英伟达改卖Token？黄仁勋GTC后发声：token就是AI新通货，值钱的不是算力，是“每度电的智商”

Core Viewpoint - NVIDIA is positioned as an "accelerated computing company" rather than merely a GPU company, emphasizing the importance of the entire technology stack in AI development [2][10][24]. Group 1: AI Competition and Token Economy - The AI competition has shifted from merely computing power to producing high-quality results quickly and cost-effectively, with the entire process needing acceleration [4][5]. - Tokens are viewed as the core currency of the AI era, where smarter tokens can command higher prices, reflecting the efficiency of the models generating them [7][8]. - NVIDIA's acquisition of Groq and the introduction of Groq LPU aim to address the challenge of generating tokens with low latency, complementing existing GPU capabilities [9][10]. Group 2: Full-Stack Approach and Industry Integration - NVIDIA is transitioning from a focus solely on chips to a comprehensive understanding of applications, necessitating a full-stack approach to accelerate software and tools used by AI [12][20]. - The company aims to build AI factories and infrastructure globally, integrating various components like networking and storage to enhance overall system performance [22][26]. - The integration of AI with existing human tools, such as Excel and SQL, requires significant acceleration to keep pace with AI's rapid processing capabilities [14][15][30][31]. Group 3: Future of AI Models and Architectures - The limitations of current models like Transformers necessitate the development of new architectures that can handle long-term memory and continuous tasks more effectively [33][36]. - AI's ability to generate economic value is linked to its improved reasoning capabilities, allowing it to perform tasks beyond mere information generation [40][41]. - The emergence of coding agents signifies a shift where AI can assist in programming, enhancing efficiency and allowing engineers to focus on higher-level problem-solving [45][46]. Group 4: Role of CPUs and System Design - CPUs remain crucial in the AI ecosystem, with NVIDIA emphasizing the need for high-performance CPUs to prevent bottlenecks in GPU utilization [53][64]. - The design of CPUs like Vera focuses on high I/O bandwidth and single-thread performance to support the demands of AI applications [64][66]. - NVIDIA's strategy includes a collaborative approach with various architectures, ensuring that the best components are utilized for optimal system performance [66][87]. Group 5: Supply Chain and Market Dynamics - The current landscape shows that nearly all aspects of the supply chain are nearing capacity, making it challenging to scale any single component significantly [92][95]. - NVIDIA's proactive supply chain planning positions it favorably to meet future demands, despite potential constraints in power and chip availability [95][96]. - The company recognizes the importance of maintaining a competitive edge in the technology stack across all layers of AI development, from infrastructure to applications [98][99].