无问芯穹发起人汪玉：Token已成为智能时代最核心的生产要素之一

Core Viewpoint - The article emphasizes the transformation of AI infrastructure from a focus on energy and computing power to the importance of Tokens as a fundamental unit in the AI production process, marking a shift towards AI 2.0 where efficiency in processing Tokens becomes crucial [3][5][6]. Group 1: AI Infrastructure Evolution - The transition from AI 1.0 to AI 2.0 involves changing the evaluation metrics of infrastructure from TOPS (Tera Operations Per Second) to Tokens per Joule (Tokens/J), highlighting the need for optimizing Token efficiency under energy consumption constraints [3][6][12]. - Tokens are identified as the core production factor in the AI era, replacing traditional data elements, and are essential for training large models and supporting multi-modal applications [5][6][7]. Group 2: Technical Challenges and Solutions - The article discusses the need for collaborative optimization between software and hardware to enhance the efficiency of Tokens/J, especially as the complexity of AI tasks increases [7][12]. - It highlights the importance of sparse matrix optimization and quantization techniques in improving neural network performance, with trends moving towards structured sparsity and real-time sparse training [9][10]. Group 3: Future Directions and Industry Collaboration - The focus is on building a multi-layered AI infrastructure that integrates bottom-level hardware, middle-layer models, and top-layer applications to enhance overall efficiency [13][14]. - The company aims to leverage AI cloud capabilities to empower various industries while facilitating the adoption of new terminal devices in everyday life, indicating a commitment to industry collaboration and innovation [15].