我们认为公司除了算力芯片以外，网络通信产品也具备较深厚布 局，网络收入已经开始高速成长。公司有望成为重要的 AI 硬件平 台型公司。下游云厂商模型继续迭代、推理需求增长，有望成为 公司增长核心驱动力。主权 AI 的需求有望贡献额外需求，降低云 厂商需求波动。我们预计公司 FY26~FY28 GAAP 净利润分别为 1111.5、1641.6、1882.8 亿美元，维持"买入"评级。 风险提示 业绩简评 2025 年 8 月 28 日公司披露 FY26Q2（25.5~25.7）业绩，FY26Q2 实现营收 467.43 亿美元，同比+55.6%，环比+6.1%；公司 FY26Q2 GAAP 毛利率为 72.4%，GAAP 净利润为 264.22 亿美元；公司 FY26Q2 Non-GAAP 毛利率为 72.7%，Non-GAAP 净利润为 257.83 亿美元。 公司指引 FY26Q2 营收为 540 亿美元（±2%），其中未考虑任何 H20 对华出货。公司指引 FY26Q2 GAAP 毛利率为 73.3%，Non-GAAP 毛 利率为 73.5%。 经营分析 数据中心业务持续增长，产品迭代稳步推进。FY26Q ...

Core Viewpoint - The company, Lingyi iTech, is expanding its business beyond being a member of the Apple supply chain, signaling a strategic shift towards AI cooling solutions and humanoid robotics, particularly through its recent recruitment for a senior engineer in NVIDIA liquid cooling technology [1][2][4]. Group 1: Company Developments - Lingyi iTech's stock price increased by over 7% on August 27, 2023, but closed at 14.78 CNY per share, marking a 63.68% increase since April 2023 [1]. - The company has been a supplier for Apple products since 2009, providing components for various devices, and is now venturing into AI cooling and humanoid robotics [2][4]. - Lingyi iTech has introduced a comprehensive cooling solution for AI infrastructure, including liquid cooling modules and systems, to meet the increasing thermal demands of high-performance AI servers [3][4]. Group 2: Market Position and Financial Performance - Lingyi iTech's revenue for Q1 2023 was 11.494 billion CNY, a year-on-year increase of 17.11%, with a net profit of 565 million CNY, up 23.52% [5][6]. - The company anticipates a net profit of 900 million to 1.14 billion CNY for the first half of 2023, representing a growth of 31.57% to 66.66% compared to the previous year [5]. - Lingyi iTech is investing at least 200 million CNY annually in robotics over the next three years, aiming to establish it as a core business segment alongside consumer electronics and automotive sectors [5].

Core Insights - The article discusses the cost comparison between H100 and GB200 NVL72 servers, highlighting that the total upfront capital cost for GB200 NVL72 is approximately 1.6 to 1.7 times that of H100 per GPU [2][3] - It emphasizes that the operational costs of GB200 NVL72 are not significantly higher than H100, primarily due to the higher power consumption of GB200 NVL72 [4][5] - The total cost of ownership (TCO) for GB200 NVL72 is about 1.6 times higher than that of H100, indicating that GB200 NVL72 needs to be at least 1.6 times faster than H100 to be competitive in terms of performance/TCO [4][5] Cost Analysis - The price of H100 servers has decreased to around $190,000, while the total capital cost for a typical hyperscaler server setup can reach $250,866 [2][3] - For GB200 NVL72, the upfront capital cost per server is approximately $3,916,824, which includes additional costs for networking, storage, and other components [3] - The capital cost per GPU for H100 is $31,358, while for GB200 NVL72, it is $54,400, reflecting a significant difference in initial investment [3] Operational Costs - The operational cost per GPU per month for H100 is $249, while for GB200 NVL72, it is $359, indicating a smaller margin in operational expenses [4][5] - The electricity cost remains constant at $0.0870 per kWh across both systems, with a utilization rate of 80% and a Power Usage Effectiveness (PUE) of 1.35 [4][5] Recommendations for Nvidia - The article suggests that Nvidia should enhance its benchmarking efforts and increase transparency to benefit the machine learning community [6][7] - It recommends expanding benchmarking beyond NeMo-MegatronLM to include native PyTorch, as many users prefer this framework [8][9] - Nvidia is advised to improve diagnostic and debugging tools for the GB200 NVL72 backplane to enhance reliability and performance [9][10] Benchmarking Insights - The performance of training models like GPT-3 175B using H100 has shown improvements in throughput and efficiency over time, with significant gains attributed to software optimizations [11][12] - The article highlights the importance of scaling in training large models, noting that weak scaling can lead to performance drops as the number of GPUs increases [15][17] - It provides detailed performance metrics for various configurations, illustrating the relationship between GPU count and training efficiency [18][21]

