AI用电的“困”与“破” | 投研报告

Core Insights - The report by Jinyuan Securities highlights the significant energy consumption associated with training advanced AI models, particularly GPT-4, which requires approximately 38.2 GWh over a training period of 95 days [1][2]. Group 1: AI Model Training Energy Consumption - The maximum rated power consumption of a single GPU can reach 1000 watts [2]. - GPT-4's training energy demand translates to an average daily consumption of about 0.40 GWh, equivalent to the daily electricity usage of approximately 40,000 households [1][2]. Group 2: Data Center Energy Consumption Trends - Global data center electricity consumption is projected to reach 415 TWh in 2024, accounting for about 1.5% of the world's total electricity consumption, with the US, Europe, and China collectively responsible for 85% of this total [2]. - In the US, data center electricity consumption is expected to grow at an annual rate of approximately 12% from 2015 to 2024, resulting in an increase of about 250 TWh, with 2024 consumption estimated at 180 TWh [2]. Group 3: Future Projections - By 2030, global data center electricity consumption is forecasted to double to approximately 945 TWh, representing nearly 3% of global electricity consumption, with the US market projected to consume about 420 TWh, a 130% increase from 2024 [2]. Group 4: Power Infrastructure Challenges - The mismatch between the manufacturing cycle of power infrastructure and the demand cycle for AI presents significant challenges [3]. - Solar energy and gas turbines are currently the only reliable power sources that can be developed in a short timeframe to meet data center demands, but solar energy's intermittent nature poses challenges for continuous operation [3]. Group 5: Energy Storage Systems - A complete electrochemical energy storage system consists of battery packs, battery management systems (BMS), energy management systems (EMS), and storage inverters, collectively referred to as "3S" [4]. - The global inverter market is projected to reach approximately $129.2 billion in 2024, with the battery energy storage system (BESS) market expected to grow at a compound annual growth rate of 18.5% from 2024 to 2030, reaching $23.3 billion by 2030 [4]. Group 6: Power Device Market Insights - The IGBT and silicon-based MOSFET markets are significant, with the IGBT market projected to grow from approximately $8.887 billion in 2024 to $16.151 billion by 2030, reflecting a compound annual growth rate of 10.47% [5]. - The market for silicon carbide (SiC) devices is also expected to grow significantly, with a combined market size projected to increase from $2.967 billion in 2024 to $9.520 billion by 2030, at a compound annual growth rate of 21.45% [5]. Group 7: Related Companies - Companies involved in this sector include SDIC (603290.SH), Yangjie Technology (300373.SZ), ChipLink Integration-U (688469.SH), Silan Microelectronics (600460.SH), and Dongwei Semiconductor (688261.SH) [6].