昨晚，国产AI芯片巨头寒武纪交出了一份"史上最强"财报——上半年营收暴增43倍，达到28.8亿元，利润也突破10亿元。要知道，去年此时，寒武纪还处于 亏损5.3亿的困境。 从巨亏到巨额盈利，寒武纪仅用了一年时间。这样的成绩即便放在整个A股市场，也实属罕见。 利好消息刺激下，今天下午一开盘，寒武纪股价便直冲1464元的历史高点，超越茅台，成为新一代"股王"。难道"酱香型科技"的戏言要在寒武纪身上成真？ 随着公司股价飙升，寒武纪的股东们也收获颇丰。 第一大股东、公司创始人之一的陈天石，身家暴涨至2000亿元，很可能已成为江西首富。 第七大股东、号称"超级散户"的章建平，凭借过人的眼光与胆识，于去年密集买入寒武纪股票，累计达600万股。以建仓成本约30亿元计算，这轮暴涨至少 为他带来50亿元的收益。 | 前十名无限售条件股东持股情况(不含通过转融通出借股份) | | | | --- | --- | --- | | 持有无限售条 | | 股份种类 | | 股东名称 | 件流通股的数 市 | 种类 | | 陈大石 | 119,530,650 | 人民币普通股 | | 北京中科算源资产管理有限公司 | 65,669,72 ...

西部某市干部变动，英伟达B200热度大减 西部某省此前大力推进智算业务，高峰时期规划近200个算力项目。近期，该省某市的核心干部变动，智 算业务负责人随之调整，导致已建成投产的算力项目在竣工验收环节遇阻——新任负责人均未批准，大批 项目搁置，直接引发市场行情降温。 对于此前的算力项目，虽仍有批准可能，但验收标准已提高。业内人士分析，在近200个项目中，最终 能"善终"的或仅5-10个。作为算力设备采购的绝对主力，当地项目受阻直接导致B200等算力设备热度大 幅下降。 H100单次维修费用数万元，英伟达GPU维保市场巨大 智算市场火热，国内也已经有上百万张英伟达高性能GPU，海量GPU的后市场业务十分可观。由于数据中 心GPU长时间高温工作，显存故障的概率很高，而维修的成本又很高。以维修一张H100为例，硬件维 修，一次维修的费用高达2-3万元，几乎是售价的10%，而固件的更新，价格更是高达10万甚至20万。 有客户资源、产业链资源的人已经准备做英伟达GPU的维保生意，不过英伟达可能也会自己下场做后市 场。更多英伟达GPU维保市场的交流 添加微信 BENSONEIT 。 浪潮、华勤技术调研服务器回收及维修市场， ...

Core Viewpoint - The introduction of UE8M0 FP8 by Deepseek for the upcoming domestic chips signifies a strategic move to enhance compatibility and efficiency in the Chinese AI ecosystem, addressing the unique requirements of domestic hardware [5][10][12]. Group 1: UE8M0 and FP8 Concept - FP8 is an 8-bit floating-point format that significantly reduces memory usage by 75% compared to 32-bit formats, enhancing computational speed and efficiency for large model training and inference [7][13]. - UE8M0 is a specific encoding format for FP8 tensor data, designed to optimize compatibility with domestic chips, differing from Nvidia's E4M3 and E5M2 formats which focus on precision and dynamic range [9][10]. - The Open Compute Project (OCP) introduced UE8M0 as part of its MXFP8 formats, aiming to standardize FP8 usage across various hardware platforms [8]. Group 2: Strategic Importance of UE8M0 - The development of UE8M0 is crucial for ensuring that domestic chips can effectively utilize FP8 without relying on foreign standards, thus reducing dependency on Nvidia's technology [12]. - Deepseek's integration of UE8M0 into its model development process aims to ensure that models can run stably on upcoming domestic chips, facilitating a smoother transition from development to deployment [11][12]. - The focus of UE8M0 is not to outperform foreign FP8 standards but to provide a viable solution that allows domestic chips to leverage FP8 efficiency [14]. Group 3: Performance and Limitations - UE8M0 can save approximately 75% in memory usage compared to FP32, allowing for larger models or increased request handling during inference [13]. - The inference throughput using UE8M0 can be about twice that of BF16, making it particularly beneficial for large-scale AI applications [13]. - However, UE8M0 is not a one-size-fits-all solution; certain calculations still require higher precision formats like BF16 or FP16, and effective calibration is necessary to avoid errors in extreme value scenarios [15].