Summary of Conference Call Notes Company and Industry - The discussion primarily revolves around Nvidia's GPU products, specifically the H100 and GB200 NVL72 models, and their performance in machine learning training environments. Core Points and Arguments 1. **Benchmarking and Performance Analysis** - The report presents benchmark results from over 2,000 H100 GPUs, analyzing metrics such as mode fops utilization (MFU), total cost of ownership (TCO), and cost per training 1 million tokens [5][6][12] - The analysis includes energy consumption comparisons, framing power efficiency in a societal context by comparing GPU energy use to average U.S. household energy usage [5][6] 2. **Cost Analysis** - The price of an H100 server has decreased to approximately $10,000, with total upfront capital costs reaching around $250,000 for a typical hyperscaler [14] - The GB200 NVL72 server costs about $1.1 million per rack, with all-in costs reaching approximately $1.5 million per rack [15] - The all-in capital cost per GPU for the GB200 NVL72 is estimated to be 1.1x to 1.7x that of the H100 [15] 3. **Operational Costs** - The operational cost per GPU for the GB200 NVL72 is not significantly higher than that of the H100, but the GB200 consumes 1200W per chip compared to 700W for the H100, impacting overall operational expenses [17][18] - Total cluster operating costs per month per GPU are $249 for H100 and $359 for GB200 NVL72, indicating a higher cost for the latter [19] 4. **Reliability Issues** - Current reliability challenges with the GB200 NVL72 are noted, with no large-scale training runs completed yet due to ongoing software maturation [7][8] - Nvidia is expected to work closely with partners to address these reliability issues, which are critical for the ecosystem's success [8] 5. **Software Improvements** - Significant improvements in training throughput have been observed, with MFU increasing from 2.5% to 5% over 12 months, attributed to software optimizations [31][33] - The cost to train GPT-175B has decreased from $218,000 in January 2022 to $12,000 by December 2022, showcasing the impact of software enhancements on cost efficiency [34] 6. **Recommendations for Nvidia** - Suggestions include expanding benchmarking efforts and increasing transparency to aid decision-making in the ML community [22][24] - Nvidia should also broaden its benchmarking focus beyond NeMo-MegatronLM to include native PyTorch frameworks [25] - Accelerating the development of diagnostics and debugging tools for the GB200 NVL72 is recommended to improve reliability [25] Other Important Content - The report emphasizes the importance of effective training and the need for Nvidia to address reliability challenges to maintain competitiveness in the GPU market [6][8] - The analysis of power consumption indicates that training large models like GPT-175B requires significant energy, equivalent to the annual consumption of multiple U.S. households [35][48] - The discussion on scaling performance highlights the differences between strong and weak scaling in compute resources, which is crucial for optimizing training processes [39][40]

