证券研究报告 | 2026年02月27日 财报重点：①计算业务同比增长 57%，GB 系列收入占数据中心总营收超 过 2/3，主要是 Blackwell 和 Blackwell Ultra 产能爬坡。②各大云厂、 数据中心运营商、模型公司等部署的 Blackwell 基础设施算力规模已经 达到 9GW。自 2023 财年推出 ChatGPT 以来，数据中心业务规模已增长近 13 倍。前五家云服务商和超大规模数据中心运营商目前贡献了公司约 50%的营收。③26 财年，主权 AI 业务营收同比增长超三倍，突破 300 亿美元。主要客户来自加拿大、法国、荷兰、新加坡和英国，预计主权 人工智能业务的机遇将持续增长。④网络业务营收同比+263%。全年来 看，网络业务收入超过 310 亿美元，比 2021 年（收购 Mellanox 的那年） 增长了十倍以上。Scale up 和 Scale out 需求创新高。 投资建议：考虑到 GB 系列产品与网络连接出货积极，同时下游 AI 需求 持续旺盛，我们上调预期，预计 2027-2028 财年公司收入为 3449/4304 亿美元（前置 3335/4279 亿美元），上调幅 ...

【国联民生海外】英伟达财报速递：业绩与指引均超预期，Blackwell加速放量 英伟达FY26Q4营收681.27亿美元，同比增长73%，比彭博一致预期超3.24%；毛利率75.2%，超越一致预期 （74.7%）；净利润395.52亿美元，同比增长79%，比彭博一致预期超5.48%；调整后每股收益1.62美元，比彭博一 致预期超5.77%。 业绩指引：预计FY27Q1总收入780亿美元（±2 【国联民生海外】英伟达财报速递：业绩与指引均超预期，Blackwell加速放量 英伟达FY26Q4营收681.27亿美元，同比增长73%，比彭博一致预期超3.24%；毛利率75.2%，超越一致预期 （74.7%）；净利润395.52亿美元，同比增长79%，比彭博一致预期超5.48%；调整后每股收益1.62美元，比彭博一 致预期超5.77%。 业绩会要点： -Blackwell系统已部署近9吉瓦基础设施，且Blackwell系统贡献了约三分之二数据中心收入。 -前五大云厂约占数据中心收入50%，但非hyperscaler增长更快。 2026年五大云厂CapEx预期已上调至接近7000亿美元（较年初上调1200亿美元） -主 ...

Core Viewpoint - Nvidia's FY26Q4 financial report significantly exceeded market expectations, showcasing robust growth in revenue and net profit, reinforcing its position as a leader in the AI chip market amidst ongoing debates about the sustainability of AI growth [4][16]. Revenue Performance - Total revenue reached $68.1 billion, a 73% year-over-year increase and a 20% quarter-over-quarter increase, surpassing the market consensus of $65.9 billion and previous guidance of $65 billion [3]. - Data center revenue for Q4 was $62.3 billion, up 75% year-over-year and 22% quarter-over-quarter, accounting for 92% of total revenue, driven by the ramp-up of Blackwell and Blackwell Ultra [5]. Profitability Metrics - GAAP gross margin was 75%, up 2 percentage points year-over-year and 1.6 percentage points quarter-over-quarter, slightly above the market expectation of 74.9% [8]. - GAAP net profit reached $42.96 billion, a 94% year-over-year increase and a 35% quarter-over-quarter increase, significantly exceeding the market expectation of $36.3 billion [8]. Business Segments - Gaming revenue for Q4 was $3.7 billion, a 47% year-over-year increase but a 13% quarter-over-quarter decline, representing 6% of total revenue [10]. - Networking revenue surged to $11 billion, a 263% year-over-year increase and a 34% quarter-over-quarter increase, marking the first time it surpassed $10 billion in a single quarter [9]. Future Guidance - For FY27Q1, Nvidia expects revenue of $78 billion, a 77% year-over-year increase, significantly above the market consensus of $72.8 billion, primarily due to the ramp-up of the Blackwell Ultra platform [11]. - The company anticipates continued growth in the data center segment, with orders exceeding previous revenue guidance of $500 billion [14]. Market Dynamics - Nvidia's management indicated that the AI market is evolving, with growth shifting from cloud computing to vertical industries and sovereign AI initiatives [13]. - The company has secured supply commitments totaling $95.2 billion, reflecting a significant increase in production capacity for semiconductors and related technologies [12]. Shareholder Returns - Nvidia repurchased $3.8 billion in shares and paid $243 million in dividends, with a remaining buyback authorization of $58.5 billion, indicating a balanced approach between shareholder returns and investment in growth [14].

