Core Insights - Goldman Sachs report highlights a significant 70% reduction in inference costs with the new TPU v7 chips from Google and Broadcom, indicating a major shift in the AI computing landscape [1][2][10]. Group 1: Cost Reduction and Implications - The 70% cost reduction signifies a fundamental change in the industry, moving beyond traditional hardware upgrades [2][5]. - The report emphasizes the importance of inference costs over training speeds, as the industry transitions from model training to deployment [4][10]. - The cost savings are attributed to three main factors: improved data transmission efficiency, tighter chip packaging, and specialized architecture of ASICs [7][8]. Group 2: Competitive Landscape - The TPU v7's cost is now comparable to NVIDIA's offerings, altering the competitive dynamics as companies reconsider their chip choices [9][10]. - The report suggests that the rise of ASICs represents a challenge to NVIDIA's dominance in the GPU market, indicating a shift towards customized solutions [11]. Group 3: Major Contracts and Market Movements - Anthropic's $21 billion order for custom ASICs marks a significant investment in dedicated AI infrastructure, reflecting a strategic shift in the industry [12][13]. - The funding for this order is backed by major players like Google and Amazon, highlighting the financial support for custom chip development [14][15]. Group 4: Role of Broadcom - Broadcom has transitioned to a key player in the AI chip market, acting as a contractor for major tech firms and providing essential interconnect technology [22][25]. - The company's business model, which includes upfront R&D fees and revenue sharing from chip sales, offers a more stable income compared to NVIDIA's model [24][27]. Group 5: Implications for China - The rise of ASICs and the reduction in inference costs may accelerate the development of China's own custom chip solutions, as companies seek alternatives to NVIDIA's GPUs [28][29]. - Chinese firms are increasingly investing in self-developed chips, aiming to create tailored solutions for their AI models [29][30]. - The report suggests that the focus should be on companies with core competencies in chip design and packaging technologies, rather than merely competing in low-cost chip production [31][34].

Core Viewpoint - The launch of Colossus 2, the world's first 1GW supercomputing cluster, marks a significant advancement in AI infrastructure, with plans to upgrade to 1.5GW by April and potentially reach 2GW, which could match the power consumption of major U.S. cities [2][12]. Group 1: Colossus 2 Overview - Colossus 2 is equipped with approximately 200,000 NVIDIA H100/H200 GPUs and around 30,000 NVIDIA GB200 NVL72 GPUs, significantly enhancing its computational power compared to its predecessor, Colossus 1, which was built in just 122 days [9][10]. - The cluster's 1GW capacity can power about 750,000 households, equivalent to the peak power demand of San Francisco [11]. - Once fully operational, Colossus 2 will house 555,000 GPUs, surpassing the GPU counts of Meta, Microsoft, and Google [13][14]. Group 2: Implications for AI Development - The advancements in Colossus 2 are expected to facilitate the development of Grok 5, which is projected to have parameters around 6 trillion, more than double that of Grok 4 [15][18]. - With the recent $20 billion funding round for xAI, the scaling capabilities for Grok 5 are increasing, leading to larger model parameters and faster training and deployment speeds [18][19]. - The rapid development of AI models is seen as a competitive advantage in the industry, emphasizing that speed is a crucial factor in the AI era [20]. Group 3: Energy Supply Concerns - The construction of large data centers like Colossus 2 is contributing to a projected annual electricity demand growth of 4.8% over the next decade, which is unprecedented for the U.S. energy system [27]. - The imbalance between rapidly increasing demand and slow supply growth is causing concerns about the stability of the power grid, leading to potential rolling blackouts for 67 million residents in 13 states during extreme weather [5][22][23]. - PJM, the regional transmission organization, is struggling to maintain supply-demand balance and has proposed measures to reduce peak demand from data centers, which have faced opposition from major tech companies [32][34].