Core Viewpoint - The article discusses the performance and cost comparison between NVIDIA's H100 and GB200 NVL72 GPUs, highlighting the potential advantages and challenges of the GB200 NVL72 in AI training environments [30][37]. Group 1: Market Predictions and Performance - After the ODM performance announcement, institutions raised the forecast for GB200/300 rack shipments in 2025 from 30,000 to 34,000, with expected shipments of 11,600 in Q3 and 15,700 in Q4 [3]. - Foxconn anticipates a 300% quarter-over-quarter increase in AI rack shipments, projecting a total of 19,500 units for the year, capturing approximately 57% of the market [3]. - By 2026, even with stable production of NVIDIA chips, downstream assemblers could potentially assemble over 60,000 racks due to an estimated 2 million Blackwell chips carried over [3]. Group 2: Cost Analysis - The total capital expenditure (Capex) for H100 servers is approximately $250,866, while for GB200 NVL72, it is around $3,916,824, making GB200 NVL72 about 1.6 to 1.7 times more expensive per GPU [12][13]. - The operational expenditure (Opex) for GB200 NVL72 is slightly higher than H100, primarily due to higher power consumption (1200W vs. 700W) [14][15]. - The total cost of ownership (TCO) for GB200 NVL72 is about 1.6 times that of H100, necessitating at least a 1.6 times performance advantage for GB200 NVL72 to be attractive for AI training [15][30]. Group 3: Reliability and Software Improvements - As of May 2025, GB200 NVL72 has not yet been widely adopted for large-scale training due to software maturity and reliability issues, with H100 and Google TPU remaining the mainstream options [11]. - The reliability of GB200 NVL72 is a significant concern, with early operators facing numerous XID 149 errors, which complicates diagnostics and maintenance [34][36]. - Software optimizations, particularly in the CUDA stack, are expected to enhance GB200 NVL72's performance significantly, but reliability remains a bottleneck [37]. Group 4: Future Outlook - By July 2025, GB200 NVL72's performance/TCO is projected to reach 1.5 times that of H100, with further improvements expected to make it a more favorable option [30][32]. - The GB200 NVL72's architecture allows for faster operations in certain scenarios, such as MoE (Mixture of Experts) models, which could enhance its competitive edge in the market [33].

Core Insights - Morgan Stanley's report compares the operational costs and profit margins of various AI solutions in inference workloads, highlighting that most multi-chip AI inference "factories" have profit margins exceeding 50%, with NVIDIA leading the pack [1][3]. Profit Margins - Among selected 100 MW AI "factories," NVIDIA's GB200 NVL72 "Blackwell" GPU platform achieved the highest profit margin of 77.6%, translating to an estimated profit of approximately $3.5 billion [3]. - Google's self-developed TPU v6e pod ranked second with a profit margin of 74.9%, while AWS's Trn2 UltraServer and Huawei's Ascend CloudMatrix 384 platform reported profit margins of 62.5% and 47.9%, respectively [3]. Performance of AMD - AMD's performance in AI inference is notably poor, with its latest MI355X platform showing a profit margin of -28.2%, and the older MI300X platform at a significantly lower -64.0% [4]. Revenue Generation - NVIDIA's GB200 NVL72 chip generates $7.5 per hour, while the HGX H200 chip produces $3.7 per hour. Huawei's Ascend CloudMatrix 384 platform generates $1.9 per hour, and AMD's MI355X platform only generates $1.7 per hour [4]. - Most other chips generate revenue between $0.5 and $2.0 per hour [4].