英伟达以一份打破纪录的财报，试图回击外界对AI泡沫的质疑。但电话会后英伟达股价转跌，此前盘 后交易中一度涨超4%。 2月25日美股盘后，英伟达公布最新单季营收达680亿美元并给出强劲指引，英伟达CEO黄仁勋在电话会 上直言，代理AI(Agentic AI)已达到拐点，算力直接转化为收入。 黄仁勋认为没有算力就无法生成token，没有token就无法实现营收增长，云服务商的巨额资本开支最终 将直接转化为收入。 从整体财务数据与供应链前瞻来看，英伟达的增长引擎依然强劲。CFO克雷斯在电话会开场中透露： 总营收达到680亿美元，同比增长73%，按环比计算较第三季度的增长进一步加速。 她明确指出，大部分收入增长依然由数据中心业务驱动。盘后英伟达一度涨超4%，但电话会后涨幅逐 渐收窄，直至转跌。 黄仁勋同时强调，企业对代理的应用飙升，Grace Blackwell搭配NVLink是推理的王者。此外，他表示 当前的太空数据中心经济还是"贫瘠的"，但形势将随着时间的推移而变化。 未来展望：预计2027财年Q1收入780亿美元，数据中心驱动大部分增长;预计到2030年，全球数据中心资 本支出将达到3-4万亿美元;已战略性确保 ...

●张佳欣 "信任"应成为产品硬指标 "代理权"越位或违背用户意愿 这种安全感的本质，是人类对"代理权"越位的深层警惕。以前，AI是一个"问答机"，人类下指令，AI来 执行。但现在，AI正在向"智能体"进化，这意味着它从被动响应转向了主动执行。 据国外黑客新闻网（The Hacker News）在1月24日发布的文章指出，AI智能体不仅仅是另一种类型 的"用户"。它们与人类用户、传统的服务账户有着本质区别，正是这些差异，导致现有的访问权限和审 批模型全面失效。在实际操作中，为了让AI能高效完成任务，系统赋予智能体的权限往往比用户本人 拥有的更高。这种"访问权限漂移"有可能导致AI在用户不知情、未授权的情况下，执行了技术层面合 法、但违背用户自主意愿的操作。 当"代理人"的技术权力在事实上大过其主人，人类在数字世界的控制权便面临被"架空"的风险。这种权 力的隐形流失，并非源于技术的恶意，而是因为系统在追求效率的过程中，悄无声息地打破了人类 的"数字主权"边界。德勤在1月21日发布的报告中指出，目前AI的"代理权"已经超越了其安全防御措 施。数据显示，全球仅有20%的公司建立了成熟的AI智能体治理模型。这种"五分之 ...

Core Insights - Mistral, a French AI startup, has achieved remarkable growth with an annual recurring revenue (ARR) exceeding $400 million, a 20-fold increase from $20 million a year ago, positioning itself as "Europe's OpenAI" [1][2] - The company plans to surpass $1 billion in ARR by the end of this year, driven by aggressive expansion among large enterprise clients, now exceeding 100 [1][2] - Mistral is investing €1.2 billion to build a new AI data center in Sweden, marking its first facility outside France, aimed at reducing reliance on external infrastructure [1][3] Vertical Integration and Infrastructure Expansion - Mistral is pursuing a vertical integration strategy by constructing and operating its own AI data centers instead of relying solely on major U.S. cloud providers [3] - The new Swedish facility will provide 23MW of computing power and is expected to be operational next year, leveraging low-carbon and relatively inexpensive local energy [4] - This infrastructure investment is projected to generate over €2 billion in revenue over the next five years, providing a predictable business model [4] Geopolitical Drivers of "Sovereign AI" Demand - There is growing concern in Europe regarding over-reliance on U.S. digital services, with over 80% of digital services and infrastructure depending on foreign providers, primarily American companies [5] - Mistral's position as the only homegrown developer of cutting-edge language models in Europe places it in a favorable position to meet the demand for data sovereignty among clients [5] - Current clients include major corporations and various European governments, with approximately 60% of revenue generated from Europe [5] Financial Position and Future Plans - Mistral's CEO indicated that the company does not require an IPO this year due to sufficient debt financing, although it may consider going public in the future to ensure independence [6] - The company is not currently pursuing an IPO, unlike competitors such as OpenAI and Anthropic, which are preparing for public offerings [5][6] Practical Applications and Market Realities - Despite the rapid growth of products like ChatGPT and Claude, Mistral's CEO expressed a pragmatic view of the market, noting that many enterprise clients are disappointed with off-the-shelf chatbot solutions [7] - There is skepticism regarding the notion that a single system can manage all business processes, emphasizing the continued relevance of traditional software companies that hold critical business data [7] - Mistral warns that startups focused solely on creating user interfaces for specific industries may find their strategies less valuable as AI technology evolves [7]

