2025年算力调度平台行业 优化计算资源，支撑AI应用 ◼ 研究背景 随着人工智能技术的迅猛发展，全球 范围内对算力的需求呈现指数级增长， 因此需要算力调度来实现跨地域、跨 平台的算力资源整合与优化调度。 ◼ 研究目标 概览标签：算力、算力调度 Computing Power Scheduling Platform Industry 計算力スケジューリングプラットフォーム業界 报告提供的任何内容（包括但不限于数据、文本、图表、图像等）均系头豹研究院独有的高度机密性文件（在报告中另 行标明出处者除外）。 ，任何人不得以任何方式擅自复制、再造、传播、出版、引用、改 编、汇编本报告内容，若有违反上述约定的行为发生，头豹研究院保留采取法律措施、追究相关人员责任的权利。头豹 研究院开展的所有商业活动均使用"头豹研究院"或"头豹"的商号、商标，头豹研究院无任何前述名称之外的其他分支机构 ，也未授权或聘用其他任何第三方代表头豹研究院开展商业活动。 头豹研究院 1 01 异构算力调度面临的挑战： ◆ 异构算力调度面临多重核心挑战：资源异 构性与软件环境碎片化显著增加调度复杂 性；跨架构任务迁移成本高导致效率低下； 缺乏统一调 ...

Data Center - Nvidia has ordered 300,000 H20 chips from TSMC due to strong demand in China, shifting from a reliance on inventory to actual supply [1] - Major internet companies like ByteDance, Tencent, and Alibaba are accelerating their data center construction demands, leading to potential increases in capital expenditures [1] - ByteDance plans to start its modular data center procurement project by July 30, 2025, which may provide guidance for the related sectors [1] - There is a forecasted increase in demand for H20 chips, with ByteDance potentially purchasing an additional 100,000 units and Tencent needing between 20,000 to 50,000 units [1] AI Infrastructure - ByteDance's procurement plan for modular data centers is a key part of its AI-driven data center strategy, aimed at enhancing computing power for AI model training and applications [2] - The company is expected to spend 160 billion yuan in capital expenditures by 2025, with approximately 90 billion yuan allocated for AI computing power procurement [2] - The procurement plan aligns with the H20 chip release, likely leading to explosive growth in demand across the liquid cooling, power generation, and server industries [2] Computing Power Scheduling - The first intelligent storage scheduling platform in China has been launched, addressing key challenges in data storage [3] - This platform integrates heterogeneous resources and dynamic scheduling, enhancing the foundation for nationwide data center construction and AI technology implementation [3] - The platform aims to optimize resource allocation and improve network stability, facilitating the efficient circulation and value extraction of data elements [3] Huawei Harmony - Huawei's Cangjie programming language will be open-sourced on July 30, featuring a compiler and standard library aimed at native intelligence and high performance [4] - Cangjie is designed to support all-scenario application development within the Harmony ecosystem, playing a crucial role in its infrastructure [4] Macro and Industry News - The Central Organization Department allocated 140 million yuan to support flood relief efforts in Beijing and other areas [5] - Chinese stock ETFs listed overseas have seen significant capital inflows since July [5] - The Beijing Stock Exchange is conducting self-inspections focusing on the technical and business preparations for batch switching of securities codes [5]

Investment Rating - The report indicates a positive investment outlook for the computing power scheduling platform industry, driven by the rapid growth in AI model applications and the increasing demand for high-performance computing resources [10][14]. Core Insights - The demand for intelligent computing power in China is experiencing unprecedented growth, particularly in large model applications, which account for nearly 60% of the demand, highlighting the significant potential of the future computing power market [14][30]. - The profitability of computing power scheduling centers heavily relies on government subsidies, which are designed to ensure local resource utilization and risk control. Successful engagement with government decision-making is crucial for companies to secure these subsidies [22][30]. - The core value of computing power scheduling platforms lies in their ability to efficiently integrate and schedule heterogeneous computing resources, significantly improving resource utilization and reducing costs for users [17][20]. Summary by Sections Computing Power Scheduling Platform's Scaling Path - The rapid growth of intelligent computing demand in China is driven by AI large models, leading to accelerated development of large-scale, high-performance computing centers [11][10]. - The profitability of computing power scheduling centers is highly dependent on government subsidies, which aim to ensure local resource utilization and risk management [22][30]. - Platforms must possess efficient integration and scheduling capabilities for heterogeneous computing resources to achieve low-cost, scalable, and marketable monetization [31][39]. Value of Computing Power Scheduling Platforms - The platforms enhance resource utilization, lower user costs, and simplify management processes, providing efficient and convenient computing power services [17][20]. - The core technologies required include resource virtualization, fine-grained slicing, real-time monitoring, and tidal scheduling, which enable low-cost and efficient utilization of resources [31][38]. Government Subsidies as Core Source - Government subsidies are essential for the profitability of computing power scheduling centers, with a structured mechanism to ensure local resource utilization and risk control [22][30]. - Companies must strategically engage early in government decision-making to influence standards and secure contracts [23][30]. Technical Features of Quality Computing Power Scheduling Platforms - Platforms need to have capabilities for fine-grained resource slicing, heterogeneous compatibility, cross-regional scheduling, real-time monitoring, and dynamic scheduling to achieve efficient resource reuse and low-cost monetization [31][38]. Core of Scaling Computing Power Centers - The core of monetizing the value of computing power platforms lies in a large and diverse customer base, which determines profitability speed and pricing potential [39][42]. - A diverse customer base allows for tiered pricing strategies to maximize revenue, while partnerships can help focus on high-margin computing power sales [42][39].

Group 1 - The State Intellectual Property Office of China has published a patent application by China Southern Power Grid Co., Ltd. for a method and device related to a computing power scheduling mechanism based on a container orchestration platform, with publication number CN120029785A and application date of February 2025 [1] - The patent describes a method that involves extracting key attribute information from configuration files of heterogeneous computing resources, adjusting component information in the container orchestration platform, and deploying heterogeneous computing resources on the adjusted platform [1] - The method aims to determine target computing resources based on the processing requirements of tasks and the status information of candidate computing resources, allowing for efficient task allocation [1] Group 2 - China Southern Power Grid Co., Ltd. was established in 2004 and is located in Guangzhou, primarily engaged in the production and supply of electricity and heat [2] - The company has a registered capital of 9,020 million RMB and has made investments in 35 enterprises, participated in 5,000 bidding projects, and holds 3,520 trademark records and 4,985 patent records [2] - Additionally, the company possesses 65 administrative licenses [2]