. Core Insights and Arguments - **Advancement of Domestic Video Generation Models**: ByteDance's Seedance 2.0 has evolved into a "director-level" tool, capable of generating multi-shot, movie-quality videos with original audio in just 60 seconds. This represents a significant breakthrough in AI video production, transforming it into a productive asset [1][2]. - **Increased Demand for Computing Power During Spring Festival**: The competition for traffic during the Spring Festival has amplified the demand for computing power. ByteDance leveraged the high visibility of the Spring Festival Gala to achieve peak concurrent usage [2]. - **AI Interaction Metrics**: On New Year's Eve, as the exclusive AI cloud partner for the Spring Festival Gala, the AI interaction volume reached 1.9 billion, with a peak of 63.3 billion tokens per minute. This indicates a substantial increase in user engagement and computing resource consumption [2]. - **Competitive Strategies of Major Companies**: Tencent distributed 1 billion yuan in red envelopes to stimulate social engagement, while Alibaba's 3 billion yuan promotional campaign integrated its ecosystem, creating a consumable closed loop. These strategies highlight the fierce competition among major players for consumer entry points, driving up concurrent requests and token consumption [2]. - **Acceleration of the Domestic Large Model and Computing Power Cycle**: The Zhipu GLM-5 model is natively adapted to the domestic GPU ecosystem, having undergone deep optimization from the core to the inference framework, ensuring compatibility with seven major domestic chip platforms. This reflects the robust demand-side traction in the domestic AI sector and the gradual improvement of the ecosystem between domestic large models and computing power [2]. Other Important but Potentially Overlooked Content - **Key Companies to Watch**: The call emphasizes several companies in the server, networking, liquid cooling/power, and connector sectors, including Inspur Information, ZTE, Unisplendour, Huaqin Technology, and others, indicating a broad interest in the supporting infrastructure of the AI industry [3]. - **Risk Warning**: There is a cautionary note regarding the potential risks associated with the industry, particularly the possibility of slower-than-expected progress in technological advancements [4].

Summary of Key Points from Conference Call Industry Overview - The conference call focuses on the domestic AI computing power industry, highlighting advancements in AI video generation and the increasing demand for computing resources during peak traffic events like the Spring Festival [1][2]. Core Insights and Arguments - The AI video generation model Seedance 2.0, developed by ByteDance, has evolved into a "director-level" tool, capable of producing multi-scene, movie-quality videos with original audio in just 60 seconds. This advancement transforms AI video into a significant productivity tool [1][2]. - During the Spring Festival, ByteDance leveraged high visibility from the New Year's Gala to achieve peak concurrent usage, emphasizing the growing demand for computing power [1][2]. - The "Doubao" AI cloud partnership with CCTV for the Spring Festival resulted in a total of 1.9 billion AI interactions on New Year's Eve, with peak activity reaching 633 billion tokens per minute. This indicates a substantial increase in consumer engagement and computing requests [3]. - Major companies like Tencent and Alibaba are competing for consumer entry points, with Tencent distributing 1 billion yuan in red envelopes and Alibaba launching a 3 billion yuan free order campaign, which enhances the consumption ecosystem and drives up computing demand [3]. - The domestic AI landscape is characterized by a strong demand-side pull, with the native adaptation of the Zhiyuan GLM-5 model to domestic GPU ecosystems, optimizing compatibility with seven major domestic chip platforms [3]. Important but Overlooked Content - Key companies to watch in the server sector include Inspur Information, ZTE, and Unisplendour; in the networking sector, focus on Huaqin Technology, Shengke Communication, Ruijie Networks, and StarNet Ruijie; for liquid cooling and power solutions, consider Invec, Shenling Environment, and Kexin Innovation Source [4]. - A risk factor mentioned is the potential for industry progress to fall short of expectations, which could impact the overall growth and investment potential in the sector [5].

深度【天风电新】AI&电力（4）-节流：电源、液冷、PCB上游的投资机会 AIDC产业链迎来量价齐升的通胀红利，电源技术迭代、HVDC架构升级、液冷成为刚需、CCL上游高端材料国产 化突破，国产厂商迎来成长期突围与出海机遇。 AI服务器电源 AI服务器功率密度提升驱动电源技术迭代，带动产品需求量与价值量同步提升，业绩即将进入加速释放阶段。 深度【天风电新】AI&电力（4）-节流：电源、液冷、PCB上游的投资机会 AIDC产业链迎来量价齐升的通胀红利，电源技术迭代、HVDC架构升级、液冷成为刚需、CCL上游高端材料国产 化突破，国产厂商迎来成长期突围与出海机遇。 重点推荐【优优绿能】、【通合科技】。 【科士达】：预计26年主业净利润10亿元，26H2 HVDC代工有望放量，增厚利润2-4亿元。 【江海股份】：27年需求预计为26年6倍，6年AIDC电容收入约10亿，26年AIDC贡献3亿净利润。 【铂科新材】：存储涨价带来稀缺通胀机会，DDR芯片电感市场空间约50亿元，预计26、27年归母净利6、8.4亿 元，对应PE 37、27X。 AI服务器电源 AI服务器功率密度提升驱动电源技术迭代，带动产品需求量与价值 ...

Investment Rating - The report maintains a "Recommended" rating for the industry, indicating an expected relative performance exceeding 15% compared to the benchmark index over the next 12 months [8]. Core Insights - The launch of Seedance 2.0 by ByteDance marks a significant advancement in AI video generation, enabling users to create videos with simple text prompts, which is expected to enhance user engagement and frequency of use over the long term [3][5]. - The model's capabilities include multi-modal input acceptance, narrative coherence, and audio-visual synchronization, addressing key industry challenges such as character consistency and audio-visual mismatch [4][5]. - The demand for computational power is projected to increase exponentially due to the anticipated high concurrency from both consumer (C-end) content creation and business (B-end) API calls [6]. Summary by Sections Seedance 2.0 Launch and Features - Seedance 2.0 is integrated into various platforms, allowing users to generate short videos easily, which is expected to drive higher usage rates [3]. - The model supports multiple input types and can generate videos with synchronized audio, enhancing the creative process for users [4]. Computational Demand Projections - The report estimates that by October 2025, the monthly usage of AI-generated content on the platform will reach 63,900 times, with a total usage time of 1,668.2 hours [6]. - The complexity of video generation is significantly higher than that of text and images, leading to a substantial increase in computational requirements [6]. - Projections indicate that the number of required H100 GPUs for video generation could range from 2.4 million to 4.284 million under different scenarios, with a market space for H100 GPUs estimated between $4.8 billion and $8.568 billion [14][12]. Long-term Market Opportunities - The report highlights the potential for domestic computational power providers to benefit from the growth of AI video generation, with specific companies recommended for investment, including chip designers and AI server manufacturers [13]. - The user base for short videos in China is expected to exceed 1 billion by the end of 2024, with increasing daily usage times projected [7][11].