智算算力交易分层定价模型研究

Research on the Tiered Pricing Model for Intelligent Computing Power Trading

摘要: 随着生成式人工智能、大模型训练、智能视频分析等应用的快速发展，全球智能算力需求呈现指数级增长。传统的算力定价模式主要基于硬件配置或使用时长，难以适应多样化、高并发的智算任务需求，存在任务适配性差与服务等级协议覆盖不足等问题。本文提出一种面向智算算力交易的分层定价模型，将定价模型划分为基础服务与增值服务两部分。基础层基于任务类型(如图像检测、自然语言处理、语音识别等)的核心先验参数构建定价模型；增值层则根据用户对时效、并发、多节点协同等服务等级需求进行浮动。结合真实智算任务日志开展实证研究，通过对比传统计费与分层计费的成本、任务完成率等指标，验证了模型的优越性。该模型在提升资源利用率、满足差异化服务需求方面具有显著优势，为算力资源的市场化配置提供了可行的技术路径。

Abstract: With the rapid development of applications such as generative AI, large model training, and intelligent video analytics, the global demand for intelligent computing power is growing exponentially. Traditional computing power pricing models, primarily based on hardware configuration or usage duration, struggle to accommodate the requirements of diverse, highly concurrent intelligent computing tasks. These models face challenges including poor task adaptability and insufficient Service Level Agreement (SLA) coverage. This paper proposes a tiered pricing model for intelligent computing power transactions, dividing the pricing structure into basic services and value-added services. The basic service tier establishes pricing models based on core prior parameters of task types (such as image detection, natural language processing, and speech recognition). The value-added tier incorporates dynamic adjustments according to users’ SLA requirements including timeliness, concurrency, and multi-node collaboration. Empirical research is conducted using real intelligent computing task logs, and the superiority of the model is verified by comparing costs, task completion rates, and other indicators between traditional billing and layered billing. This model demonstrates significant advantages in enhancing resource utilization and meeting differentiated service demands, providing a viable technical pathway for the market-based allocation of computing power resources.