Core Insights - The article emphasizes that NVIDIA remains the dominant player in AI computing power, providing significantly better performance per dollar compared to AMD [1][30]. - A report from Signal65 reveals that under certain conditions, NVIDIA's cost for generating the same number of tokens is only one-fifteenth of AMD's [4][30]. Performance Comparison - NVIDIA's platform offers 15 times the performance per dollar compared to AMD when generating tokens [1][30]. - The report indicates that for complex models, NVIDIA's advantages become more pronounced, especially in the context of the MoE (Mixture of Experts) architecture [16][24]. MoE Architecture - The MoE architecture allows models to split parameters into specialized "expert" sub-networks, activating only a small portion for each token, which reduces computational costs [10][11]. - However, communication delays between GPUs can lead to idle time, increasing costs for service providers [13][14]. Cost Analysis - Despite NVIDIA's higher pricing, the overall cost-effectiveness is better due to its superior performance. For instance, the GB200 NVL72 costs $16 per GPU per hour, while AMD's MI355X costs $8.60, making NVIDIA's price 1.86 times higher [27][30]. - The report concludes that at 75 tokens per second per user, the performance advantage of NVIDIA is 28 times, resulting in a cost per token that is one-fifteenth of AMD's [30][35]. Future Outlook - AMD's competitiveness is not entirely negated, as its MI325X and MI355X still have applications in dense models and capacity-driven scenarios [38]. - AMD is developing a cabinet-level solution, Helios, which may narrow the performance gap in the next 12 months [39].

Core Viewpoint - The global AI technology is shifting from training to inference, driving a new growth opportunity in the hardware supply chain [2] Summary by Category Overall - The transition from training-dominated AI to inference-driven AI is significantly increasing the demand for inference computing power, driven by the iteration of multimodal large models like Google's Gemini 3 Pro and OpenAI's Sora 2 [2] - Major cloud service providers (CSPs) are expected to increase capital expenditures, with a forecast of $431 billion by 2025, a 65% year-on-year increase, and potentially reaching $602 billion by 2026 [2] - Sovereign AI initiatives are being launched globally, such as the U.S. "Gateway to the Stars" plan with an investment of approximately $500 billion and the EU's plan to invest $21.5 billion in AI super factories, contributing to a high-growth phase in global AI infrastructure [2] - By 2030, global AI data center capacity is projected to reach 156 GW, accounting for 71% of total data center demand [2] Cloud Side - PCB: AI servers are bringing clear value increases, with Nvidia's DGX H100 single GPU corresponding to a PCB value of $211, a 21% increase from the previous generation; the GB200 NVL72 raises the single GPU value to $346 [3] - The domestic high-end PCB capacity is expected to be released in 2026 to support downstream demand, driving upgrades in upstream materials [3] - Storage: The structural supply-demand imbalance due to AI demand has led to significant price increases in DRAM and NAND Flash, with a shift in investment focus towards high-value products expected in 2026 [3] - KVCache technology is accelerating the replacement of HDDs with QLC SSDs, with a projected 30% penetration rate in the enterprise SSD market by 2026 [3] Optical Interconnect - Optical interconnect technology is entering a new era as a key component of AI computing clusters, with optical switches meeting the interconnection needs of large-scale AI clusters due to their high bandwidth, low latency, and low power consumption [4] - The MEMS-based technology route currently dominates, with domestic manufacturers actively engaging in various segments of the global supply chain [4] End Side - AI Phones: The AI phone market is expected to maintain moderate growth in 2025, with competition shifting towards end-side AI capabilities [5] - The operating systems of mobile phones are evolving from "application launchers" to "system-level intelligent agents," with flagship chips from Apple and Android continuously enhancing NPU computing power [5] - AR Glasses: The integration of AI and AR in smart glasses is seen as the future of wearable devices, with the market experiencing rapid growth [5] - The optical imaging module solutions for AR glasses are expected to favor light guide technology due to its advantages in clarity and size, while LCOS remains the mainstream for consumer products [5] Recommendations - The company suggests focusing on sectors benefiting from the shift to inference computing and hardware upgrades, including: - PCB and upstream materials: Shenghong Technology, Huitian Technology, Jingwang Electronics, Guanghe Technology, Dongcai Technology [6] - Storage and equipment: Beijing Junzheng, Zhaoyi Innovation, Jucheng Co., Jingzhida [6] - Optical interconnect: Yintan Zhikong, Saiwei Electronics [6] - End-side AI: GoerTek, Luxshare Precision, Baiwei Storage, Longqi Technology, Crystal Optoelectronics, Zhongke Lanyun, Howey Group, Sunny Optical Technology [6]

