阿里云栖大会观展反馈 20250928 摘要 国产算力商业闭环于 2024 年 10 月形成并强化，受益于豆包模型接近 GPT-4 能力及华为、寒武纪芯片突破，带动 AI 硬件崛起，国内推理和应 用端需求显著增长，字节跳动 token 消耗量超 Google，第三方 IDC 企 业如世纪互联业绩提升。 阿里云栖大会显示，阿里数据中心装机容量到 2032 年将达 2022 年的 10 倍，每年新增 6-7GW。阿里超节点盘久采用 CPU 与 GPU 分离架构， GPU 可使用阿里平头哥，CPU 选择多样，并公开 A Link 协议，类似华 为 UB 协议，促进计算存储网络互通。 国产 AI 硬件生态迅速发展，中兴发布凌云 SKY UP 交换芯片，字节跳动 或将推出相应芯片，表明国内交换芯片及模型训练与推理需求旺盛。阿 里钉钉平台拥有数千万中小企业客户，其 OA、业务流及供应链系统部 署于阿里云与钉钉之上，为 AGI 商业闭环提供沃土。 阿里、华为、海光公开互联协议，有助于 GPU 和交换芯片企业集成，实 现高效互联互通。阿里 Alink 系统实现小于 150 纳秒的 PTP 低延迟和单 柜 PB 级带宽，采用 ...

Core Insights - The construction of data centers in the U.S. is facing a dual challenge of power scarcity and community opposition, exacerbated by rising electricity prices driven by increased demand from these centers [1][6][8] Group 1: Electricity Price Surge - The capacity price in the PJM interconnection has skyrocketed from $2.2 billion in the 2023/2024 delivery year to $16.1 billion in the 2026/2027 delivery year, indicating a significant increase in electricity costs [3] - Capacity prices in the PJM "rest of market" area surged from $29 per megawatt-day in the 2024/2025 delivery year to $269 per megawatt-day in the 2025/2026 delivery year, marking a more than fivefold increase within a year [3] - This surge in electricity prices has resulted in an average bill increase of 18% to 25% for residents in the PJM region [3][6] Group 2: Policy Responses and Community Resistance - At least 12 states are considering new policies to ensure data centers bear the full costs of their electricity consumption, aiming to prevent the financial burden from falling on ordinary consumers [1][7] - Local policymakers are under pressure to create special rate structures that internalize the costs associated with data centers, reflecting a significant policy shift [7] - Community opposition, driven by concerns over rising electricity costs, water resource depletion, and noise pollution, is increasingly becoming a threat to data center projects [8][9] Group 3: Energy Solutions - Short-term solutions to the electricity demand crisis include solar and energy storage, which accounted for 80% of new generation capacity in the U.S. in 2024 [2][10] - The U.S. Energy Information Administration (EIA) reported that 48.6 GW of new capacity was added, with approximately 80% coming from solar and storage [10] - In the long term, natural gas and nuclear energy are viewed as essential components for ensuring stable power supply, with large tech companies exploring direct power agreements with nuclear energy providers [10][14]

如果您希望可以时常见面，欢迎标星收藏哦~ 来源：内容来自半导体芯闻综合 。 到 2030 年，人工智能机架的能耗将是传统机架的 20 至 30 倍。 预计同期平均机架容量将稳步上升至 30-50kW 冷却和配电成为未来数据中心的战略重点 机 架 长 期 以 来 被 认 为 是 数 据 中 心 的 基 本 单 元 ， 但 随 着 人 工 智 能 的 兴 起 ， 机 架 正 在 发 生 改 变 ， Lennox 数据中心解决方案公司提供的一张新图表（上图）显示了这种变化正在迅速展开。 该公司预测，以前以人工智能为中心的机架仅消耗几千瓦的电力，到 2030 年，其用电量可能达到 1MW，这一规模曾经只有整个设施才能达到。 预计同期数据中心机架的平均功率将达到 30-50kW，反映出计算密度的稳步上升，与 AI 工作负 载的对比十分鲜明。 电力输送和冷却的新需求 据预测，单个 AI 机架所消耗的能源是通用机架的 20 到 30 倍，这对电力输送和冷却基础设施提 出了新的需求。 普尔弗描述了目前正在进行的行业合作水平。"制造商、工程师和最终用户的合作比以往任何时候 都更加紧密，分享见解，并在实验室和实际部署中共同进行实 ...

通信行业点评 液冷技术再突破，IDC 热管理"变纪元"开启 2025 年 09 月 28 日 ➢ 微软发布液冷新技术，致力突破 IDC 功耗。根据 IDC 圈 2025 年 9 月 25 日 咨讯，微软董事长兼首席执行官 Satya Nadella 在社交媒体宣布，微软团队取得 了一项巨大突破：一种使用微流控技术的新型液冷方法，为比传统方法更高效、 更可持续、更密集的的数据中心打开了大门。 ➢ 微软该项技术有望重塑 IDC 液冷格局。我们认为液冷引入复杂流体系统， 使得越来越多的芯片服务器"出厂即液冷"。若微软的微流控技术实现规模化，必 将对 IDC 行业带来"二次冲击"。与冷板和浸没式液冷的"外部接触"不同，微 流控将冷却液直接嵌入芯片硅层，热传导效率提升 2-3 倍。而内置流体与外界的 交换的耦合可能反倒没有冷板紧密，同时，芯片的高散热能力实际上会提升对机 房整体散热能力的需求。 ➢ 看好芯片性能持续迭代下带来的 AI 热管理&基建侧机遇。根据 IEA 预测： 全球为满足数据中心的用电需求，电力供给将从 2024 年的约 460TWh 增长， 至 2030 年超过 1000TWh，即数据中心整体的电力需求 ...

