2024年第52计算机行业周报：从DeepSeek-V3看AI算力需求变迁

Investment Rating - The report maintains a "Positive" investment rating for the industry [24]. Core Insights - The establishment of the Low Altitude Economic Development Department is expected to accelerate the development of the low-altitude economy, which has been recognized as a new growth engine by the government [14][15]. - The launch of DeepSeek-V3, a domestic AI model, demonstrates significant advancements in AI computing power, with training completed using only 2.778 million H800 GPU hours, indicating potential for cost-effective model development and increased demand for inference computing power [6][50]. - Shanghai aims to build a world-class AI industry ecosystem by 2025, targeting an intelligent computing power scale of over 100 EFLOPS, which will benefit the domestic computing power industry [18][41]. Summary by Sections Market Overview - The overall market saw a slight increase, with the Shanghai Composite Index closing at 3400.14 points, up 0.95%, while the computer sector experienced a significant decline of 4.66% [5][29]. Key Recommendations - Focus on domestic computing power, particularly recommending leading AI model companies like iFlytek and AI chip leaders like Cambricon [6][50]. Recent Developments - The Low Altitude Economic Development Department has been officially established, which is expected to lead to a more organized development of the low-altitude economy [14][15]. - The "Molding Shanghai" initiative aims to enhance AI capabilities, with a goal of achieving a computing power scale of 100 EFLOPS by the end of 2025 [18][41]. Performance Metrics - DeepSeek-V3 has outperformed several existing models in various benchmarks, showcasing its competitive edge in the AI landscape [42][44]. - The training cost for DeepSeek-V3 was approximately $5.576 million, significantly lower than other comparable models, indicating a trend towards more efficient AI model training [82]. Future Outlook - The report anticipates a substantial increase in inference computing demand, with projections indicating that by 2027, inference workloads will account for 72.6% of AI server workloads [48][71].