——英伟达 FY26Q4 业绩点评 | [姓名table_Authors] | 电话 | 邮箱 | 登记编号 | | --- | --- | --- | --- | | 秦和平(分析师) | 0755-23976666 | qinheping@gtht.com | S0880523110003 | | 刁云鹏(研究助理) | 021-38674878 | diaoyunpeng@gtht.com | S0880125070016 | 本报告导读： 长期收入上调，毛利率维持坚挺。Agent 应用拐点已至，英伟达将凭借最优的 Token 成本，持续领跑 AI 基础设施。 投资要点： | 财务摘要(百万美元) | 2025A | 2026A | 2027E | 2028E | 2029E | | --- | --- | --- | --- | --- | --- | | 营业收入 | 130,497 | 215,938 | 380,098 | 523,847 | 637,418 | | 同比增长 | 114.2% | 65.5% | 76.0% | 37.8% | 21.7% | | 毛利润 | 97,859 | ...

指引超预期，Token 经济学的最佳增长引擎 英伟达(NVDA.O) ——英伟达 FY26Q4 业绩点评 | [姓名table_Authors] | 电话 | 邮箱 | 登记编号 | | --- | --- | --- | --- | | 秦和平(分析师) | 0755-23976666 | qinheping@gtht.com | S0880523110003 | | 刁云鹏(研究助理) | 021-38674878 | diaoyunpeng@gtht.com | S0880125070016 | 本报告导读： 长期收入上调，毛利率维持坚挺。Agent 应用拐点已至，英伟达将凭借最优的 Token 成本，持续领跑 AI 基础设施。 投资要点： | 财务摘要(百万美元) | 2025A | 2026A | 2027E | 2028E | 2029E | | --- | --- | --- | --- | --- | --- | | 营业收入 | 130,497 | 215,938 | 380,098 | 523,847 | 637,418 | | 同比增长 | 114.2% | 65.5% | 76.0% ...

Core Insights - Nvidia has made significant progress in AI inference with its Blackwell architecture, achieving a milestone in "token economics" [1] - The company has implemented an "extreme hardware-software co-design" strategy, optimizing hardware efficiency for complex AI inference workloads, reducing the cost of token generation to one-tenth compared to the previous Hopper architecture [1] Industry Applications - Several inference service providers, including Baseten, DeepInfra, Fireworks AI, and Together AI, are utilizing the Blackwell platform to host open-source models [2] - These companies have successfully achieved cross-industry cost reductions by combining cutting-edge open-source intelligent models, Blackwell's hardware advantages, and their own optimized inference stacks [2] - For instance, Sentient Labs, focusing on multi-agent workflows, reported a cost efficiency improvement of 25% to 50% compared to the Hopper era, while companies in the gaming sector, like Latitude, have achieved lower latency and more reliable responses [2] Technical Specifications - The core of Blackwell's efficiency lies in its flagship system, the GB200 NVL72, which features a configuration of 72 interconnected chips and up to 30TB of high-speed shared memory [6][7] - This design is well-suited for the current mainstream "Mixture of Experts (MoE)" architecture, allowing for efficient splitting and parallel processing of token batches across multiple GPUs [6][7]

