Core Viewpoint - The article discusses the recent decision by Lu Chen Technology to suspend its DeepSeek API service, primarily due to cost considerations, despite DeepSeek's high theoretical profit margin of 545% [1][2][3]. Group 1: Cost Considerations - Lu Chen Technology's decision to halt DeepSeek API access is largely attributed to the high costs associated with providing stable service, which smaller MaaS providers struggle to manage compared to larger cloud companies [6][9]. - The theoretical profit margin of 545% reported by DeepSeek is based on a scenario where user demand is maximized, which is not typical for standard MaaS products that require significantly more resources to maintain stable output [3][4]. - The cost of providing DeepSeek services is exacerbated by the need for redundant computing resources to handle unpredictable user demand, leading to higher operational costs for smaller providers [9][10]. Group 2: Industry Impact - The suspension of DeepSeek API services by Lu Chen Technology reflects broader challenges faced by smaller MaaS companies in the wake of DeepSeek's competitive pricing and open-source initiatives, which threaten their business models [10]. - As DeepSeek continues to open-source its technology, many third-party MaaS providers are finding it increasingly difficult to maintain a competitive edge, leading to a potential disruption in the industry [10]. - The article highlights that numerous companies across various sectors, including technology, finance, and government, have integrated DeepSeek, indicating its widespread influence and the potential risks for smaller players in the market [8].

以下文章来源于飞跑的鹿 ，作者RunningLu 飞跑的鹿 . 公司研究 | 产业挖掘 | 价值投资 | 长期主义每日分享上市公司及产业链 | 市场调味剂 01 产业链全景图 03 什么是DeepSeek DeepSeek 是一家人工智能技术公司，中文即"深度求索"。 它家的模型在自然语言处理和代码 生成等领域相当能打，像DeepSeek-v3在多个数学基准测试和代码能力测试中，就超越了众多 竞品。 真正让DeepSeek火出圈的是2024年12月26日，这家公司宣布上线并同步开源的 DeepSeek-V3 模型。它以 1/11的算力、仅2000个GPU芯片 训练出性能 超越GPT-4o的大模型 。其 总训练成 本只有557.6万美元 ，而 GPT-4o的约为1亿美元 ，使用 25000个GPU芯片 。 双方的成本至少 是10倍的差距。 在性能上，DeepSeek-V3在数学、代码能力和中文知识问答方面还超过了ChatGPT-4o， 以下即为两者的模型参数对比： 04 为什么被称为AI界的拼多多 据公司介绍，在数学、代码、自然语言推理等任务上， DeepSeek-R1 性能比肩已经OpenAlo1 正式版，这 ...

This article has been reviewed according to Science X's editorial process and policies . Editors have highlighted the following attributes while ensuring the content's credibility: Singapore media have linked a local fraud case to the alleged movement of Nvidia chips to be used by Chinese AI firm DeepSeek. Servers that may contain AI-powering Nvidia chips shipped from the United States to Singapore ended up in Malaysia, but their actual final destination remains a mystery, the city-state's interior ministe ...

证券研究报告 计算机 DeepSeek 再开源，关注 AI 应用变化 华泰研究 2025 年 3 月 03 日│中国内地 动态点评 2 月 24 日起 DeepSeek 连续 6 天开源，在之前放出的模型参数、技术报告 基础上，再次发布了 Infra 层的核心代码，涉及 MLA、通信-计算、矩阵乘法 运算、专家负载、文件存取等模块优化，旨在提高模型本身和硬件的效率， 且国产 GPU 适配进展顺利。据 DeepSeek 数据，若将 Web、APP 和 API 的所有用户请求均以 R1 定价计费，则每日总收入将为 562,027 美元，成本 利润率为 545%。若考虑 V3 定价、夜间打折等因素，付费 token 占比 50% 情况下我们测算成本利润率有望达到 108%，优化效果明显。我们认为，模 型层的持续优化，有望持续降低应用层成本、提高应用表现。建议关注 2B 和 2C 应用中拥有用户、数据和场景优势的公司。 DeepSeek 在原先开源的基础上，再次开源 Infra 核心代码 此前 DeepSeek 在核心的 V3/R1 模型上，已经开源了模型权重，使得全球 用户均可自行下载、部署和推理，并且配备了较为详 ...