Core Insights - The AI inference business is demonstrating remarkable profitability amid intense competition in the AI sector, with a recent Morgan Stanley report providing a comprehensive analysis of the global AI computing market's economic returns [1][3][8] Company Performance - A standard "AI inference factory" shows average profit margins exceeding 50%, with Nvidia's GB200 chip leading at nearly 78% profit margin, followed by Google's TPU v6e pod at 74.9% and Huawei's solutions also performing well [1][3][5] - AMD's AI platforms, specifically the MI300X and MI355X, are facing significant losses with profit margins of -28.2% and -64.0% respectively, attributed to high costs and low output efficiency [5][8] Market Dynamics - The report introduces a "100MW AI factory model" that evaluates total ownership costs, including infrastructure, hardware, and operational costs, using token output as a revenue measure [7] - The future AI landscape will focus on building technology ecosystems and next-generation product layouts, with Nvidia solidifying its lead through a clear roadmap for its next platform, "Rubin," expected to enter mass production in Q2 2026 [8]

Core Viewpoint - The article emphasizes the growing trend of liquid cooling technology in data centers, particularly in relation to AI applications and the performance improvements it offers over traditional cooling methods. Group 1: Liquid Cooling Technology - Liquid cooling servers outperform air-cooled versions by 25% in performance and reduce power consumption by 30% [1] - The adoption of liquid cooling is expected to rise significantly, with projections indicating that over 65% of new systems will utilize this technology by next year [1] - The cost-effectiveness of liquid cooling solutions is highlighted, as they are seen as a way to lower overall expenses while maintaining high performance [1][6] Group 2: Company Performance and Projections - NVIDIA is projected to ship approximately 30,000 units of the GB200 and 10,000 units of the GB300, with an additional 200,000 units of the B200 single card expected [5] - The GB300 is anticipated to see increased shipments, with expectations raised to 100,000 units [5] - The server assembly sector is experiencing a resurgence, indicating a positive outlook for companies involved in this space [5] Group 3: Industry Trends - All Azure regions are now equipped to support liquid cooling, enhancing the flexibility and efficiency of data center operations [3] - The industry is moving towards building gigawatt and multi-gigawatt data centers, with over 2 gigawatts of new capacity established in the past year [4] - The trend towards liquid cooling is accelerating, with domestic manufacturers beginning to capture market share [6]

Group 1 - The core viewpoint of the article highlights the introduction of the AI computing cluster solution Cloud Matrix 384 at the WAIC2025 conference, which utilizes a fully interconnected topology for efficient collaboration [1] - The Atlas 900 A3 SuperPoD features a super-node architecture with ultra-large bandwidth of 392GB/s, ultra-low latency of less than 1 microsecond, and exceptional performance of 300 PFLOPs, enhancing the training performance of models like LLaMA3 by over 2.5 times compared to traditional clusters [1] - The Ascend 910C outperforms NVIDIA's GB200 NVL72 in system-level BF16 computing power (300 PFLOPS), HBM capacity (49.2TB), and bandwidth (1229TB/s) [1] Group 2 - The semiconductor equipment ETF (159516) tracks the semiconductor materials and equipment index (931743), focusing on the materials and equipment sectors within the semiconductor industry, covering companies involved in manufacturing, processing, and testing [1] - The index constituents are primarily companies with significant technological advantages and core competitiveness in semiconductor materials and equipment, aiming to reflect the overall performance and industry development trends of listed companies in this segment [1] - Investors without stock accounts can consider the Guotai Zhongzheng Semiconductor Materials and Equipment Theme ETF Initiated Link A (019632) and Link C (019633) [1]

Group 1 - Amazon's cloud computing division has developed specialized hardware for cooling next-generation NVIDIA GPUs, which are widely used for AI-related computing tasks [1] - The high energy consumption of NVIDIA GPUs necessitates additional cooling equipment for companies utilizing these processors [1] - Amazon previously considered building data centers with liquid cooling systems but found existing solutions inadequate for their scale [1] Group 2 - The newly developed In-Row Heat Exchanger (IRHX) can be integrated into existing and new data centers to support high-power NVIDIA GPUs [1] - Customers can now access a new AWS service through the P6e compute instance, which utilizes NVIDIA's high-density computing hardware [2] - Amazon has reduced reliance on third-party suppliers by developing its own infrastructure hardware, contributing to improved profitability [2]