Core Viewpoint - BP's subsidiary Castrol is betting on the Chinese liquid cooling server market by launching an end-to-end liquid cooling solution aimed at enhancing the efficiency of data centers in China [1][4]. Group 1: Market Trends - The data center industry is facing significant energy consumption challenges, particularly with the rise of AI applications, leading to increased power demands from high-end GPU/AI chips, which have surged from approximately 500-700W to levels of 1000W and even 2000W [1][3]. - The power density of server cabinets has exceeded 50kW, with leading AI cabinets commonly reaching over 100kW, making liquid cooling the only viable solution for effective heat management beyond this threshold [1][3]. - The Chinese data center market is rapidly transitioning to liquid cooling, with IDC forecasting a market size of $2.37 billion in 2024, representing a 67% increase from 2023 [7]. Group 2: Technical Insights - Liquid cooling technology is seen as a major upgrade for data centers, addressing the need for improved energy efficiency and lower operational energy consumption, with PUE (Power Usage Effectiveness) being a critical metric [5][6]. - The cooling liquid is a unique medium that interacts with all components of the data center infrastructure, necessitating in-depth research and understanding of its properties, including material compatibility and lifecycle changes [4][6]. - Castrol emphasizes an end-to-end solution that integrates the entire system, addressing the complexities of liquid cooling deployment, operation, and retirement while ensuring compatibility with existing equipment [4][6]. Group 3: Financial Considerations - The cost structure of liquid cooling can be divided into CAPEX (capital expenditures) and OPEX (operational expenditures), with CAPEX covering design, hardware, and deployment, while OPEX benefits from reduced energy and maintenance costs due to improved efficiency [6]. - The company is investing heavily in compatibility testing for various chip and server types to ensure confidence among clients and facilitate the adoption of liquid cooling solutions [6]. Group 4: Strategic Importance of China - China is identified as a key strategic market for Castrol due to its competitive supply chain and partnerships with local firms like Inspur and Yixin Technology [7]. - The liquid cooling server market in China is projected to grow at a compound annual growth rate (CAGR) of 46.8% from 2024 to 2029, with the market size expected to reach $16.2 billion by 2029 [7].

英国石油公司（BP）旗下润滑油巨头嘉实多正在押宝中国液冷服务器市场。近日，嘉实多宣布，面向 中国市场发布其端到端液冷解决方案，推动液冷在中国数据中心从设计到交付完整落地。 数据中心行业被形容为"不冒烟的钢厂"，一直受高能耗问题困扰，人工智能应用爆发下，能耗问题更加 突出。嘉实多热管理事业部全球总裁黄建棠告诉21世纪经济报道记者，观察近两年数据中心市场，高端 GPU/AI芯片功耗已从约500–700W跃升至1000W，甚至2000W级，机柜功率密度轻松越过50kW，头部 AI机柜普遍达100kW以上。为在此密度下控制散热与PUE，七八十千瓦尚可采用液/风混合，但一旦超 过100kW，风冷基本失效，液冷将成为唯一可行路径，其根本原因在于液体比热容与导热能力远高于空 气、可从芯片源头高效带走热量。 黄建棠进一步分析，在训练与大规模推理场景中，算力主要由GPU芯片承担，CPU更多负责调度与通用 负载。目前，数据中心的IT功耗/资本开支明显向GPU芯片倾斜，常见情况下有60%–80%的IT功耗集中 在GPU加速节点，其余分配给CPU、存储与网络等部件，从散热能力与热负载匹配来看，高功耗的加速 节点必然优先采用液冷。 从 ...

Core Viewpoint - The penetration rate of magnetic levitation compressors is expected to increase, benefiting Hanzhong Precision Machinery's products due to rising demand for cooling systems in data centers driven by AI computing power growth [1] Group 1: Magnetic Levitation Compressors - The collaboration between NVIDIA and OpenAI includes a significant investment of up to $100 billion (approximately 711.5 billion RMB) for building large data centers, which will drive the demand for upgraded cooling systems [1] - Traditional cooling technologies face energy consumption bottlenecks, making magnetic levitation compressors a key technology for energy-saving retrofits in data centers [1] - Hanzhong Precision Machinery offers a total of 9 models in the RTM series of magnetic levitation refrigeration compressors suitable for data center applications, indicating potential for substantial benefits in the future [1] Group 2: Vacuum Pumps and Industry Outlook - Currently, the company's vacuum pumps are primarily used in the photovoltaic industry, specifically in crystal pulling and battery cell segments [1] - As capital expenditures in the photovoltaic and semiconductor industries are expected to recover, the company is likely to benefit significantly from this trend [1] - The report maintains a "strongly recommended" rating for the company, reflecting positive expectations for its future performance [1]

