专访云天励飞董事长陈宁：AI推理时代已至，推理芯片崛起将是中国科技复兴巨大机遇

Core Insights - The article discusses the ongoing transformation in the AI industry, highlighting the shift from training to inference as a pivotal moment for the sector, with 2025 anticipated as a year of significant AI application growth [2][3]. Industry Overview - The AI industry is evolving through three distinct phases: 1. The "Intelligent Perception" era (2012-2020), characterized by fragmented solutions driven by small models [3]. 2. The "AIGC" era (2020-2025), where large models demonstrate impressive content generation capabilities but struggle to find profitable business models [3]. 3. The upcoming "Agentic AI" era, where intelligent agents will integrate large models, operating systems, and hardware to perform complex tasks independently, marking a true industrial revolution [3][4]. Market Dynamics - The transition to inference-focused computing is seen as a fundamental shift, requiring a focus on cost-effectiveness and market economics rather than just performance [3][4]. - The emergence of dedicated inference chips is expected to disrupt Nvidia's dominance established during the training era, as companies like Google and Broadcom pivot towards specialized inference solutions [5][6]. Opportunities for China - China is positioned to capitalize on the inference chip market, as it faces fewer barriers compared to the training sector, where it lags behind Nvidia due to advanced process limitations and high CUDA ecosystem barriers [5][6]. - The rise of inference chips is viewed as a significant opportunity for China's technological resurgence, aligning with its strengths in providing high-cost performance products [5][6]. Technological Innovations - The introduction of the GPNPU architecture aims to address the unique demands of inference tasks, optimizing performance, storage bandwidth, and capacity while reducing costs [6]. - The goal is to lower the total cost of ownership (TCO) for users by enhancing energy efficiency and minimizing operational costs through innovative chip technologies [6]. Future Projections - The demand for inference computing is expected to surge, with projections indicating that the daily token processing volume could reach 100 trillion by mid-next year, necessitating significant infrastructure investments [7]. - Companies are urged to reduce the comprehensive cost of processing "million tokens" to one cent, which will require architectural and technological innovations [7].