Core Insights - OpenClaw, an open-source AI Agent project, has rapidly gained popularity, with monthly user visits skyrocketing to 2.63 million, a 10,000% increase in just two weeks, leading to significant demand for cloud services and MaaS orders [1] - Major cloud providers like Alibaba Cloud, ByteDance's Volcano Engine, and Baidu Intelligent Cloud are competing fiercely to support OpenClaw deployments, indicating a shift in the cloud market dynamics [1][2] - The MaaS market is projected to grow significantly, with a year-on-year increase of 421.2% expected by mid-2025, driven by the rise of AI Agents and the associated token economy [2][3] Group 1: Market Dynamics - ByteDance's Volcano Engine leads the domestic market with a 46% share of the public cloud model market, actively expanding its sales team to capitalize on the growing demand for MaaS [2][8] - Baidu's Wenxin model has seen substantial growth, with daily API calls reaching 500 million in Q4 2023, and projected to exceed 1.65 billion by December 2024 [4] - The competition among cloud providers is intensifying as they seek to capture market share in the rapidly expanding MaaS sector, with Alibaba focusing on reclaiming lost ground [15][17] Group 2: Technological Advancements - The concept of "Token efficiency" is becoming crucial as the demand for AI applications grows, emphasizing the need for cloud providers to offer low latency, high stability, and transparent cost structures [3] - Companies are exploring new architectures to improve token efficiency, which is essential for handling complex tasks with fewer tokens [3] - The integration of AI models with cloud services and proprietary chips is seen as a strategic advantage for companies like Alibaba, aiming to create a comprehensive AI cloud ecosystem [17][18] Group 3: Future Projections - The token economy is expected to reach trillions in value as the demand for AI Agents increases, with OpenClaw serving as a catalyst for this growth [2][3] - By 2027, the daily token consumption for ByteDance's Doubao model is predicted to exceed 100 trillion, highlighting the immense potential for growth in the MaaS market [10] - The competition will not only be about token consumption but also about the ability to provide efficient and effective AI solutions that meet developer needs [21]

Core Insights - The report highlights a significant shift in AI inference economics, where the focus has moved from raw chip performance to the intelligence output per dollar spent [1][4][46] - NVIDIA continues to dominate the market, with its GB200 NVL72 outperforming AMD's MI350X by a factor of 28 in throughput [1][5][18] AI Inference Economics - The key metric for evaluating AI infrastructure has transitioned to "how much intelligence can be obtained for each dollar" [4][6][46] - In high-interaction scenarios, the cost per token for DeepSeek R1 can be reduced to 1/15th of other solutions [2][20] Model Architecture - The report discusses the evolution from dense models to mixture of experts (MoE) models, which activate only the most relevant parameters for each token, improving efficiency [9][11][46] - MoE models are becoming the standard for top open-source large language models (LLMs), with 12 out of the top 16 models utilizing this architecture [11][14] Performance Comparison - In terms of performance, the GB200 NVL72 shows a significant advantage over AMD's MI355X, achieving up to 28 times the performance in certain scenarios [18][24][30] - The report indicates that as interaction rates increase, the performance gap between NVIDIA and AMD platforms widens, with NVIDIA's solutions becoming increasingly efficient [30][37] Cost Efficiency - Despite the higher hourly cost of the GB200 NVL72, its advanced architecture and software capabilities lead to a lower cost per token, making it more economical in the long run [20][41][45] - The analysis shows that the GB200 NVL72 can achieve a performance per dollar advantage of approximately 12 times compared to its competitors [42][44] Future Trends - The future of AI models is expected to lean towards larger and more complex MoE architectures, with platform-level design becoming a critical factor for success [46][47] - Companies like OpenAI, Meta, and Anthropic are likely to continue evolving their flagship models in the direction of MoE and inference, maintaining NVIDIA's competitive edge [46]

Group 1 - The core viewpoint of the news highlights the positive market expectations surrounding the acquisition of Zhenai Meijia by Tanjin Technology, which has led to a surge in Zhenai Meijia's stock price and market capitalization reaching 7.4 billion yuan [1] - Tanjin Technology has established a comprehensive capability loop covering the entire AI Agent ecosystem, with its self-developed "Taiqing" enterprise-level model and "Konghu" data cloud foundation, supporting commercial applications [2] - The acquisition of Manus by Meta is seen as a strategic move to enhance its product ecosystem, focusing on AI Agent applications, which aligns with the broader trend of AI Agent commercialization [1][2] Group 2 - The rapid growth in the AI sector is underscored by the significant increase in revenue and market value of related companies, with predictions that 2026 may mark the year of widespread AI Agent adoption [2] - Tanjin's B2C AI Agent has achieved a daily token consumption exceeding 50 billion, with a monthly growth rate of nearly 20%, positioning it among the top 10 in the To B sector [2] - Industry experts suggest that AI Agents are evolving from mere tools to foundational infrastructure, potentially creating higher-dimensional commercial value through deep integration with existing workflows and intelligent intercommunication between agents [3]

