机械设备行业点评报告 Google Gemini 3 表现超预期，看好 AI 算力 需求的成长性 证券研究报告·行业点评报告·机械设备 增持（维持） [Table_Tag] [投资要点 Table_Summary] ◼ 事件：Google Gemini 3 发布，表现超市场预期 近期 Google 发布了最新的多模态大模型 Gemini 3，展现出超市场预 期的得分能力与多模态理解能力。 ◼ Google Gemini 3 在 Benchmark 测试上实现断层领先 Gemini 3 问世前，顶尖大模型之间的 Benchmark 测试得分差距微小， Gemini 3 推出后在 Benchmark 测试上的得分实现了断层式领先：①HLE 测 试基础思考能力得分 37.5%（无工具），领先于 Gemini 2.5 Pro 的 21.6% 和 GPT-5.1 的 26.5%；②多模态理解能力上，ScreenSpot-Pro（截图理解） 测试得分 72.7%，远超 Claude Sonnet 4.5 的 36.2%和 GPT-5.1 的 3.5%。 Gemini 3 的另一大亮点是"生成式 UI"的强大审美能力，逐步 ...

Summary of Key Points from Conference Call Industry Overview - The conference call discusses the application of power devices in AI data center power supplies, focusing on high-end AI servers like NVIDIA's NVL72, which utilize multiple 5.5 kW power modules, achieving a total cabinet power of 132 kW [1][3][6]. Core Insights and Arguments - **Power Supply Units (PSUs)**: The current architecture of server power supplies can be categorized into embedded power supplies and cabinet power supplies. High-end AI servers like the NVIDIA NVL72 utilize multiple 5.5 kW modules, with each kW priced around 2,100 RMB [3][6]. - **Future Trends**: The development of 800V High Voltage Direct Current (HVDC) technology is expected to increase PSU power from 5.5 kW to 12 kW or even 19 kW in the coming years, necessitating the use of advanced semiconductor materials like silicon carbide (SiC) and gallium nitride (GaN) to enhance efficiency and reduce size [1][7]. - **Material Applications**: SiC is primarily used in high-voltage applications, while GaN is favored for lower voltage applications due to its higher efficiency and lower thermal loss. The next-generation Ruby solution will support 800V HVDC input [1][10][11]. - **Cost Considerations**: SiC devices are more expensive, with 650V SiC devices priced around 50 RMB each, while 1,200V SiC devices cost about 90 RMB. GaN devices are similarly priced between 50 to 60 RMB [23][24]. Additional Important Content - **Domestic Market Development**: Domestic manufacturers like Great Wall Electronics are beginning to adopt GaN technology in their 7.5 kW or 8 kW power modules, bridging the gap between standard servers and high-end AI servers [25]. - **Current Usage of IGBT and SiC**: IGBT remains dominant in current server power systems, but there is a gradual shift towards SiC in future 800V HVDC systems [15]. - **Market Dynamics**: The power market is still in an experimental phase, with high reliability requirements leading to lower price sensitivity among customers [40]. - **Competition and Supply Chain**: Domestic suppliers like Dongwei Semiconductor and SIDA Semiconductor are providing products for local enterprises, but they have not yet entered the core supply chain dominated by international players like NVIDIA [38][39]. Conclusion - The conference highlights the ongoing transition in power supply technology for AI servers, emphasizing the importance of advanced materials like SiC and GaN, the evolving market dynamics, and the competitive landscape in both domestic and international contexts. The future of server power supplies will likely see increased efficiency and power capacity driven by technological advancements and market demands.