文 | 半导体产业纵横 AI服务器需求还在持续上升。 大摩预测，仅英伟达平台，AI服务器机柜的需求就将从2025年的约2.8万台，在2026年跃升至至少6万 台，实现超过一倍的增长。与此同时，AMD的Helios服务器机架项目（基于MI400系列）也获得了良好 进展，进一步加剧了市场对先进AI硬件的需求。 | | | | | 英伟达AI服务器路线 | | | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | | HIDD | H200 | 8200 | GE200 NVLT2 | GB300 NVL72 | VR200 NVL144 | VR200 MVL144 CPX | UVR300 | | 发布时间 | 22年Q4 | 24年Q2 | 25年Q1 | 24年Q4 | 25年Q3 | 26年H2 | 26年東 | 27年H2 | | CPU | x86 | x86 | x86 | Grace | Grace | Vera | Vera | Vera | | 时提形式 | CoWo5-S | CoWo5 S | CoWoS-L | C ...

作者按：re：Invent是全球最大云厂商亚马逊AWS一年一度的战略和技术大会。re：Invent 2025于12月1日-5日在美国拉斯维加斯召开。我们 在现场，记录美国云计算行业这一年最新的变化。 以下是正文： 美国西部时间12月3日，全球最大云计算厂商亚马逊AWS在re:Invent 2025发布了新款AI芯片和新款自研模型。 （亚马逊AWS CEO马特·加曼在re:Invent 2025发布Trainium 3） 亚马逊AWS CEO（首席执行官）马特·加曼（Matt Garman）宣布，亚马逊AWS自研的AI芯片Trainium 3已经上线，这是一款3nm（纳米）工艺制程的芯 片。亚马逊AWS下一代AI芯片Trainium 4也在研发设计阶段。 亚马逊AWS还发布了Amazon Nova 2系列的自研模型。其中包括Lite（轻量化版）、Pro（专业版）、Sonic（声音模型）、Omni（多模态推理模型）这四 款模型。目前Amazon Nova系列已有数万企业客户，包括海信、印孚瑟斯（印度第二大IT企业）等。 01 亚马逊AWS已部署百万枚自研AI芯片 马特·加曼提到，目前亚马逊已经部署了超过100万枚 ...

Core Viewpoint - Nvidia's third-quarter revenue growth accelerated for the first time in two years, with data center revenue reaching a record high, reflecting strong demand for AI infrastructure [2][11][12] Financial Performance - Revenue: In Q3, Nvidia reported revenue of $57.01 billion, a year-on-year increase of approximately 62%, surpassing analyst expectations of $55.19 billion [6][11] - EPS: The adjusted non-GAAP EPS for Q3 was $1.30, a 60% year-on-year increase, exceeding analyst expectations of $1.26 [7] - Gross Margin: The adjusted gross margin for Q3 was 73.6%, slightly below the expected 74.0%, but the guidance for Q4 indicates an increase to 75% [7][16] Segment Performance - Data Center: Q3 revenue from data centers was $51.2 billion, a 66% year-on-year increase, exceeding analyst expectations [8][12] - Gaming and AI PC: Revenue from gaming and AI PC was $4.3 billion, a 30% year-on-year increase [8] - Professional Visualization: Revenue was $760 million, a 56% year-on-year increase [8] - Automotive and Robotics: Revenue was $592 million, a 32% year-on-year increase [8] Guidance and Future Outlook - Revenue Guidance: For Q4, Nvidia expects revenue of $65 billion, indicating a year-on-year growth of over 65% [10][14] - Gross Margin Guidance: The expected gross margin for Q4 is 75%, marking the first year-on-year increase in six quarters [16] - Future Revenue Potential: Nvidia's CFO stated that new chips are expected to generate $500 billion in revenue over the next few quarters [18][19] Market Position and Strategy - Nvidia's strong customer base includes major companies like Microsoft, Amazon, Alphabet, and Meta, which collectively account for over 40% of its sales [18] - The company has secured $500 billion in chip orders for 2025 and 2026, indicating robust future demand [18][21] - Nvidia's strategy includes investing in AI infrastructure and maintaining strong partnerships to enhance its market position [21]