Core Insights - Aixin Yuan Zhi, a company specializing in edge AI inference chips, officially listed on the Hong Kong Stock Exchange on February 10, 2026, becoming the first Chinese edge AI chip stock [4][44]. - The investment journey of Yaotu Capital, which began in 2017, reflects a long-term belief in the potential of edge AI chips, driven by the intelligentization of hardware in consumer electronics and automotive sectors [54][45]. - The emergence of Aixin Yuan Zhi and other companies marks a significant shift in China's AI chip industry, transitioning from cloud computing to edge computing [4][47]. Company Development - Aixin Yuan Zhi successfully captured market share in visual terminal computing and quickly entered the smart automotive sector in 2021, securing multiple orders from car manufacturers, with a cumulative shipment of nearly one million units over four years [49][50]. - The company has attracted strategic investments from major players like Weihua Chuangxin, Meituan, Tencent, and government investment platforms from Ningbo and Chongqing, enabling substantial funding for high-end smart driving chip development [49][50]. - Aixin Yuan Zhi plans to release a significant chip supporting urban NOA (Navigation on Autopilot) in Q2 2026, aiming to penetrate the L2+ high-level autonomous driving market [49][50]. Investment Logic - The investment logic for edge AI chips is based on the belief that Chinese hardware manufacturers will become global players, thus creating a substantial market for domestic edge AI chips [54][54]. - The company’s success is attributed to three key strategies: building a team that integrates AI algorithms and chip design, early investment in the smart automotive market, and leveraging industry collaboration for funding [50][57]. Market Dynamics - The edge AI chip market in China benefits from rich application scenarios, a complete supply chain, and diverse customers, providing opportunities for world-class enterprises to emerge [53][70]. - The competitive landscape indicates that the moat for edge chips lies in ecosystem development rather than performance alone, necessitating collaboration with solution providers to penetrate fragmented markets [53][73]. - The rise of open-source models has significantly lowered adaptation barriers for domestic AI chips, enabling faster scene implementation and forming a "sovereign AI" ecosystem [53][39]. Future Trends - The year 2025 is anticipated to be a significant year for GPUs, while the future will see a rise in network interconnectivity and communication chips, with several companies expected to achieve valuations exceeding $10 billion [53][64]. - The evolution of AI models continuously defines new requirements for AI chips, creating opportunities for technological innovation [66][66]. - The demand for edge AI chips is driven by the need for low latency, strong privacy, and high energy efficiency, particularly in sensitive fields like autonomous driving and robotics [62][63].