Investment Rating - The report maintains a "Positive" investment rating for the supernode industry, driven by the explosive growth of model parameters and the shift in computing power demand from single points to system-level integration [3]. Core Insights - The supernode trend is characterized by a dual expansion of high-density single-cabinet and multi-cabinet interconnection, balancing communication protocols and engineering costs [4][5]. - Domestic supernode solutions, represented by Huawei's CloudMatrix 384, achieve a breakthrough in computing power scale, surpassing single-card performance limitations [4][5]. - The industrialization of supernodes will reshape the computing power industry chain, creating investment opportunities in server integration, optical communication, and liquid cooling penetration [4][5][6]. - Current market perceptions underestimate the cost-performance advantages of domestic solutions in inference scenarios and overlook the transformative impact of computing network architecture on the industry chain [4][7]. Summary by Sections 1. Supernode: New Trends in AI Computing Networks - The growth of large model parameters and architectural changes necessitate understanding the two dimensions of computing power expansion: Scale-up and Scale-out [15]. - Scale-up focuses on tightly coupled hardware, while Scale-out emphasizes elastic expansion to support loosely coupled tasks [15][18]. 2. Huawei's Response to Supernode Challenges - Huawei's CloudMatrix 384 represents a domestic paradigm for cross-cabinet supernodes, achieving a computing power scale 1.7 times that of NVIDIA's NVL72 [4][5][6]. - The design of supernodes must balance model training and inference performance with engineering costs, particularly in multi-GPU inference scenarios [69][77]. 3. Impact on the Industry Chain - The industrialization of supernodes will lead to a more refined division of labor across the computing power industry chain, with significant implications for server integration and optical communication [6][4]. - The demand for optical modules driven by Huawei's CloudMatrix is expected to reach a ratio of 1:18 compared to GPU demand [6]. 4. Key Company Valuations - The report suggests focusing on companies involved in optical communication, network devices, data center supply chains, copper connections, and AI chip and server suppliers [5][6].

Core Viewpoints - The article discusses the rapid growth and potential of the computing power leasing industry, particularly through the lens of CoreWeave, a significant player in this sector [2][11]. Company Overview - CoreWeave was established in 2017, originally as a cryptocurrency mining company, and has since pivoted to focus on AI cloud and infrastructure services, operating 32 data centers by the end of 2024 [2][3]. - The company has deployed over 250,000 GPUs, primarily NVIDIA products, and is a key provider of high-performance infrastructure services [2][3]. Business Model - CoreWeave offers three main services: bare-metal GPU leasing, management software services, and application services, with a focus on GPU leasing as the core offering [3][4]. - Revenue is generated primarily through two models: commitment contracts (96% of revenue) and on-demand payment, allowing flexibility for clients [4][5]. Financial Performance - In 2024, CoreWeave's revenue reached $1.915 billion, a year-over-year increase of over seven times, with Q1 2025 revenue at $982 million, reflecting a fourfold increase [8][9]. - The company has a remaining performance obligation of $15.1 billion, indicating strong future revenue potential [8]. Competitive Advantages - CoreWeave has optimized GPU utilization rates and efficiency, achieving significant performance improvements in AI training and inference tasks [7]. - The company has established strong relationships with NVIDIA, ensuring priority access to cutting-edge chips and technology [6][7]. Market Outlook - The AI infrastructure market is projected to grow from $79 billion in 2023 to $399 billion by 2028, with a compound annual growth rate of 38%, highlighting the industry's potential [11]. - The computing power leasing sector is expected to play a crucial role in the digital economy, driven by increasing demand for AI capabilities [11][14]. Future Growth Strategies - CoreWeave plans to expand its customer base, explore new industries, and enhance vertical integration with strategic partnerships [10]. - The management aims to leverage existing contracts and maintain a low leverage asset structure to support growth [10].