以下文章来源于解码LP ，作者吴琼 解码LP . 投资界（PEdaily.cn）旗下，专注募资动态 宽容失败。 作者 I 吴琼 报道 I 投资界-解码LP 投资界-解码LP获悉，近日，海南省科学技术厅发布最新征集调查通知，针对《海南省科 技创新容错免责实施办法（试行）》（简称《实施办法》）征求意见。 其中对满足容错免责的情形予以界定，并创新性提出企业及青年科研人员的"首次免 责"。 眼下全国各地掀起容错探索潮，海南此举也指向同一个目标——鼓励创新，宽容失败。 海南尝试 首提青年科研人员"首次免责" 时间回到20 24年底，海南省发布《进一步强化企业科技创新主体地位改革若干措施》 （简称《若干措施》），针对企业在科技创新决策中的参与度不够、研发投入占比不 高、企业牵头承担科研项目比例较低、企业在"产学研"合作中作用发挥不够、企业创新 服务支撑力度不足等科技创新存在的突出问题，支持企业科技创新发展所需。 该《若干措施》主要提出了七个方面2 8条改革措施。在健全符合创新规律的监管制度方 面，针对性提出建立支持企业创新的容错免责制度、建立政府投资基金尽职免责制度等 创新举措，并将尽快推动出台具体配套措施。 这一次，针 ...

图片来源：阿里巴巴 阿里正在开源其 Wan2.1 系列中的四个模型，这是该公司基础 AI 模型的最新版本，能够根据文本和图像输入生成图像和视频。 开源与 OpenAI 等创建的专 有模型形成对比。 自 DeepSeek 在一月份震撼全球市场以来，开源 AI 技术便成为了焦点。阿里巴巴于2月26日周三表示，已将其视频生成人工智能模型免费开放使用，进一 步加剧了与 OpenAI 等竞争对手的竞争。 图片来源：阿里云 其 Wan2.1 系列模型将通过阿里云的 Model Scope 和 Hugging Face （一个庞大的 AI 模型库）提供，全球的学术界、研究人员和商业机构均可访问。 阿里巴巴在香港上市的股票收盘上涨近 5% 。 目前围绕 AI 模型是否会商品化的问题正展开激烈讨论。 中国企业尤其积极推动开源模型的发展，阿里巴巴和 DeepSeek 的模型如今跻身全球最受欢迎之列。阿里巴巴于 2023 年 8 月发布了其首个开源模型，而 Meta 在美国则以其 Llama 模型引领开源潮流。 阿里巴巴的股票今年表现强劲，其香港上市股票截至 2025 年已上涨 66%。 本文翻译自： CNBC https://w ...

Core Viewpoint - DeepSeek has made significant advancements in optimizing parallel computing strategies and has introduced new models that enhance performance and reduce costs in AI applications [2][3][5][7]. Group 1: Optimized Parallelism Strategies - DeepSeek announced the release of Optimized Parallelism Strategies aimed at improving computational efficiency, reducing resource waste, and maximizing system performance through effective task allocation and resource coordination [3][5]. - The strategies are designed for high-performance parallel execution in multi-core, distributed, or heterogeneous systems, balancing computation, communication, and storage overhead [5]. Group 2: New Model Releases - NVIDIA has open-sourced the first optimized DeepSeek-R1 model on the Blackwell architecture, achieving a 25-fold increase in inference speed and a 20-fold reduction in cost per token [3][7]. - The DeepSeek-R1 model's local deployment has garnered significant attention, with its inference throughput reaching 21,088 tokens per second, compared to 844 tokens per second for the H100, marking a substantial performance improvement [7]. Group 3: Cost Reduction Initiatives - DeepSeek announced a significant reduction in API call prices during nighttime hours, with DeepSeek-V3 prices cut to 50% and DeepSeek-R1 to as low as 25%, with reductions up to 75% [6]. Group 4: Additional Open Source Contributions - DeepSeek has continued its open-source initiatives by releasing FlashMLA, DeepEP, and DeepGEMM, which are optimized for NVIDIA GPUs and designed to support various AI model training and inference tasks [9].