以下文章来源于未尽研究 ，作者未尽研究 未尽研究 . AI，新能源，合成生物，地缘X 本文来自微信公众号： 未尽研究 （ID：Weijin_Research） ，作者：未尽研究，题图来自：AI生成 从V3到V3.1，DeepSeek正在探索出一条"算力自由"之路。 从魔改PTX到使用UE8M0 FP8 Scale的参数精度，DeepSeek先榨取英伟达GPU算力，再适配国产芯 片，可能会在软硬件协同方面带来新的突破， 进一步提高训练效率，最多可以减少75%的内存使用 ，从而在实际应用中减少对进口先进GPU芯片的依赖。 DeepSeek正在与下一代国产GPU芯片厂商一起，向算力自主又迈进一步。正是这样一种令人激动的 前景，激活了科技色彩愈发浓厚的中国资本市场。 DeepSeek发布了V3.1，而不是广受期待的V4或者R2，连R1也消失了。 DeepSeek变成了一个混合推 理架构，即一个模型同时支持思考模式和非思考模式。 这是一个趋势，在V3.1发布一周之前，GPT- 5发布了，这是一个"统一的系统"，包括一个对话模型，一个思考模型，和一个实时路由，用来决定 如何结合对话与思考。 这次升级提高了DeepSeek ...

Group 1 - The core argument of the articles highlights that the disparity in electricity supply is a critical factor determining the future competitiveness of AI technology, with the U.S. facing significant challenges due to its weak power grid [1] - AI models like GPT-3 consume substantial amounts of electricity, with a single training session equivalent to the annual electricity usage of 120 American households, and daily operations consuming 500,000 kWh, enough to power 20,000 households for a day [4] - High-end AI chips, such as NVIDIA's H100, have a significant energy footprint, with each chip consuming approximately the annual electricity of three households, and projected sales of 4-5 million units in 2024 could lead to an annual consumption equivalent to that of 12 million households [4] Group 2 - The U.S. power grid operates with only 15% reserve capacity, leading to shortages and outages, particularly in states like Texas and California, which hampers the ability of AI companies to secure sufficient energy for their data centers [5] - In contrast, China excels in energy production and transmission, with diverse sources including wind, hydro, nuclear, and solar power, and is currently constructing 42 nuclear power plants, with the Yalong River Hydropower Station expected to generate 300 billion kWh annually, three times that of the Three Gorges Dam [7] - China's unique ultra-high voltage transmission technology allows for significantly lower industrial electricity prices, with rates as low as 0.3 yuan per kWh compared to California's 1.2 yuan, providing a substantial competitive advantage in the AI technology race [7]

Group 1 - The core viewpoint of the article is that investors who are heavily invested in both US and Chinese assets have seen greater returns this year compared to those focused solely on US stocks, as Chinese assets have significantly outperformed US stocks [1] - The S&P 500 index has only increased by 9.6% this year, while the US dollar index has depreciated by 9.8%, indicating that gains in US stocks may not offset currency losses for global asset allocators [1] - There is a notable trend of South Korean retail investors increasing their investments in Hong Kong and A-shares, with a record investment of over $5.4 billion, surpassing Japan as their second-largest overseas investment destination [2] Group 2 - The article highlights the performance of specific Chinese stocks, such as Xiaomi and BYD, which have seen significant net inflows from South Korean investors, indicating a shift in investment focus towards undervalued Chinese assets [2] - The performance of liquid cooling stocks in the A-share market has been exceptional, with companies like Shenling Environment and Yingweike seeing increases of 60% and 83% respectively, while the US counterpart Vertiv only rose by 5% during the same period [3] - The article discusses the strong performance of Nvidia-related stocks in the A-share market, with companies like Industrial Fulian and Shenghong Technology experiencing substantial gains, suggesting that global asset allocation thinking can enhance investment returns [5] Group 3 - The China Banking Index has outperformed the CSI 300 index this year, with a year-to-date increase of 9.8% compared to 6.8% for the latter, indicating a strong performance of Chinese banking stocks [6] - The article mentions a specific fund, Anzheng Changying, which focuses on a diversified asset allocation strategy including A-share dividends, gold, and US stocks, achieving an annualized return of 12.5% since 2013 [10] - The A-share market has seen significant gains in dividend-paying stocks, particularly in the banking sector, which have outperformed major US tech stocks, highlighting the potential of A-shares as a viable investment option [11] Group 4 - The article emphasizes the importance of global asset allocation, suggesting that diversification across different markets can mitigate risks associated with market volatility, as seen during the recent downturns in the US market [13] - The investment strategy of combining Chinese assets with gold and US stocks is presented as a way to reduce overall portfolio volatility and enhance returns in the current uncertain economic environment [14] - The article concludes that a well-rounded investment approach that includes Chinese assets is essential for long-term success in the financial markets [14]