Summary of Key Points from the Conference Call Industry Overview - The conference focused on the AI supply chain, particularly developments in AI chip technology and infrastructure at the Taiwan Open Compute Project (OCP) seminar held on August 7, 2025 [1][2][9]. Core Insights - **AI Chip Technology**: AI chip designers are advancing in scale-up technology, with UALink and Ethernet being key competitors. Broadcom highlighted Ethernet's flexibility and low latency of 250ns, while AMD emphasized UALink's latency specifications for AI workload performance [2][10]. - **Profitability of AI Factories**: Analysis indicates that a 100MW AI factory can generate profits at a rate of US$0.2 per million tokens, potentially yielding annual profits of approximately US$893 million and revenues of about US$1.45 billion [3][43]. - **Market Shift**: The AI market is transitioning towards inference-dominated applications, which are expected to constitute 85% of future market demand [3]. Company-Specific Developments - **NVIDIA's Rubin Chip**: The Rubin chip is on schedule, with the first silicon expected from TSMC in October 2025. Engineering samples are anticipated in Q4 2025, with mass production slated for Q2 2026 [4][43]. - **AI Semi Stock Recommendations**: Morgan Stanley maintains an "Overweight" (OW) rating on several semiconductor companies, including NVIDIA, Broadcom, TSMC, and Samsung, indicating a positive outlook for these stocks [5][52]. Financial Metrics and Analysis - **Total Cost of Ownership (TCO)**: The TCO for a 100MW AI inference facility is estimated to range from US$330 million to US$807 million annually, with upfront hardware investments between US$367 million and US$2.273 billion [31][45]. - **Revenue Generation**: The analysis suggests that NVIDIA's GB200 NVL72 pod leads in performance and profitability among AI processors, with a significant advantage in computing power and memory capability [43][47]. Additional Insights - **Electricity Supply Constraints**: The electricity supply is a critical factor for AI data centers, with a 100MW capacity allowing for approximately 750 server racks [18]. - **Growing Demand for AI Inference**: Major cloud service providers (CSPs) are experiencing rapid growth in AI inference demand, with Google processing over 980 trillion tokens in July 2025, a significant increase from previous months [68]. Conclusion - The AI semiconductor industry is poised for growth, driven by advancements in chip technology and increasing demand for AI applications. Companies like NVIDIA and Broadcom are well-positioned to capitalize on these trends, with robust profitability metrics and strategic developments in their product offerings [43][52].

Core Viewpoint - Huawei demonstrates strong technological capabilities in the semiconductor industry, particularly with its Ascend series chips and the recent launch of CM384, positioning itself as a leader in domestic AI chips [2][3]. Group 1: Financial Performance - In 2024, Huawei achieved a total revenue of RMB 862.072 billion, representing a year-on-year growth of 22.4% [5]. - The smart automotive solutions segment saw a remarkable revenue increase of 474.4%, while terminal business and digital energy businesses grew by 38.3% and 24.4%, respectively [5]. - Revenue from the Chinese market reached RMB 615.264 billion, driven by digitalization, intelligence, and low-carbon transformation [5]. Group 2: Huawei Cloud - The overall public cloud market in China is projected to reach USD 24.11 billion in the second half of 2024, with IaaS accounting for USD 13.21 billion, representing a year-on-year growth of 14.4% [6]. - Huawei Cloud holds a 13.2% market share in the Chinese IaaS market, making it the second-largest cloud provider after Alibaba Cloud [6]. - Huawei Cloud's revenue growth rate reached 24.4%, the highest among major cloud vendors in China [6]. Group 3: Ascend Chips - The CloudMatrix 384 super node integrates 384 Ascend 910 chips, achieving a cluster performance of 300 PFLOPS, which is 1.7 times that of Nvidia's GB200 NVL72 [10]. - The single-chip performance of Huawei's Ascend 910C is approximately 780 TFLOPS, which is one-third of Nvidia's GB200 [10][11]. - The Ascend computing system encompasses a comprehensive ecosystem from hardware to software, aiming to meet various AI computing needs [15][20]. Group 4: HarmonyOS - HarmonyOS features a self-developed microkernel, AI-native capabilities, distributed collaboration, and privacy protection, distinguishing it from Android and iOS [12]. - The microkernel architecture enhances performance and fluidity, while the distributed soft bus technology allows seamless connectivity among devices [12][13]. Group 5: Kirin Chips - The Kirin 9020 chip has reached high-end processor standards, comparable to a downclocked Snapdragon 8 Gen 2 [23]. - The Kirin X90 chip, based on the ARMv9 instruction set, features a 16-core design with a frequency exceeding 4.2GHz, achieving a 40% improvement in energy efficiency [25][26]. Group 6: Kunpeng Chips - Kunpeng processors are designed for servers and data centers, focusing on high performance, low power consumption, and scalability [27]. - The Kunpeng ecosystem strategy emphasizes hardware openness, software open-source, enabling partners, and talent development [29].