Core Insights - The competition among leading AI labs in reasoning models has intensified, with no clear SOTA leader emerging yet [1][3][10] - The release of Claude 3.7 Sonnet's hybrid reasoning model is expected to set a new standard for future AI models [13][16][17] Group 1: Reasoning Models Overview - OpenAI's o3-mini excels in mathematical reasoning but lacks in creative content generation compared to Grok and DeepSeek models [3][4] - Grok 3 Think has rapidly caught up to o3-mini, demonstrating strong reasoning capabilities and faster inference speed [4][5] - Claude 3.7 Sonnet leads in solving real-world coding problems, significantly outperforming others in engineering code tasks [5][19] - Gemini 2.0 Flash is underappreciated, showing strong multimodal understanding but lacking standout features [6][7] - DeepSeek R1 has made innovations despite limited resources, but currently lags behind top labs [7][8] Group 2: Base Model Competition - Grok 3 is perceived to potentially surpass GPT-4.5 in base model capabilities, with user feedback indicating a preference for Grok [10][11] - The importance of high-quality base models for reinforcement learning in reasoning models is emphasized, countering doubts about diminishing returns [12] Group 3: Hybrid Reasoning Model - Claude 3.7 Sonnet's hybrid reasoning model combines LLM and reasoning capabilities, likely influencing future AI model releases [13][16] - Users can toggle between fast and slow thinking modes, enhancing the model's adaptability [14][15] Group 4: AI Coding Developments - Claude 3.7 Sonnet has significantly improved coding capabilities, allowing for longer and more reliable code outputs [20][21] - Claude Code is positioned as a foundational tool for AI coding products, focusing on backend capabilities rather than direct user competition [22][23] Group 5: Action Scaling and Learning - The action scaling capability in Claude 3.7 allows for iterative problem-solving, crucial for effective AI agent deployment [25][26] - Continuous learning and dynamic fine-tuning are identified as key challenges for developing personalized AI agents [28] Group 6: Product Form and User Experience - OpenAI's Deep Research is recognized as the first PMF product in the RL scaling paradigm, offering superior user experience and task completion accuracy [29][30] - The ability to control research depth and breadth through configurable parameters is highlighted as a significant advancement [31][32]

Investment Rating - The report assigns a "Buy" rating for stocks, indicating an expected relative return of over 20% within six months [39] Core Insights - The media sector saw a decline in the Shenwan Media Index by 1.8%, ranking 30th, while the Shanghai Composite Index rose by 0.97% and the ChiNext Index increased by 2.99% during the week of February 17 to February 21, 2025 [8][9] - The gaming sector experienced a 3.57% increase, while the film industry saw a significant drop of 10.86% [9][10] - The total box office for the film market reached 155.42 billion yuan in February 2025, marking a 40% year-on-year increase [23][25] Summary by Sections Market Review - The Shenwan Media Index decreased by 1.8% during the week, while the gaming sector rose by 3.57% and the film sector fell by 10.86% [8][9] - Monthly performance showed the gaming sector up by 22.86%, while the film sector increased by 29.3% [9] AI Developments - Tencent's AI assistant "Tencent Yuanbao" has rapidly iterated and upgraded its features, achieving the second position in the Apple App Store free app download rankings in China [17] - The xAI company, founded by Elon Musk, released the Grok 3 AI model, achieving a score of 1400 in the competitive arena, showcasing its capabilities in various fields [16] Film Industry Performance - "Nezha: The Devil's Child" has surpassed 13.4 billion yuan in cumulative box office, becoming the highest-grossing animated film globally [4] - The film "Detective Chinatown 1900" has also performed well, crossing 3.3 billion yuan in box office [4] Gaming Sector Insights - The National Press and Publication Administration issued 110 domestic game licenses and 3 import licenses in February [30] - Tencent's "Peace Elite" announced the integration of DeepSeek for AI-driven digital representation [30][31] Company Recommendations - Companies to watch in the AI+ industry include Tencent, Century Huatong, Zhejiang Wenlian, and others in various sectors such as education, marketing, and e-commerce [3][4]

Core Insights - DeepSeek's five core technologies (FlashMLA, DeepEP, DeepGEMM, DualPipe & EPLB, 3FS file system) were showcased during a five-day "Open Source Week," achieving significant global attention [1] - JD Cloud announced full-stack adaptation of these technologies, resulting in a 50% performance improvement in inference scenarios [1][2] Group 1: Technology Enhancements - Flash MLA optimizes GPU memory and computational resources, addressing resource wastage in traditional methods for processing variable-length sequences [1] - The vGPU AI computing platform supports Flash MLA's FP8 format, reducing single Token's KV Cache memory usage by 57 times compared to Multi-head Attention, ensuring high throughput and low latency under high concurrency [1] Group 2: Communication and Performance - JD Cloud's vGPU AI computing platform fully supports distributed inference using the DeepEP communication library, significantly enhancing inference throughput [2] - By integrating DeepEP, JD Cloud utilizes NVLink for intra-machine communication and NVSHMEM for inter-machine communication, improving GPU resource utilization and reducing performance bottlenecks [2] Group 3: Local Deployment and Adaptation - JD Cloud has assisted multiple local governments in deploying DeepSeek based on existing infrastructure, allowing local enterprises to access the service without resource investment [3] - The platform has achieved comprehensive domestic chip adaptation, ensuring self-control from foundational computing to large model applications, including over ten domestic AI computing solutions [2]

