来关注周末要闻。 宏观•要闻 央行、证监会等五部门联合召开座谈会 为深入学习贯彻习近平总书记在民营企业座谈会上的重要讲话精神，落实党中央、国务院关于金融支持 民营企业发展的决策部署，2025年2月28日，中国人民银行、全国工商联、金融监管总局、中国证监 会、国家外汇局联合召开金融支持民营企业高质量发展座谈会。中国人民银行党委书记、行长潘功胜， 中央统战部副部长、全国工商联党组书记沈莹，金融监管总局党委委员、副局长周亮，中国证监会党委 委员、副主席李超出席会议并讲话。 会议要求，要实施好适度宽松的货币政策，发挥好结构性货币政策工具作用，强化监管引领，引导金融 机构"一视同仁"对待各类所有制企业，增加对民营和小微企业信贷投放。执行好金融支持民营经济25条 举措，健全民营中小企业增信制度，加快出台规范供应链金融业务政策文件。强化债券市场制度建设和 产品创新，持续发挥"第二支箭"的撬动引领作用。抓好"科创板八条""服务现代化产业体系十六条""并 购六条"等政策落实落地，支持民营企业通过资本市场发展壮大。金融机构要强化金融服务能力建设， 进一步畅通民营企业股、债、贷等多元化融资渠道，加大各类金融资源要素投入，将民营企业 ...

The. Good afternoon, investors. It's Friday again. The working day of the week is coming to an end. Welcome to our live broadcast. Today's live broadcast is military-themed. Mai Pei Qi's Little Bridge. Sorry, we are late by half a minute. We are very punctual. Without further ado, let's talk about Now. Since the Spring Festival until now, DeepSeek has been very popular. It has attracted the attention of the whole technology industry to AI and artificial intelligence. In this live broadcast, we will also int ...

如何看待过去一周 DeepSeek 开源项目的特点及影响 20250301 摘要 • DeepSeek 开源周发布数据准备、模型训练和模型推理三大类项目，针对 英伟达 Hopper GPU 架构优化，提升经济效益和用户体验，理论利润率可 达 545%，展示全栈优化能力。 • Flash MLA 针对 Hopper 架构 GPU 优化，减少可变长度序列服务场景的 计算开销，内存绑定达 3,000GB/s，计算绑定达 580KB look 速度，逼近 H800 GPU 上限。 • DBP 是首个开源专家通信并行库，优化 MOE 混合专家模型通信效率，提 升不同专家之间的信息传递速度，加速大规模计算。 • DateEngineer 专为 FP8 精度通用矩阵乘法设计，支持普通和专家混合模 型分组运算，在 H 系列 GPU 卡上实现极致的 FP8 运算速度。 • 双向流水线并行算法通过减少 V3 和 IE 训练时计算与通信重叠时间中的气 泡来提升效率，提高显卡利用率和训练速度，目前仅支持 NVIDIA H 系列 显卡。 • EPLB 通过冗余专家策略优化硬件利用率，在推理阶段减少资源浪费，提 高推理速度和效率，具有较 ...

先是宣布："尊敬的用户，潞晨云将在一周后停止提供DeepSeek API服务，请尽快用完您的余额。如果 没用完，我们全额退款。" 后又发布消息："感谢网友的热心提醒，Colossal-AI此前发布对DeepSeek-R1（671B）模型的LoRA微 调，在参数加载过程中因参数名称不匹配的Bug导致Loss异常，已在GitHub线上修复。" 这家公司突然宣布停止提供DeepSeek API服务的消息引发较大关注。 除此之外，潞晨科技创始人尤洋和硅基流动创始人袁进辉，在社交平台公开互怼的内容也引来了大量吃 瓜群众。 这是两家大模型服务公司。尤洋质疑硅基流动的商业模式和技术水平，袁进辉则回应并指责潞晨科技存 在抄袭行为。随后，潞晨科技宣布将暂停DeepSeek API服务。 事件发生的背景是：春节期间DeepSeek访问量暴增经常无法访问，硅基流动趁此机会发出了DeepSeek API并对自己的产品做了一大波宣传，因此发出了不少代金券，其网站访问量也随之大增。 周末，DeepSeek大消息不断。 潞晨科技宣布停止DeepSeek API服务 3月1日，潞晨科技官微发布了两则消息。 硅基流动PR稿里说自己有300万用 ...