Core Insights - The article discusses the significant investment and development of artificial intelligence (AI) infrastructure in Malaysia, particularly through a collaboration between NVIDIA and YTL Power International, which has led to the establishment of a large data center in Johor, Malaysia [2][4]. Group 1: Investment and Development - NVIDIA announced a $4.3 billion AI infrastructure development plan with YTL Power, with $2.4 billion already allocated for a 200 MW data center cluster [2]. - The data center in Johor is equipped with high-performance supercomputers utilizing NVIDIA's advanced GB200 chips, capable of processing 1.8 TB of information per second [1]. - The facility is part of a larger strategy to position Malaysia as a digital hub, aligning with the country's ambition to become a leading AI nation by 2030 [4]. Group 2: Regional Competition and Growth - Analysts predict that Malaysia's data center capacity could surpass that of Singapore within the next five years, driven by the AI boom and the global demand for AI infrastructure [5]. - Major tech companies, including Amazon, Google, and Microsoft, are expected to invest $240 billion in expanding their operations in the Asia-Pacific region over the next five years [7]. - The Asia-Pacific data center capacity is projected to grow from 12 GW in 2024 to over 29 GW by 2030, making it the second-largest data center market globally [7]. Group 3: Broader Industry Trends - High-profile collaborations and investments are emerging across the region, with notable projects in India, South Korea, and Thailand, indicating a regional trend towards expanding AI and data center capabilities [10][12][14]. - The surge in investment is driving up stock prices for data center companies, with DCI Indonesia's market value exceeding $37 billion, making it the second-highest valued company in Indonesia [15]. - The rapid development of data centers raises concerns about resource consumption, particularly regarding electricity and water supply, highlighting the need for sustainable energy solutions [19].

Core Insights - Morgan Stanley's report highlights the increasing value of liquid cooling components in AI systems, with the GB300 NVL72 system's cooling component valued at $49,860, a 20% increase from the GB200 NVL72 system [1] - The next-generation Vera Rubin NVL144 platform is expected to see a 17% increase in cooling component costs, reaching approximately $55,710, driven by rising cooling demands for compute and switch trays [1][4] Industry Trends - The demand for liquid cooling solutions is surging due to the exponential increase in data center computing density, as traditional air cooling methods are inadequate for high-density computing equipment [5] - NVIDIA's GPUs are experiencing significant power increases, with TDPs projected to reach 2,300W for the Vera Rubin platform and 3,600W for the VR300 platform, making cooling capabilities a critical bottleneck for performance [5] Market Growth - IDC forecasts that China's liquid cooling server market will reach $3.39 billion by 2025, with a year-on-year growth of 42.6%, and a compound annual growth rate of 48% from 2025 to 2029, potentially exceeding $16.2 billion by 2028 [6] - NVIDIA's related chip cooling demand is expected to be a major driver of market growth [6] Stock Performance - Several liquid cooling concept stocks have seen significant price increases this year, with companies like Siyuan New Materials, Yinvik, and Kexin New Energy reporting over 50% revenue growth year-on-year [7] - Notable companies such as Ice Wheel Environment and Silver Wheel Co. have received extensive institutional research interest, indicating strong market confidence in the liquid cooling sector [7][10]