Core Insights - The article discusses the projected demand for CoWoS wafers, predicting that global demand will reach 1 million pieces by 2026, with TSMC dominating the capacity allocation and Nvidia securing 60% of the CoWoS capacity [1][2]. Group 1: TSMC and CoWoS Technology - TSMC is expected to produce approximately 510,000 CoWoS wafers for Nvidia's next-generation Rubin architecture AI chips, which will account for about 60% of the global market demand [1]. - The CoWoS technology is crucial for enhancing signal transmission efficiency and chip density while reducing power consumption and heat dissipation, making it the standard packaging method for high-end AI chips [3]. Group 2: US Manufacturing Expansion - TSMC plans to build an advanced packaging facility in Arizona, which will include CoWoS, SoIC, and CoW technologies, with 60% of the capacity dedicated to Nvidia [2]. - The establishment of the US facility aims to strengthen the local supply chain, mitigate geopolitical risks, and address the increasing demand for advanced packaging technologies driven by AI and high-performance computing chips [2]. Group 3: Investment and Future Projections - Since Trump's second term, TSMC has announced a total investment plan of up to $100 billion, covering wafer fabs, R&D centers, and advanced packaging facilities [2]. - The anticipated output from Nvidia's chips could reach 5.4 million units by 2026, with 2.4 million units coming from the Rubin platform [1].

Core Viewpoint - The article discusses the illegal export of Nvidia's advanced AI chips, particularly the B200 GPU, to China despite U.S. export restrictions, highlighting the emergence of a black market for these products [1][2][3]. Group 1: Nvidia's AI Chips and Black Market Activity - Following the tightening of U.S. export controls on AI chips to China, at least $1 billion worth of restricted Nvidia advanced AI processors have been shipped to mainland China [1]. - The B200 GPU has become the most popular chip in China's semiconductor black market, widely used by major U.S. companies like OpenAI, Google, and Meta for training AI systems [1][2]. - Despite the ban on selling advanced AI chips to China, it is legal for Chinese entities to receive and sell these chips as long as they pay the relevant border tariffs [1][2]. Group 2: Distribution and Sales Channels - A company named "Gate of the Era" has emerged as a major distributor of the B200, having sold nearly $400 million worth of these products [3]. - The B200 racks are sold at prices ranging from 3 million to 3.5 million RMB (approximately $489,000), which is lower than the initial price of over 4 million RMB [3]. - The sales of these chips are facilitated through various distributors in provinces like Guangdong, Zhejiang, and Anhui, with significant quantities being sold to data center providers [2][3]. Group 3: Market Dynamics and Future Outlook - The demand for Nvidia's B200 chips remains high due to their performance and relative ease of maintenance, despite U.S. export controls [11]. - Following the easing of the H20 export ban, the black market sales of B200 and other restricted Nvidia chips have reportedly decreased as companies weigh their options [13]. - Southeast Asian countries are becoming key transit points for Chinese companies to acquire restricted chips, with potential tightening of export controls being discussed by the U.S. government [13][15]. Group 4: Repair and Maintenance Services - There is a growing demand for repair services for Nvidia's high-end chips, with some companies in China specializing in the maintenance of H100 and A100 chips that have entered the market through special channels [17]. - The average monthly repair volume for these AI chips has reached 500 units, indicating a significant market need for maintenance services [17][18]. - The introduction of the H20 chip has seen limited market acceptance due to its high price and inability to meet the demands for training large language models [18].