Core Viewpoint - Alibaba is actively advancing a three-year plan to invest 380 billion yuan in artificial intelligence (AI) infrastructure, with a focus on enhancing its data center capabilities to support the anticipated arrival of super artificial intelligence (ASI) by 2032 [3][4]. Investment in AI Data Centers - The overall investment in AI data centers (AIDC) in China is projected to increase from 187.5 billion yuan in 2024 to 482.6 billion yuan by 2027 [3]. - Major tech companies like Tencent, ByteDance, Huawei, and Baidu are accelerating their investments in AIDC as AI technology continues to advance [3]. Power Supply and Stability - The stability of power supply is critical for the operation of data centers, with power outages being the leading cause of downtime, accounting for 54% of incidents [6]. - AI training requires unpredictable power loads, which can fluctuate significantly in short time frames, posing challenges for power grid management [6]. Energy Management Initiatives - ByteDance is launching a project to procure large-scale lithium battery storage systems and renewable energy sources to ensure stable power supply for its data centers, aiming to reduce reliance on traditional power grids [7][9]. - The project aligns with China's dual carbon goals, emphasizing the need for data centers to utilize renewable energy sources effectively [9]. Growth of Energy Storage Solutions - The demand for energy storage solutions in data centers is expected to rise due to the need for stable and green power supply, with a target for renewable energy usage in new data centers to exceed 80% by the end of 2025 [9]. - Companies like Sungrow Power Supply have established AIDC divisions, indicating a shift towards integrated energy systems that combine storage and renewable energy solutions [10][11]. Market Trends and Challenges - The AIDC energy storage market is anticipated to grow significantly, with projections indicating a compound annual growth rate exceeding 50% compared to traditional data center storage markets by 2028 [11]. - However, challenges remain in terms of technology, policy, and business models, including energy conversion losses, reliability, and market mechanisms [12].

Core Insights - The collaboration model of "energy storage + renewable energy" is likely to become a standard for Artificial Intelligence Data Centers (AIDC) in China [1][5] - Alibaba is investing significantly in AI infrastructure, with plans to increase energy consumption at its global data centers by tenfold by 2032 [1] - The overall investment in AIDC in China is projected to grow from 187.5 billion yuan in 2024 to 482.6 billion yuan by 2027 [1] Group 1: Importance of Stable Power Supply - Data centers, as the core infrastructure for AI, heavily rely on stable power supply, with power issues causing 54% of downtime incidents [2] - AI training has unique power load characteristics, making it challenging to predict electricity demand [2] - To ensure reliable power, data centers often use dedicated substations or power lines, avoiding shared power sources [2] Group 2: Development of Energy Storage in Data Centers - ByteDance is procuring large-scale lithium battery storage systems and renewable energy equipment to ensure uninterrupted power supply for its data centers [3] - The project aims to respond to carbon neutrality policies and reduce reliance on traditional power grids [3] - The demand for energy storage solutions is increasing due to the need for stable and green power supply in data centers [4] Group 3: Market Trends and Financial Performance - The energy storage business for data centers has shown significant growth, with companies like Double Ended Co. reporting a 113% increase in sales revenue from AIDC storage solutions [5] - Sunshine Power has shifted its main business focus to energy storage systems, indicating a trend towards integrated energy solutions for AIDC [5][6] - The global AIDC energy storage market is expected to reach several tens of GWh by 2028, with a compound annual growth rate exceeding 50% compared to traditional data center storage [6] Group 4: Challenges and Future Outlook - Despite the potential for energy storage to become standard, challenges remain in technology, policy, and business models [6] - Issues such as energy conversion losses, system reliability, and market mechanisms need to be addressed for widespread adoption [6] - Collaboration among industry players, policymakers, and grid companies is essential for overcoming these challenges [6]

Core Insights - The spending on data center construction is experiencing unprecedented growth, soon surpassing the growth rate of office building construction costs [1] - As of July this year, annualized spending on data center construction in the U.S. has surged to $41 billion, nearly matching the construction costs of all private office buildings [2] - Since July 2014, the increase in data center construction spending has reached 2200%, with major tech companies announcing significant new investment plans [3] Group 1 - The release of ChatGPT at the end of 2022 has significantly accelerated this trend, prompting official statistical agencies to categorize data centers as a separate statistical category [2] - Notable investors, such as David Einhorn, have warned that this cycle could lead to substantial capital destruction [3] - Tech giants including Nvidia, OpenAI, Oracle, SoftBank, and Meta have announced new construction plans totaling hundreds of billions of dollars, which will further elevate spending figures [3] Group 2 - Nvidia and OpenAI are planning a $100 billion super data center project with a power capacity of 10 gigawatts [3] - OpenAI, Oracle, and SoftBank have also announced their $500 billion "Stargate" project, while Meta is advancing its "Titan Cluster" data center initiative [3]