Core Insights - The essence of AI competition has shifted to an "electricity competition," with investment trends moving from traditional IT infrastructure to non-IT infrastructure such as power, cooling, and materials [1][6][12] - By 2030, China's non-IT infrastructure investment related to AI is expected to reach 800 billion RMB, with power systems dominating at 38%, followed by metals for data center construction at 12% and advanced cooling systems at 10% [2][9] Investment Trends - The investment wave in AI infrastructure is expanding beyond traditional chips and servers to include essential non-IT infrastructure [2] - Total capital expenditure for AI in China is projected to grow to 2-2.5 trillion RMB by 2030, with non-IT infrastructure accounting for one-third of this total [2] Power Consumption and Data Centers - The energy consumption of data centers in China is expected to grow at a compound annual growth rate (CAGR) of 18%, increasing from 102 TWh in 2024 to 277 TWh by 2030, representing 29% of global data center electricity consumption [3][9] - The rapid increase in power consumption is driven by the accelerated adoption of AI data centers (AIDC), which have significantly higher power requirements than traditional data centers [4] Key Drivers of Investment - The report identifies three main drivers for the surge in investment: the proliferation of AIDC, the deployment of high-performance chips, and the increasing power density of server cabinets [4][5][9] Opportunities in Power Supply - China has significant advantages in AI power infrastructure, including ample generation capacity, lower industrial electricity prices (30-60% lower than developed markets), a leading position in renewable energy supply chains, and a relatively young and robust power grid [12] - Five major investment opportunities are highlighted: nuclear power, electrical equipment, battery energy storage systems (BESS), diesel generators, and advanced power supply technologies [13][17][20][22][24] Cooling and Materials - Efficient cooling and essential raw materials are critical for AI infrastructure, with significant investment potential [25] - Liquid cooling technology is expected to grow rapidly, with a projected market size of 79 billion RMB by 2030, driven by the need for efficient heat management in high-density AI environments [26] - The demand for key metals such as copper and aluminum is also expected to rise, with copper consumption in AI data centers projected to reach approximately 1 million tons by 2030, accounting for 5-6% of national demand [27]

Core Insights - The article emphasizes that electricity supply is becoming a critical bottleneck for AI development, with China potentially having an advantage over the U.S. in this regard [1][3]. Group 1: Electricity Demand and Supply - The demand for electricity in AI data centers is expected to grow exponentially, with predictions indicating that by 2030, a single AI data center could require up to 8 GW of power, equivalent to eight large nuclear reactors [2]. - The existing electricity infrastructure in the U.S. is struggling to meet this surging demand, with some data centers in Northern Virginia facing power supply wait times of up to seven years [2]. - Companies like xAI are resorting to renting portable gas generators due to prolonged electricity supply wait times, leading to increased operational costs [2]. Group 2: U.S. Electricity Grid Challenges - The U.S. electricity grid is under significant strain, with Bernstein predicting that the average annual growth rate of electricity demand will reach 2.3% from 2024 to 2030, with regional variations as high as 13.4% in Texas [7]. - Investment in the electricity grid is primarily focused on maintaining existing infrastructure rather than expanding capacity, with only 28% of distribution investment allocated for expansion [8]. - The article highlights the risk of grid failures, citing historical outages that stemmed from insufficient investment in infrastructure [11]. Group 3: Innovations in Power Supply - Solid State Transformers (SST) are proposed as a solution to the electricity supply challenges faced by data centers, offering higher efficiency and reduced space requirements compared to traditional systems [16][21]. - SST technology can achieve an efficiency of up to 98%, significantly improving power delivery for AI workloads and reducing copper usage by 45% [21][27]. - The market potential for SST is substantial, with estimates suggesting a market size of 800-1000 billion yuan by 2030 if penetration reaches 20% in new AI data centers [27]. Group 4: Future Outlook - The article suggests that as the demand for electricity continues to rise, technology companies will likely engage in large-scale procurement of power resources, further straining the already tight electricity grid [7]. - The transition to renewable energy sources poses additional challenges for grid stability, necessitating innovative solutions to balance supply and demand [11].

Core Viewpoint - The investment opportunities in the artificial intelligence (AI) industry chain can be better grasped through a dual layout of chips and applications [1] Group 1: Chip Engine - The domestic computing power market is experiencing significant growth, with a projected market size of $50 billion and an annual compound growth rate of approximately 30% [3] - The domestic GPU market presents substantial investment opportunities, with several domestic computing power companies seeing notable increases in market value despite previous losses [2][3] - The penetration rate of domestic computing power chips is expected to rise from 20-30% this year to over 50% in the coming years, driven by advancements in domestic chip technology and restrictions on foreign chip procurement [4][6] Group 2: Application Engine - The application side is benefiting from the continuous evolution of large models, with domestic companies like Meituan releasing new versions of their AI models [8] - The demand for applications, including humanoid robots and AI-enabled consumer electronics, is expected to grow as the capabilities of large models improve, leading to clearer business models and enhanced monetization [9] - The AI market is transitioning from a training focus to a reasoning focus, with reasoning-side demand likely to dominate future AI computing power needs [9]