Core Viewpoint - The emergence of DeepSeek challenges traditional beliefs about AI model development, demonstrating that a financial startup from China can innovate in AI, contrary to the notion that only large tech companies or research institutions can do so [1][4][5]. Group 1: AI Economics: Scaling Laws vs. Scale Effects - DeepSeek's success indicates a shift in understanding the barriers to AI model development, particularly reducing the constraints of computational power through algorithm optimization [8][9]. - Scaling laws suggest that increasing model parameters, training data, and computational resources leads to diminishing returns in AI performance, while scale effects highlight that larger scales can reduce unit costs and improve efficiency [10][11]. - The interplay between scaling laws and scale effects is crucial for understanding DeepSeek's breakthrough, as algorithmic advancements can enhance the marginal returns of computational investments [12][14]. Group 2: Latecomer Advantage vs. First-Mover Advantage - The distinction between scaling laws and scale effects provides insights into the competitive landscape of AI, where latecomers like China can potentially catch up due to higher marginal returns on resource investments [16][22]. - The AI development index shows that the U.S. and China dominate the global AI landscape, with both countries possessing significant scale advantages, albeit in different areas [18][22]. - The competition between the U.S. and China in AI is characterized by differing strengths, with the U.S. focusing on computational resources and China leveraging its talent pool and application scenarios [19][22]. Group 3: Open Source Promoting External Scale Economies - DeepSeek's open-source model reduces commercial barriers, facilitating broader adoption and innovation in AI applications, which can accelerate the "AI+" process [24][26]. - The open-source approach allows for greater external scale economies, benefiting a wider range of participants compared to closed-source models, which tend to concentrate profits among fewer entities [25][28]. - The potential market size for AI applications is estimated to be about twice that of the computational and model layers combined, indicating significant growth opportunities [27]. Group 4: Innovation Development: From Supply and Assets to Demand and Talent - The success of DeepSeek raises questions about the role of traditional research institutions in innovation, suggesting that market-driven demands may lead to more successful outcomes in technology development [30][31]. - The integration of technological and industrial innovation is essential for sustainable growth, emphasizing the need for a shift from a supply-side focus to a demand-side approach that values talent and market needs [32][33]. - The importance of talent incentives and a diverse innovation ecosystem is highlighted, as smaller firms may be more agile in pursuing disruptive innovations compared to larger corporations [34][36]. Group 5: From Fintech to Tech Finance - The relationship between finance and technology is re-evaluated, with the success of DeepSeek illustrating how financial firms can leverage technological advancements to enhance their competitive edge [36][39]. - The role of capital markets in fostering innovation ecosystems is emphasized, suggesting that a diverse range of participants is necessary for achieving external scale economies [38][39].

Nvidia CEO Jensen Huang said Wednesday (Feb. 26) that sales of the company’s most advanced chip architecture hit a record in the fourth quarter, and it is a harbinger of even more demand ahead because the artificial intelligence (AI) era has just begun.“AI is advancing at light speed,” Huang said in a conference call with Wall Street analysts to discuss earnings. “We’re just at the start of the age of AI.” The architecture for the Blackwell GPU, which was delayed for two months due to technical issues, book ...

Core Insights - Nvidia's CEO Jensen Huang emphasized that next-generation AI will require 100 times more computational power than previous models due to new reasoning approaches that involve step-by-step question answering [1] - Nvidia reported a significant revenue increase of 78% year-over-year, reaching $39.33 billion, with data center revenue, primarily from AI-focused GPUs, soaring 93% to $35.6 billion, now representing over 90% of total revenue [2] - Despite strong earnings, Nvidia's stock experienced a 17% drop on January 27, attributed to concerns over potential performance gains from competitors like DeepSeek, which suggested lower infrastructure costs for AI [3] Company Performance - Nvidia's fourth-quarter earnings exceeded analysts' expectations, showcasing robust growth in both overall and data center revenues [2] - The data center segment, crucial for AI workloads, has become the dominant revenue source for Nvidia, highlighting the company's leadership in the GPU market [2] Competitive Landscape - Huang countered claims from DeepSeek regarding the feasibility of achieving high AI performance with lower infrastructure costs, asserting that reasoning models will necessitate more chips [3] - DeepSeek's open-sourced reasoning model was acknowledged by Huang as a significant advancement in the field, indicating the competitive pressure Nvidia faces [4]

发生了哪些财经大事？ 2月26日下午，中金公司与中国银河证券的合并传闻引起了市场高度关注。晚间，两家公司双 双对该传闻进行了回应。 26日晚间，中金公司发布公告称，经公司与控股股东中央汇金投资有限责任公司确认，控股股 东不存在筹划上述传闻所称事项或其他涉及公司的应披露而未披露的重大事项。中国银河也发 布澄清公告，未得到任何来自政府部门、监管机构或公司控股股东、实际控制人有关上述传闻 的书面或口头的信息。 盘面上，尾盘中金公司、中国银河双双直线拉升封涨停。中国银河港股涨17.11%，中金公司港 股拉升超19%。 金价突然大跳水！ 2月25日，国际金价出现大幅跳水！截至26日07:00，伦敦金现一度跌破2900美元/盎司，最大跌 幅约2%，最低报2890.2美元/盎司；COMEX黄金失守2930美元/盎司，跌幅1.17%，收报2928.6 美元/盎司；沪金大跌1.18%，收报678.8元/克。2月26日，周大福、老凤祥、周六福等品牌足金 饰品的价格跌至886元/克。此前，周大福等足金饰品最高价为895元/克。相当于买一个30克的 金手镯省出270元。 净利大增80%！ 英伟达最新财报公布 英伟达公布财报显示，英伟 ...

