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 global AI training competition, ignited by ChatGPT, is leading to a significant industrial transformation, with 2025 anticipated as the year of explosive AI application growth. The demand for reasoning computing power is surging, creating a sharp contradiction with high costs [1] - The CEO of CloudWalk Technology, Chen Ning, emphasizes that AI is a key driver of technological breakthroughs in the next five years, with China narrowing the gap in algorithms and having advantages in application, data, energy, and system integration [3] - The reasoning chip sector is seen as crucial for China to "overtake" in the AI landscape, marking a fundamental shift from training to reasoning in computing paradigms [4][5] Industry Phases - The development of the AI industry can be divided into three phases: 1. The "Intelligent Perception" era (2012-2020), characterized by fragmented solutions driven by small models 2. The AIGC (AI Generated Content) era (2020-2025), where large models demonstrate impressive content generation capabilities 3. The upcoming "Agentic AI" era (starting in 2025), where intelligent agents will integrate large models, operating systems, and hardware to perform complex tasks independently [4] Reasoning Chip Potential - Chen Ning highlights that the transition to reasoning requires a focus on market economics and high cost-performance ratios, contrasting with the training phase's emphasis on performance and iteration speed [5] - The emergence of independent reasoning chips is breaking Nvidia's monopoly established during the training era, as companies like Google and Broadcom are investing in specialized reasoning chips [6] New Chip Architecture - CloudWalk Technology proposes a new chip architecture called GPNPU, which aims to integrate three core capabilities: compatibility with CUDA ecosystems, optimization of matrix calculations, and advanced packaging technologies to reduce costs and memory bottlenecks [7] - The GPNPU aims to achieve a better balance between computing power, storage bandwidth, and capacity, addressing the diverse needs of future reasoning chip applications [7] Future Demand Scenarios - Chen Ning predicts explosive demand for reasoning capabilities, citing the example of the Doubao model, which processes 50 trillion tokens daily, with potential growth to 100 trillion tokens by mid-next year [8] - To support the industrialization of AI, there is a need to reduce the comprehensive cost of reasoning to a "penny" level per million tokens, achievable through architectural and technological innovations [8]