Core Insights - The first round of evaluations for South Korea's autonomous AI foundational model selection revealed that LG AI Research Institute scored the highest at 90.2, advancing to the next round, while NAVER Cloud was unexpectedly eliminated due to not meeting originality standards [1][2][3] - The evaluation aimed to objectively assess the capabilities of South Korea's top AI models, reflecting the country's aspirations to develop independent core technologies in the AI field [1][3] Group 1: Evaluation Results - LG AI Research Institute achieved the highest score in all assessment categories, including benchmark tests, expert evaluations, and user feedback [2] - Upstage and SK Telecom also advanced, with Upstage demonstrating competitive performance with relatively fewer parameters [2] - NAVER Cloud was disqualified for using the Chinese open-source AI model Tongyi Qianwen, violating the competition's requirements for model independence [3][11] Group 2: Controversies and Reactions - Three out of five participating teams were recently accused of plagiarism related to Chinese AI technologies, raising concerns about the fairness of the evaluation process [2][3] - The controversy has sparked significant public backlash in South Korea, with citizens expressing disappointment over the perceived decline in technological superiority [3][12] Group 3: Government Initiatives and Funding - The South Korean government plans to invest approximately 530 billion KRW (about 383 million USD) to support the five selected teams, with a focus on providing hardware and ensuring data security for AI training [7][8] - Future evaluations will occur every six months, with one team being eliminated each round until a final winner is determined [8] Group 4: Challenges in the AI Industry - South Korea's AI industry faces significant challenges, including a lack of original large models, over-reliance on external open-source technologies, and slow application of AI across various sectors [12][13][16] - The small population and language barriers limit data accumulation and market size, hindering the development of competitive AI models [16][17] Group 5: Future Directions - To overcome current challenges, South Korea's AI industry must focus on core technology development and hardware support, while fostering a more innovative ecosystem free from conglomerate dominance [30][31] - The government has introduced new laws to enhance the AI development environment, aiming to address the fundamental issues hindering the industry's growth [25][27]

Core Insights - The rise of AI agents has led to concerns about the erosion of human control in the digital realm, as these systems may operate beyond user intent and authority [2][4] - Trust has emerged as a critical factor in AI product design, shifting from a soft advantage to a hard metric that influences user engagement and control [2][7] Group 1: AI Evolution and User Control - AI is transitioning from a passive responder to an active agent, which raises questions about the delegation of decision-making authority [3][4] - Current governance models for AI are inadequate, with only 20% of companies having established mature AI governance frameworks, leaving most users vulnerable [4] Group 2: Redefining Human-Machine Authority - The concept of "meaningful oversight" is being introduced to ensure that AI systems are transparent and understandable to users, rather than operating as opaque black boxes [5][6] - A dual authorization framework is being promoted, separating AI's access to data from its ability to take action, thereby restoring decision-making power to humans [6] Group 3: Trust as a Product Metric - The younger generation, growing up with AI, is increasingly questioning the trade-offs of data sharing with cloud giants, leading to a demand for localized and private AI solutions [7] - The next generation of users prioritizes autonomy and control over their data and AI behavior, indicating a shift in what constitutes value in AI products [7]

Core Insights - The focus of world leaders at this year's Davos Forum has shifted from hype to investment returns, emphasizing the need for effective real-world deployment of AI [2] - Discussions on "sovereign AI," open ecosystems, and the risks of over-reliance on any single country or company have gained prominence [2] Group 1: AI Deployment and Investment Returns - There is a growing pressure for AI to deliver tangible results, with leaders expressing the need for AI to provide actual returns rather than just exploring possibilities [3] - The conversation has moved towards enhancing worker productivity and transforming workflows rather than merely replacing jobs [6] - A significant tech company executive indicated a target of increasing business growth by $40 billion over the next three to five years without hiring additional staff, highlighting the role of AI in boosting individual productivity [7] Group 2: Sovereign AI - "Sovereign AI" refers to nations seeking greater control over their AI futures to address geopolitical uncertainties and the dominance of large tech companies [4] - Different countries emphasize various aspects of the AI stack, such as computing, data, models, applications, and talent, based on their specific goals [4] - Research is being conducted to define the different angles of AI sovereignty and the benefits associated with each perspective [5] Group 3: Open Ecosystems and Trust - There is a preference for open ecosystems, where open data and models can enhance transparency and trust in technology, thereby accelerating innovation [5] - Collaboration with institutions like ETH Zurich is part of a broader global effort to work on open models and related initiatives [5] Group 4: AI Agents and Ethical Considerations - AI agents are emerging in two forms: practical implementations within companies and independent agents negotiating information and funding online [8] - Caution is advised regarding the latter, especially concerning personal or financial data, as significant research and infrastructure are needed to build trust [8] - Ethical education and professional standards for those building AI systems are essential, along with regulatory frameworks to address potential shortcuts or cheating [11]