Core Viewpoint - The liquid cooling sector is poised to become a key investment direction in the AI computing revolution, with expectations of rapid growth in penetration rates and market demand [2][9][33]. Group 1: Market Dynamics - In August, the AI market sentiment expanded from optical modules and PCBs to include server power supplies and liquid cooling, with stocks like Invec and others reaching historical highs [1][14]. - The liquid cooling segment, previously seen as a laggard in AI trends, is now gaining fundamental support, indicating a shift from mere thematic speculation to a more robust market presence [1][33]. - The anticipated release of NVIDIA's GB200 and GB300 series chips is expected to drive the adoption of liquid cooling solutions, marking 2025 as a pivotal year for the sector [9][24]. Group 2: Technological Advancements - Liquid cooling technology is becoming a standard for next-generation computing products, with significant improvements in performance and efficiency compared to traditional air cooling [2][5]. - The transition from air cooling to liquid cooling is driven by the increasing power requirements of GPUs, with the B200 and B300 chips necessitating liquid cooling solutions due to their high power outputs [5][6]. - The introduction of new products supporting liquid cooling, such as AMD's MI350 series and Broadcom's Ethernet switches, signifies a broader acceptance and integration of liquid cooling technologies [7][9]. Group 3: Industry Growth Projections - TrendForce projects that the penetration rate of liquid cooling in AI data centers will increase from 14% in 2024 to 33% in 2025, driven by the mass production of NVIDIA's new server models [8][9]. - The market for liquid cooling components, including cold plates and heat exchange systems, is expected to expand significantly, with estimates suggesting a potential market size of 100 billion RMB by 2026 [12][9]. - The liquid cooling industry is anticipated to mirror the growth trajectories of optical modules and PCBs, with a strong likelihood of emerging leaders in the market [33][24]. Group 4: Competitive Landscape - Companies like Invec are positioned as key players in the liquid cooling supply chain, with significant stock price increases reflecting market confidence in their growth potential [14][17]. - The competitive landscape is characterized by two types of companies: system integrators like Invec and manufacturers focusing on individual components, each following distinct growth paths [25][26]. - The entry of domestic manufacturers into the global liquid cooling market is expected to follow the patterns seen in the PCB industry, with potential for substantial market share gains [19][24].

Core Viewpoint - The liquid cooling sector is poised to become a core investment direction in the AI computing revolution, with increasing penetration rates and a shift from speculative trading to fundamental support [4][6][10]. Group 1: Market Dynamics - In August, the AI market expansion spread from optical modules and PCBs to server power supplies and liquid cooling [1]. - Companies like Invec, Oulu Tong, and Megmeet have seen their stock prices reach historical highs, indicating a growing interest in the liquid cooling sector [2]. - The liquid cooling concept has been developing for nearly two years, but the actual market implementation has been slow [5][6]. Group 2: Technological Advancements - The demand for liquid cooling technology is becoming urgent with the release of NVIDIA's Blackwell series chips, which require advanced cooling solutions [8][10]. - The transition from air cooling to liquid cooling is driven by the increasing power requirements of GPUs, with the B200 and B300 chips necessitating liquid cooling solutions [7][8]. - The penetration rate of liquid cooling in AI data centers is expected to rise significantly, from 14% in 2024 to 33% in 2025, as major cloud service providers upgrade their infrastructure [12]. Group 3: Investment Opportunities - The liquid cooling market is projected to reach 100 billion RMB by 2026, driven by the increasing demand for AI servers and the expansion of liquid cooling solutions [15]. - Companies in the liquid cooling supply chain, such as Invec, have seen substantial stock price increases, with Invec's stock rising nearly 98% in August alone [17]. - The liquid cooling sector is expected to replicate the success of the optical module and PCB markets, with potential for significant growth and investment opportunities [20][27]. Group 4: Competitive Landscape - The liquid cooling industry can be divided into two categories: system integrators like Invec and manufacturers focusing on individual components [28]. - The entry of domestic manufacturers into the global liquid cooling supply chain is anticipated, similar to the previous trends seen in the PCB market [22][24]. - The high entry barriers in the liquid cooling sector are primarily due to the need for advanced system integration capabilities, which many manufacturers currently lack [32][35].