Summary of Huawei's CloudMatrix 8 Conference Call Company and Industry - **Company**: Huawei - **Industry**: Semiconductor and AI Computing Key Points and Arguments Product Overview - Huawei introduced the **CloudMatrix 8**, a powerful domestic solution in China built using the **Ascend 10C** chip, competing directly with Nvidia's **GB200 NVL72** [3][4] - The CloudMatrix 8 architecture is noted for its engineering advantages at the system level, not just at the chip level, with innovations across accelerator, networking, optics, and software layers [4] Performance Metrics - The CloudMatrix 8 can deliver **300 PFLOPS** of dense BF16 compute, nearly double that of the **GB200 NVL72** [10] - Key specifications comparison: - **BF16 dense PFLOPS**: CloudMatrix 300 vs. GB200 180 - **HBM capacity**: CloudMatrix 49.2 TB vs. GB200 13.8 TB - **HBM bandwidth**: CloudMatrix 1.229 TB/s vs. GB200 576 TB/s - **All-in System Power**: CloudMatrix 559,378 W vs. GB200 145,000 W [10][53] Power Consumption and Efficiency - The CloudMatrix 8 consumes significantly more power, drawing approximately **500 kW**, which is over **3.9 times** that of the GB200 NVL72 [51] - Despite higher power consumption, Huawei's system is designed to leverage China's abundant energy resources, allowing for scaling without power constraints [13][54] Supply Chain and Production Challenges - Huawei's Ascend chips are primarily produced by TSMC, with significant reliance on foreign production for components like HBM and wafers [16][19] - The company has reportedly circumvented sanctions to acquire necessary components, including **$500 million** worth of 7nm wafers [17] - Domestic production capabilities are improving, with SMC ramping up capacity, but foreign reliance remains a critical issue [24][27] Strategic Implications - The advancements in Huawei's technology are seen as a response to U.S. export controls, highlighting the importance of AI competitiveness as a national security concern [9] - The CloudMatrix 8's design reflects a strategic focus on scaling up capabilities, leveraging domestic strengths in networking and infrastructure software [11][15] Market Positioning - Huawei's CloudMatrix 8 is positioned as a competitive alternative to Nvidia's offerings, with a focus on system-level performance rather than just chip performance [5][6] - The architecture's design allows for significant scaling, which is crucial for meeting the demands of AI workloads [28][30] Conclusion - Huawei's CloudMatrix 8 represents a significant advancement in China's AI computing capabilities, with a focus on system-level innovations and leveraging domestic resources, despite challenges in supply chain and power efficiency [54]

Group 1 - By 2035, approximately 32% of global semiconductor production may be affected by climate change-related copper supply disruptions, which is four times the current level [1] - Chile, the largest copper producer, is facing water shortages that are slowing down copper production, with most of the 17 countries supplying copper for the chip industry expected to face drought risks by 2035 [1][3] - The low recycling rate of copper is causing urgency among chip manufacturers, highlighting the critical role of copper in semiconductor manufacturing [1] Group 2 - Copper is difficult to replace in the semiconductor industry, primarily used for interconnect lines due to its superior conductivity and low resistance compared to aluminum [1][2] - The global copper price has fluctuated significantly over the years, with a structural bull market observed from 2021 to 2022 driven by demand from electric vehicles and renewable energy [3] - The transition from optical to copper cables by companies like NVIDIA is expected to significantly increase copper demand, particularly in AI data centers [4][5] Group 3 - The global power demand for data centers is projected to grow at a compound annual growth rate of 15%, leading to an estimated additional demand of 2.6 million tons of copper by 2030 [6] - A copper supply gap of 4 million tons is anticipated by 2030 due to the growth of electric vehicles and renewable energy, coupled with limited new copper mine supplies [6][8] - The increasing use of copper in electric vehicles is notable, with estimates indicating that each electric vehicle requires between 80 to 83 kilograms of copper [8] Group 4 - The copper interconnect process is essential in integrated circuit manufacturing, utilizing techniques like the Damascene process to achieve embedded copper filling [7] - Copper serves as a critical material in various layers of chip interconnects, providing low resistance and excellent performance in modern semiconductor applications [7] - The demand for copper is further driven by the rapid growth of AI technologies, with the AI market expected to grow significantly, increasing the need for copper in computing infrastructure [9]