证券时报网讯，今年伊始，来自中国的开源AI模型DeepSeek火爆全球。 这段时间，就有不少高校引入了"DeepSeek"，不仅实现了24小时答疑解惑、精准回应校园生活诉求，还 大幅提升了老师的办公科研效率，智慧教学与研究场景正在变为现实。 新学期开始，中国人民大学在自主研发的教学科研行政平台中接入了满血版"DeepSeek"，在学校算力的 支持下，老师们多了一个专属"智能助手"。 还有一些师生将课件、研究等"喂料"给DeepSeek，"训练"出专属的课程知识库、数据库。 ...

Summary of Key Points from the Expert Call on DeepSeek and Embodied AI Industry Overview - The discussion focuses on the **humanoid robot industry** and the role of **embodied AI** in enhancing the commercial value of humanoid robots [1][2]. Core Insights and Arguments 1. **Importance of Embodied AI**: Embodied AI is critical for humanoid robots, as it determines their intelligence and commercial value. The process involves four steps: sensing, decision-making, planning, and execution [2][3]. 2. **Sub-models of Humanoid Robots**: - **Strong Reasoning Model**: Enhances decision-making and planning, reducing model hallucinations [3]. - **Multi-modal Large Model**: Processes various types of information, allowing robots to perceive their environment [3]. - **World Model**: Provides a physical understanding of the world, improving human-robot interaction [4]. 3. **Challenges in Developing Embodied AI**: - **Data Collection Difficulties**: Obtaining physical world data is challenging, impacting model training and optimization. Simulated data is not yet viable for training due to discrepancies with real-world data [5]. - **Insufficient Reasoning Capabilities**: Current embodied AI models are in early development stages, lacking the reasoning capabilities needed for complex real-world interactions [6]. 4. **Technological Advancements from DeepSeek**: - **Chain-of-Thought Models**: DeepSeek's models improve decision-making and planning efficiency for humanoid robots [7][9]. - **Enhanced Understanding**: These models can surpass human experts in specific tasks, such as solving math and science problems [9]. - **Cost Efficiency**: DeepSeek's architecture reduces model parameters and training costs, accelerating the development of embodied AI [10][13]. 5. **Future Expectations**: - **Timeline for All-Purpose Humanoid Robots**: Full realization of all-purpose humanoid robots is projected to be 5-10 years away due to data collection and computational challenges [14]. - **Increased R&D Investment**: The breakthrough from DeepSeek is expected to boost R&D in specialized scenarios, enhancing technology advancements in embodied AI [14]. Value Chain Positioning 1. **Sensing and Data Collection**: Essential for humanoid robots to interact with their environment. Data collection factories are being established in major Chinese cities, benefiting sensor and dexterous hand companies [15]. 2. **Edge-side Computing Chips**: With the rise of edge-side embodied AI, there is a growing demand for chips that support this technology [17]. 3. **Motion Control Units**: As humanoid robots perform more complex tasks, the accuracy of motion control systems becomes increasingly critical [18]. Additional Important Insights - The emergence of strong inference models due to DeepSeek's advancements is expected to enhance market confidence in achieving embodied AI [14]. - The development of a positive ecosystem for embodied AI is being fostered by the use of DeepSeek datasets and methodologies by various edge-side models [12].

Flash | 23 Feb 2025 16:27:43 ET │ 13 pages Citi Research is a division of Citigroup Global Markets Inc. (the "Firm"), which does and seeks to do business with companies covered in its research reports. As a result, investors should be aware that the Firm may have a conflict of interest that could affect the objectivity of this report. Investors should consider this report as only a single factor in making their investment decision. Certain products (not inconsistent with the author's published research) are ...

24 February 2025 Equity Research Report China Technology Equities Servers: Riding on the wave of DeepSeek DeepSeek is driving Cloud capex in China. AI inferencing workloads have been surging, triggered by the growing popularity of DeepSeek R1 in recent weeks. We see that individual user demand has been growing, and the AI model has been increasingly deployed by enterprises and government entities. DeepSeek R1 model's innovations in algorithm reduce inferencing costs sharply and also lead to much less comput ...

