Summary of Conference Call on Domestic AI Chip Industry Industry Overview - The domestic AI chip industry is rapidly evolving as internet companies in China shift towards domestic AI chips to counter restrictions on high-end products from NVIDIA. Major players like ByteDance, Tencent, and Alibaba are increasing their procurement of products from companies such as Cambricon, Suiyuan, and their own TPUs and Pingtouge products to meet large-scale training needs and localization requirements [1][5]. Key Points and Arguments - **Performance Comparison**: Domestic AI chips are gradually approaching NVIDIA's high-end products, specifically targeting the A100/A800 series. The next generation is expected to achieve 60%-80% of the performance level of the H200 series, with inference capabilities already comparable to the H20 series [1][3]. - **Supply Chain Challenges**: The domestic AI chip industry faces a bottleneck in 7nm process capacity, necessitating multi-chip interconnects to enhance performance. Power management improvements are required, with air-cooled modules needing to reach 800-1,000 watts and liquid-cooled modules 1,200-1,500 watts [1][4]. - **Diverse Solutions**: Different companies have varying approaches to supernode solutions. For instance, Kunlunxin uses Ethernet switches, while Cambricon employs ByteDance's proprietary protocol. Other companies like Muxi and Tianshu rely on established solutions like P3E [1][9]. - **Market Demand**: The demand for domestic AI chips is increasing as companies like ByteDance plan to scale their operations significantly, with expectations of reaching 200,000 cards by 2025 and doubling to 500,000 by 2026. The overall market shipment is projected to be between 500,000 to 800,000 cards [2][5]. - **Ecosystem Development**: The general GPU route has advantages due to CUDA compatibility, while proprietary routes face ecological bottlenecks. Continuous updates are necessary to keep pace with the latest versions [4][6]. Additional Important Insights - **Government Initiatives**: The construction of government intelligence computing centers has slowed down, leading chip manufacturers to focus more on industry applications and internet demands. Key regions include Shanghai, Hefei, Hangzhou, Shenzhen, and Beijing [10][11]. - **International Operations**: Some domestic internet giants have established computing centers overseas to utilize advanced chips for model training, subsequently bringing the trained models back to China for fine-tuning and inference [16][17]. - **Competitive Landscape**: Companies like Huawei, Cambricon, and Haiguang are competing in the chip market, with Huawei's 910C currently holding an advantage over others like Haiguang's BW1,000 [19]. - **Future Trends**: The demand for domestic GPUs is expected to rise, particularly in sectors like finance and energy, as restrictions on imported products increase [7][8]. This summary encapsulates the key discussions and insights from the conference call regarding the domestic AI chip industry, highlighting the competitive landscape, performance benchmarks, and evolving market dynamics.