Group 1 - The core viewpoint of the article highlights the competitive performance of a new generation of domestic GPUs, which reportedly achieves 60%-70% of the computational power of NVIDIA's H100, while outperforming it in certain efficiency metrics [1][4][5] - The domestic GPU shows superior computational efficiency in tasks such as image classification (ResNet-50 v1.5) and maintains parity with H100 in object recognition training (Mask R-CNN) [1][4] - The article discusses the shifting dynamics in the computing power market, particularly the decline in project scale for companies like Lianhua Zixing, which has transitioned from large-scale projects to smaller server rental services [7][8] Group 2 - A leading model startup has raised nearly 6 billion yuan in funding over six months, with over 3 billion yuan allocated for computing power procurement [10][13] - The article notes that many local investments in the model company come with conditions requiring funds to be invested in local industry development, particularly in computing power consumption [12] - The collaboration between a domestic chip company and an electronics company has resulted in a large-scale computing cluster, benefiting both parties financially through a dual procurement model [14][15][16][17] Group 3 - The second-hand server market is experiencing growth, with a leading platform reporting 2 billion yuan in revenue from server recovery services [18][19] - The article mentions challenges faced by domestic AI servers in achieving large-scale deployment due to high costs, despite government support for domestic computing power development [20][21] - The article highlights the fluctuating demand for computing power in western provinces, with significant disparities in consumption among different regions [23][24] Group 4 - A chip company has invested 4 billion yuan in building its own computing centers to ensure a closed-loop sales model for its chips [25][26] - A major manufacturer has purchased GB200 systems, which are reported to be equivalent to 30-40 units of H200, indicating a trend towards more efficient computing solutions [27] - A southern manufacturer has acquired 80,000 NVIDIA network cards, suggesting a substantial increase in GPU procurement [28]

Summary of Conference Call on Qorev and the Computing Power Leasing Industry Company Overview - **Company Name**: Qorev - **Founded**: 2017, originally as a cryptocurrency mining company named Atlantic - **Headquarters**: United States - **Business Focus**: Transitioned from cryptocurrency mining to AI cloud and infrastructure services since 2019 [1][2] Industry Insights - **Industry**: Computing Power Leasing - **Market Dynamics**: The computing power leasing industry is experiencing rapid growth, driven by increasing demand for AI applications and infrastructure [20][21] - **Competitors**: Over 100 new cloud computing service providers have emerged in the last six months, indicating a highly competitive landscape [15] Core Business Model - **Service Offerings**: - **Technical Infrastructure and Services**: Bare-metal GPU leasing, allowing direct access to H100 and A100 chips [2][3] - **Management Software**: Tools for managing AI workloads [3] - **Application Services**: Includes services like SUNK and Tensor Racer for enhanced efficiency [3] - **Revenue Model**: - **Contractual Revenue**: 96% of revenue comes from committed contracts, primarily with large AI labs and enterprises like Microsoft [4][10] - **On-Demand Pricing**: A retail-like model for clients needing additional computing power [5] Financial Performance - **Revenue Growth**: - 2024 revenue projected at $19.15 billion, a sevenfold increase year-over-year [10] - Q1 2025 revenue reached $9.82 billion, a fourfold increase [10] - **Order Backlog**: Remaining orders valued at $150 billion, a 53% year-over-year increase [10] - **Profitability**: - 2023 loss of $863 million, reduced to $315 million in Q1 2025 due to IPO-related costs [11] - Gross margin improved to 73.3% in Q1 2025, up from 74% in the previous year [11][12] Competitive Advantages - **GPU Utilization**: Qorev's floating-point utilization exceeds industry averages by over 20% [7][8] - **Strategic Partnerships**: Strong relationships with NVIDIA ensure priority access to cutting-edge chips [6] - **Power Supply Agreements**: Secured agreements for 500 MW capacity, with 360 MW currently available [9] Future Growth Strategies - **Expansion Plans**: Focus on scaling existing contracts and exploring new industries such as banking and pharmaceuticals [13][14] - **International Growth**: Plans to expand into North America, Europe, and Asia-Pacific [13][14] - **Operational Focus**: Aiming to optimize power supply and maintain low leverage through strategic financing [14] Market Trends - **Pricing Dynamics**: H100 rental prices are expected to decline as production ramps up, with NVL72 offering lower costs [16][17] - **Contractual Preferences**: Emphasis on securing long-term contracts (3-5 years) to ensure resource availability and profitability [19] Conclusion - The computing power leasing industry is poised for significant growth, driven by AI demand and evolving market dynamics. Qorev's strategic positioning, robust financial performance, and competitive advantages position it well for future success in this rapidly expanding sector [20][21]