Summary of the Investor Presentation on Greater China Tech Semiconductor Industry Overview - **Industry Focus**: Greater China Tech Semiconductor, specifically the AI semiconductor sector - **Key Drivers**: Positive outlook on cloud AI demand remains unchanged, with expectations of similar AI capital expenditure intensity from US Cloud Service Providers (CSPs) [1][2] Core Insights - **China AI Market**: A more constructive view on the China AI market, with MediaTek rated as Overweight (OW) [1][2] - **AI Semiconductor Demand**: The demand for AI semiconductors is expected to accelerate due to generative AI applications across various sectors [5] - **DeepSeek Impact**: The introduction of DeepSeek may trigger an edge AI replacement cycle, as it demonstrates cheaper inferencing capabilities [5] - **Supply Chain Dynamics**: The ramp-up of China’s foundry supply is becoming stronger and more sustainable, leading to a slower recovery in mature node foundry utilization [5] Financial Metrics and Valuation - **Valuation Comparisons**: - TSMC (Ticker: 2330.TW) has a current price of 1,080.0 TWD with a target price of 1,388.0 TWD, indicating a 29% upside [6] - MediaTek (Ticker: 2454.TW) has a current price of 1,500.0 TWD with a target price of 1,688.0 TWD, indicating a 13% upside [7] - **P/E Ratios**: - TSMC's P/E ratio is projected to decrease from 33.4 in 2023 to 17.7 in 2025, reflecting expected earnings growth [6] - MediaTek's P/E ratio is expected to improve from -35% in 2023 to 24% in 2025 [7] Market Trends - **AI vs. Non-AI Semiconductor Growth**: Non-AI semiconductor growth is projected at only 10% year-over-year in 2024, excluding NVIDIA's AI GPU revenue [9] - **Inventory Levels**: A decrease in semiconductor supply chain days of inventory is noted, which historically correlates with an increase in semiconductor stock prices [10][11] Long-term Demand Drivers - **Tech Diffusion**: The reacceleration of AI semiconductor demand is driven by the proliferation of generative AI across various industries [5] - **Tech Deflation**: Anticipated "price elasticity" is expected to stimulate demand for technology products [5] Additional Insights - **AI Computing Wafer Consumption**: Expected to reach up to $15.3 billion in 2025, with NVIDIA being the primary consumer [26] - **HBM Consumption**: Projected to reach up to 20.5 billion gigabits in 2025, with NVIDIA consuming the majority of the supply [29] This summary encapsulates the key points from the investor presentation, highlighting the positive outlook for the AI semiconductor sector in Greater China, the financial metrics of key players, and the broader market trends influencing the industry.

Summary of the Conference Call Transcript Company and Industry Involved - **Company**: Microsoft Corp. (MSFT) - **Industry**: Data Centre Equipment Core Points and Arguments 1. **Microsoft's Capex and Capacity Intentions**: Further clarification is needed from Microsoft regarding its capital expenditure (capex) and broader capacity intentions, as the current situation appears to be Microsoft-specific rather than indicative of a broader industry trend [2][3][8] 2. **Impact of Lease Cancellations**: Microsoft has cancelled leases for two data centres totaling a few hundred megawatts and opted not to convert statements of qualification into leases for up to 1 gigawatt of capacity, reallocating some projects from international to domestic [3][8] 3. **Microsoft's Market Share**: Microsoft accounts for approximately 4% of global data centre capacity and about 8% of planned future additions, indicating a significant role in the data centre market [4][9] 4. **Capex/Sales Ratio**: Microsoft has a high capex/sales ratio of around 23%, second only to Meta at 28%, suggesting aggressive investment in infrastructure [4][9] 5. **Current Market Reaction**: The current share price reaction of Schneider Electric, which has seen a 4% decline, is viewed as a potential buying opportunity rather than a sign of a fundamental shift in the market [2][8] 6. **Demand Trends**: Initial assessments indicate that there are no signs of demand slowdown in the data centre sector, with capacity utilization data reaching new all-time highs [8][9] 7. **Future Capacity Plans**: Microsoft is expected to maintain its ambitious data centre buildout plans, with projections indicating it will account for a significant share of future capacity additions [4][9] Other Important but Possibly Overlooked Content 1. **Competitive Landscape**: The competitive pressure from cloud-native leaders like Amazon and Google poses risks to Microsoft's growth, particularly in the context of its legacy businesses transitioning to cloud-centric models [24] 2. **Regulatory Risks**: Microsoft faces legal and regulatory risks due to its dominant position in the PC operating systems and productivity applications market, which could impact its investment strategies [24] 3. **Analyst Ratings**: Microsoft currently holds a "Buy" rating with a price target of $480, reflecting positive sentiment among analysts despite the recent lease cancellations [38][49] This summary encapsulates the key insights from the conference call, focusing on Microsoft's strategic positioning within the data centre equipment industry and the implications of its recent decisions.