2月26日，金融监管总局、国家发展改革委联合召开金融资产投资公司（下称"AIC"）股权投资试点座 谈会。会议提出，金融监管总局将加强与国家发展改革委的协同配合，推动优化股权投资环境，调动更 多资金和资源支持试点。 扩大AIC股权投资试点是稳经济增长一揽子增量措施之一。去年9月24日，金融监管总局宣布将AIC股权 投资试点的区域由上海扩大至北京、天津、上海、重庆、南京、杭州、合肥、济南、武汉、长沙、广 州、成都、西安、宁波、厦门、青岛、深圳、苏州等18个大中型城市。同时，适当放宽股权投资金额和 比例限制，将表内投资占比由原来的4%提高到10%，投资单只私募基金的占比由原来的20%提高到 30%。 会议指出，试点以来，AIC迅速响应，已实现18个试点城市签约全覆盖，签约金额超过3500亿元；认真 落实"投早、投小、投长期、投硬科技"要求，撬动社会资金参与试点，基金设立、募资和项目投资等各 项工作顺利有序推进，取得积极成效。 科创100、科创200指数方案即将优化 重要的消息有哪些 两部门：推动优化股权投资环境 调动更多资金资源支持试点 昨日，上海证券交易所与中证指数公司联合发布公告称，拟对上证科创板100指数和上 ...

要闻提示 NEWS REMIND 1.传奔驰中国裁员15%！赔偿N+9，全球9万员工奖金打折 2. 纵目科技创始人回应跑路传闻：没有失联，正在寻求海外融资，可能性更多元 3. 前员工谈梁文锋：他视我们为专家，一起学习，没有996，下放管理权 4. 网传菜鸟开启大规模裁员：比例20-50%，周一谈周四走，被证实消息不实 5.苏州知名芯片公司合芯科技爆雷，董事长跑路被限高，自研高端CPU成泡影 6.TikTok吸金能力再创记录，成为历史首个年应用内消费达60亿美元的产品，是第二名的2倍多 7.曝苹果首款折叠屏iPhone誓要「零折痕」，三星、安费诺联手攻关 8.英伟达凌晨发布财报，业绩超预期，黄仁勋感叹：需求惊人 今日头条 HEADLINE NEWS 传奔驰中国裁员15%！赔偿N+9，全球9万员工奖金打折 2月26日，车fans创始人孙少军透露，BBA中的一家厂商已开始人员优化，并且今年是第一年，之后两年会持 续调整，"这家车企今年优化人员占比超过10%，给N+10补偿。"同一天，多个信源表示奔驰中国区开启了大 规模裁员，并对被裁员工给出了N+9的丰厚赔偿金。对于这些消息，截至发稿前，官方并未给出回应。 据悉，小 ...

Core Viewpoint - The AI large model sector is experiencing a new wave of price reductions, driven by competition and advancements in algorithms and computing cost control [4][5]. Group 1: Price Reductions and Promotions - DeepSeek has launched a time-limited discount for its API calls, reducing DeepSeek-V3 to 50% of its original price and DeepSeek-R1 to 25% [2]. - Major companies like ByteDance, Alibaba Cloud, Tencent, and iFlytek have announced significant price cuts for their models, with ByteDance's Doubao visual understanding model dropping to 0.003 yuan per thousand tokens, an 85% reduction from the industry average [4]. - Free service strategies are being adopted by several firms, with Baidu's Wenxin Yiyan and OpenAI's ChatGPT offering free access starting April 1 [4]. Group 2: Technological Advancements - NVIDIA has open-sourced a new optimization scheme based on the Blackwell architecture, achieving a 20-fold reduction in single token costs and a 25-fold performance increase compared to the previous H100 model [2][3]. - The new FP4 quantization model reaches 99.8% performance of FP8 models in the MMLU general intelligence benchmark, showcasing significant advancements in model efficiency [3]. Group 3: Market Reactions and Investments - The A-share market has seen a surge in large model concept stocks, with 69 companies experiencing an increase, and a net inflow of 90.02 billion yuan in leveraged funds [5]. - Notable stock increases include Capital Online rising by 13.71% and several others by over 5% [5]. - China Unicom has optimized its "adaptive slow thinking" model for DeepSeek series, saving approximately 30% in inference computation [5]. Group 4: Industry Dynamics - The open-sourcing of DeepSeek's cost-reduction methodology is expected to lower training and inference costs across the industry, stimulating rapid expansion of AI applications [5]. - The competitive landscape is intensifying, with companies focusing on algorithm optimization and iteration as key strategies alongside price reductions [4].