上周三，沐曦股份登陆科创板，开盘暴涨超568%，市值瞬间突破3000亿元。1.9万股民集体狂欢，因为中一签就能暴赚30万，这要比前几天上市的摩尔线 程还要夸张。当下，两家公司市值合计超6000亿元，整个资本市场洋溢着一股近乎狂热的乐观气息。 更令人兴奋的是，就在沐曦上市的同一天，壁仞科技传来港股IPO已通过聆讯的消息，最快2026年初挂牌。这相当于给了前两次没赶上的投资者第三次机 会，国产GPU板块即将掀起新一轮高潮。中国GPU的春天，似乎终于来了。 但冷静下来看，这场国产GPU的狂欢，似乎只有投资者赚得盆满钵满。 根据招股书中披露的财务状况显示，摩尔线程、沐曦、壁仞目前均处于大额亏损状态。其中，摩尔线程2025年前三季度亏损7.24亿，沐曦为3.46亿，壁仞 在今年上半年亏损为16.01亿。更重要的是，目前三家公司都处在加大研发的阶段，商业化也都刚刚起步，根本不存在短期盈利的可能。 可在投资者眼中，这点亏损根本不算事。 因为三家公司的GPU可是AI时代的硬通货。从2023年大模型热潮爆发以来，全球算力供不应求。最大受益者英伟达在每个报告期都赚爆，成为人类历史 上市值最高的公司。各大投行打着计算器，接连给国产 ...

Core Insights - The article highlights a significant shift in the procurement strategy of AI computing power by top universities, moving from traditional hardware specifications to practical performance and compatibility with mainstream AI models like DeepSeek and Qwen [1][10]. Procurement Strategy - The procurement document emphasizes the importance of practical performance and compatibility over mere hardware specifications, marking a departure from the previous focus on "parameter stacking" [3][10]. - The core requirements include a focus on domestic production, ensuring real-world performance, and compatibility with existing CUDA ecosystems [3][10]. Technical Specifications - The procurement specifies the need for 13 GPU servers, with detailed requirements for CPU, memory, storage, and network capabilities, emphasizing the use of domestic components [4][6][8]. - Key performance indicators include the ability to run specific AI models under defined conditions, such as maintaining low latency and high throughput during inference tasks [9][11]. Trends in AI Computing Power Procurement - The article identifies three major trends in GPU server procurement by universities: 1. Transitioning from "indicator comparison" to "model testing" [10]. 2. Moving from a "NVIDIA-dominated ecosystem" to a "domestic compatible ecosystem" [10]. 3. Shifting focus from "hardware procurement" to "computing power system construction" [11]. Market Implications - The procurement strategy indicates a growing demand for domestic AI computing solutions that can effectively replace imported technologies, thereby fostering the development of a robust domestic AI ecosystem [12]. - Companies like Haiguang, Biran, and Muxi are highlighted as potential suppliers capable of meeting these stringent requirements, showcasing advancements in their GPU and CPU technologies [11][12].

Overview of AI Computing Infrastructure in China - The report by Analysys focuses on the development status, core driving factors, key trends, and stakeholder recommendations for AI computing infrastructure in China by 2025 [1] - The evolution path of computing infrastructure is shifting from "scale expansion" to "quality and efficiency improvement" [1] National Strategy and Scale Position - The "East Data West Computing" project is central to the national strategy, with plans to build national computing hubs in eight regions including Beijing-Tianjin-Hebei and the Yangtze River Delta, and to establish ten data center clusters [5] - As of 2024, the number of operational computing center racks in China is expected to reach 8.3 million, with a total computing power exceeding 280 EFLOPS, making it the second largest globally [7] - Intelligent computing power accounts for over 30% of the total, with a growth of nearly 13 times since 2019, averaging an annual growth rate of about 90% [7] Development Environment - National policies are solidifying top-level design, with local governments setting clear goals for intelligent computing construction [12] - Technological advancements in AI chips and cooling technologies are reducing Power Usage Effectiveness (PUE) [17] - The demand for computing power is surging due to generative AI, with applications expanding from the internet to traditional industries like finance and healthcare [19] - The supply of computing power is transitioning from heavy asset investment to platform-based services, lowering barriers for SMEs [21] Development Progress and Core Drivers - The development stages include an exploration phase (~2019), a market activation phase (2020-2022), and a high-speed growth phase (2023-2028) [26][34] - Five core driving factors include the iteration of large models, policy and capital linkage, industrial application scaling, long-tail computing power release, and cloud scheduling technology [35][36][37][38] Key Trends for AI Computing Infrastructure by 2025 - Trend 1: Accelerated breakthroughs in autonomous controllable computing power, with a goal for over 70% of computing power in Shanghai to be domestically controlled by 2027 [39] - Trend 2: Green computing becoming a hard constraint, with new data centers required to meet specific PUE standards [41] - Trend 3: Deepening cross-regional computing interconnection, enhancing the national backbone network for free flow of computing power [44] - Trend 4: Dual-track development of intelligent computing cloud platforms, offering comprehensive and vertical services [46] - Trend 5: New demands driven by large language models and knowledge bases, increasing the need for specialized computing centers [48] - Trend 6: Accelerated cloud computing for inclusivity, with a projected 80% growth in the smart computing service market by 2024 [49] Stakeholder Recommendations - Government: Strengthen regional computing network planning and provide financial/tax incentives for green computing and autonomous technology development [51] - Enterprises: Supply side should create high-quality computing cloud platforms, while demand side should prioritize cloud leasing over self-built solutions [52] - Industrial Parks: Develop intelligent computing industry clusters with supporting green energy and high-speed networks [53] - Ecosystem: Collaborate among chip, server, and cloud platform companies to tackle key technologies and establish industry standards [54]