非GPU赛道，洗牌

Core Viewpoint - The rise of non-GPU chip forces is unstoppable, indicating a significant shift in the global computing power industry, traditionally dominated by NVIDIA GPUs [1][6]. Group 1: Recent Developments in the Computing Power Industry - Shanghai's leading GPU company, Muxi Co., recently listed on the STAR Market, with its stock price surging by 687.79% to a market cap of 329.88 billion yuan [1]. - Google’s TPU has secured orders worth over 100 billion yuan, breaking the GPU monopoly in the computing power market, while Broadcom's CEO revealed a total order of 21 billion USD (approximately 148.6 billion yuan) from Anthropic [2]. - In China, the computing power industry is also heating up, with AI chip company Qingwei Intelligence securing over 2 billion yuan in financing, supported by a rare investment lineup [2]. Group 2: Market Trends and Dynamics - The global computing power market is experiencing a transformation, with the long-standing NVIDIA GPU monopoly beginning to loosen [2][6]. - The demand for general computing power has led to NVIDIA GPUs dominating the market, but alternatives like Google’s TPU and Amazon’s Trainium3 are starting to replace GPUs in specific scenarios [2][8]. - By the first half of 2025, non-GPU computing cards are expected to account for 30% of the domestic market [2]. Group 3: Investment Movements - Intel is reportedly planning to acquire AI chip unicorn SambaNova for 1.6 billion USD (approximately 11.29 billion yuan) to regain competitiveness in the AI era [3]. - Non-GPU unicorn Groq has raised over 3 billion USD (approximately 21.3 billion yuan) in funding over the past two years [3]. Group 4: Industry Structure and Future Directions - The computing power industry is expected to face path differentiation, driven by demand, technology, and ecosystem development [8][10]. - The need for efficient computing solutions is pushing companies to seek alternatives to the traditional GPU-centric model, especially as AI applications diversify across various industries [8][9]. - The traditional von Neumann architecture is facing challenges, necessitating architectural innovations to overcome performance limitations [9]. Group 5: Non-GPU Market Growth - Gartner predicts that by 2027, the demand for AI inference applications will lead to AI accelerators (typically non-GPU AI-specific chips) surpassing GPU shipments [16]. - In the first half of this year, China's non-GPU chip market has shown significant growth, with projections indicating a market share of nearly 50% by 2028 [16]. Group 6: Domestic Chip Companies and Their Strategies - Key players in the domestic non-GPU sector include Kunlun Chip, Cambricon, and Qingwei Intelligence, each representing different technological routes [25][29]. - Cambricon and Kunlun Chip focus on ASIC routes, while Qingwei Intelligence emphasizes reconfigurable computing architectures [29][30]. - The ASIC architecture, exemplified by Google’s TPU, offers high performance and efficiency, but requires significant time and resources for customization [30][31]. Group 7: Reconfigurable Computing Advantages - Reconfigurable computing is gaining momentum, addressing the inefficiencies of GPUs and the rigidity of ASICs, thus balancing performance and cost [32][37]. - Qingwei Intelligence's reconfigurable chips have achieved over 30 million units shipped, with significant orders expected in the coming years [32]. - The technology supports efficient inter-chip communication, avoiding bandwidth bottlenecks and communication delays inherent in traditional architectures [33]. Group 8: Conclusion - The AI computing power landscape is evolving towards a diversified and heterogeneous integration, with GPUs maintaining dominance in general-purpose applications while non-GPU routes rapidly rise in AI inference and specialized computing needs [39][41].