Core Insights - DeepSeek has revealed its operational costs and theoretical revenue during its open-source week, indicating a daily total cost of $87,072 and a potential revenue of $562,027, leading to a theoretical profit margin of 545% [4][11][12] - However, actual revenue is significantly lower due to lower pricing for DeepSeek-V3 compared to R1, free access to web and app services, and discounts during off-peak hours [12] Cost and Revenue Analysis - Daily total cost is calculated at $87,072, assuming a rental cost of $2 per hour for each H800 GPU [5][11] - The theoretical daily revenue, if all tokens were charged at DeepSeek-R1 rates, would be $562,027, resulting in a theoretical net profit of $474,955 [11][12] - Actual revenue is impacted by various factors, including lower pricing for DeepSeek-V3 and limited monetization of services [12] System Architecture and Performance - DeepSeek employs a cross-node expert parallelism (EP) strategy to enhance throughput and reduce latency, addressing the complexity introduced by EP [2][15] - The system achieved a peak node utilization of 278 and an average utilization of 226.75 nodes during the 24-hour period analyzed [5] - Total input tokens processed were 608 billion, with 56.3% hitting the KVCache [7] Technical Specifications - Each H800 node provides an average input throughput of approximately 73.7k tokens per second during the prefill phase and 14.8k tokens per second during decoding [9] - The system utilizes a combination of FP8 and BF16 formats for matrix calculations and dispatch transmissions to ensure service quality [5] Load Balancing Strategies - DeepSeek implements load balancing across GPUs to prevent performance bottlenecks, ensuring equitable distribution of computational and communication loads [22][23] - The optimization goals include balancing core-attention computation loads and dispatch sending volumes across different GPUs [23][24] - The expert parallel load balancer aims to minimize the maximum dispatch reception load across all GPUs [26]

Core Insights - China's technology sector is experiencing significant advancements, with major companies like Tencent, Alibaba, Baidu, JD.com, and PDD Holdings leading the charge in AI and emerging technologies [2][8] Group 1: Technological Advancements - DeepSeek, an AI startup, is at the forefront of China's tech revolution, recently launching its R2 model, which enhances coding capabilities and multilingual reasoning [3] - China's semiconductor industry holds over 25% of the global market share in semiconductor packaging and more than 50% in advanced packaging, leveraging technologies like 2.5D/3D stacking [4] - Robotics innovations were showcased at CES 2025, with Unitree Robotics presenting humanoid and quadrupedal robots, highlighting China's rapid progress in this field [5] - Electric vehicle technology is advancing, with companies like Zeekr and Great Wall Motor displaying innovative models, supported by suppliers like Hesai, whose lidar units have dropped in price from $80,000 in 2017 to around $200 in 2025 [6] - Augmented reality is gaining traction, with companies like Xreal and Rokid presenting advanced AR glasses and eyewear, reflecting China's comprehensive approach to technological innovation [7] Group 2: Company-Specific Developments - Tencent has launched its Hunyuan Turbo S model, which delivers responses within a second, significantly outperforming competitors and matching capabilities of DeepSeek's models [10][11] - Alibaba is investing $53 billion in cloud and AI infrastructure over the next three years, positioning itself as a leader in AI with the upcoming release of its QwQ-Max-Preview model [14][15] - Baidu is focusing on autonomous driving through a partnership with CATL to develop competitive driverless vehicles and plans to launch its upgraded Ernie 4.5 AI model [16][17][18]

Investment Rating - The report indicates a strong demand for AI technology talent, highlighting a significant talent shortage in the industry, which suggests a positive investment outlook for companies focusing on AI technology [2][19]. Core Insights - The AI technology sector is experiencing a rapid increase in demand for highly educated professionals, with nearly 47% of positions requiring master's or doctoral degrees, significantly higher than the overall job market [3][9]. - The average annual salary for AI technology positions is notably high, with over 30% of roles offering salaries above 500,000, compared to less than 10% in the overall job market [4][11]. - The report identifies a talent gap of approximately 4 million in the AI sector as of 2023, emphasizing the high value and demand for AI professionals [5]. Talent Demand Analysis - AI technology roles are characterized by a high demand for algorithm engineers, who account for 67.17% of the talent demand, followed by image algorithms and machine vision [6][7]. - The demand for deep learning and machine learning professionals is increasing, reflecting the growing application of these technologies across various industries [8]. - The highest demand for AI talent is found in the internet industry, which accounts for 30.37% of the total demand, followed by electronics and semiconductor sectors [13][15]. Regional Distribution - The Yangtze River Delta region shows the highest demand for AI technology talent, with 40.11% of the total demand, while major cities like Beijing, Shanghai, and Shenzhen lead in individual demand [16]. Talent Characteristics - AI technology professionals are predominantly young, with 59.90% under the age of 30, and a high percentage (72.99%) holding advanced degrees from prestigious institutions [17][18]. - The talent shortage index for AI technology reached 3.24 in January 2025, indicating a significant supply-demand imbalance, particularly for roles in search algorithms and recommendation algorithms [19].