Workflow
【电新】AI服务器电源乘势而起——电力AI系列报告三(殷中枢/郝骞/和霖)

Core Viewpoint - The rapid development of AI is significantly increasing its electricity demand, with projections indicating that by 2026, AI's power consumption will be at least ten times that of 2023, leading to growth opportunities for server power supply systems [2] Group 1: AI Power Demand - The International Energy Agency (IEA) forecasts that AI industry electricity consumption will surge, with the U.S. Department of Energy predicting data center power demand in the U.S. to reach between 74-132 GW by 2028 [2] - Global competition in AI is intensifying, with significant investment plans announced by the U.S. and the EU, which are expected to further drive AI electricity demand [2] Group 2: Server Power Supply Upgrades - AI data centers require a transition of power from high-voltage electricity to chip-level voltage, involving three types of power supplies: UPS, AC-DC server power, and DC-DC chip power [3] - Current trends show that server power supplies are upgrading towards higher efficiency and power density, with future applications of 8kW and 12kW server power supplies anticipated [3] Group 3: Market Size and Growth - The market size for NVIDIA's AI server AC-DC power supply is projected to reach between 35.1 billion to 45.5 billion yuan by 2025, with a year-on-year growth rate of 43%-86% [4] - Major players in the AI server power supply market include Delta and Lite-On Technology, with revenues of approximately 1.82 billion yuan and 900 million yuan respectively in 2023 [4] Group 4: Components and Opportunities - The introduction of BBU (Battery Backup Unit) and supercapacitors in NVIDIA's GB300 server rack is expected to create development opportunities, as these components ensure stable power supply during outages [5] - The estimated value of BBU and supercapacitors in the GB200 NVL72 rack is approximately $10,800 and $7,500 respectively, indicating a low sensitivity of end-user cloud companies to these costs [5]