Core Insights - The article emphasizes the significance of Tokens in measuring the performance and commercial viability of AI models, shifting the focus from what AI can do to quantifying its efficiency, cost, and value [1][14][16] Group 1: Token Consumption and Revenue - Token consumption is closely linked to computational power, which in turn correlates with revenue for model providers [2] - OpenAI's token usage on Microsoft Azure is projected to increase from 0.55 trillion to 4.40 trillion daily tokens between June 2024 and June 2025, with annual revenue expected to rise from $5.5 billion to over $10 billion [3] Group 2: Consumer and Business Applications - Major contributors to consumer token consumption include AI features in high-traffic applications like Google Search and Douyin, with Google’s AI Overview feature projected to consume between 1.6 trillion and 9.6 trillion tokens daily [4][5] - ChatGPT remains a significant driver of token consumption, with a combined monthly active user base of 1.015 billion across app and web platforms as of July 2025 [7] Group 3: Business Applications and Market Penetration - Business applications are seeing high penetration rates, with OpenAI's B2B revenue expected to account for 54% of its annual recurring revenue by 2025 [9] - Google has reported over 85,000 enterprise customers for its Gemini model, leading to a 35-fold increase in token consumption [9] Group 4: Technological Advancements - The increase in token consumption is attributed to advancements in reasoning capabilities, multi-modality, agent-based systems, and longer context lengths, which enhance the practical application of AI [10][12] - New models like GPT-5 and Grok4 are designed to improve AI's usability in complex scenarios, thereby increasing token consumption [11] Group 5: Pricing Dynamics - Despite the increase in token consumption, the pricing for tokens is decreasing due to competitive pricing strategies and optimization of computational costs by model providers [13] - The introduction of tiered pricing models allows smaller clients to access AI capabilities, further driving token consumption [13] Group 6: Economic Implications - Understanding token economics provides insights into cost-effectiveness, technological efficiency, and the evolution of application scenarios, marking a shift towards a more mature and industrialized AI sector [14][16]

Core Insights - The article reveals a fundamental challenge faced by the AI industry: the scarcity of computational resources [1] - It analyzes the contrasting strategies of DeepSeek and Anthropic in navigating this challenge [4][42] - The report emphasizes the importance of balancing technological breakthroughs and commercial success within limited computational resources [58] Group 1: AI Service Pricing Dynamics - AI service pricing is fundamentally a trade-off among three performance metrics: latency, throughput, and context window [2][3] - Adjusting these three parameters allows service providers to achieve any price level, making simple price comparisons less meaningful [30] - DeepSeek's extreme configuration sacrifices user experience for low pricing and maximized R&D resources [4][39] Group 2: DeepSeek's Market Performance - After the initial launch, DeepSeek experienced a significant drop in its own platform's user base, with a 29% decrease in monthly active users [15][12] - In contrast, the usage of DeepSeek models on third-party platforms surged nearly 20 times, indicating a shift in user preference [16][20] - The low pricing strategy of DeepSeek, at $0.55 per million tokens for input and $2.19 for output, initially attracted users but could not sustain long-term engagement [6][7] Group 3: Token Economics - Tokens are the fundamental units in AI, and their pricing is influenced by the service provider's ability to manage latency, throughput, and context window [21][22] - DeepSeek's official service has become less competitive in terms of latency compared to other providers, leading to a decline in its market share [33] - The context window offered by DeepSeek is the smallest among major providers, limiting its effectiveness in applications requiring extensive memory [34] Group 4: Anthropic's Resource Constraints - Anthropic faces similar computational resource challenges, particularly after the success of its programming tools, which increased demand for resources [44][45] - The API output speed of Anthropic's Claude has decreased by 30%, reflecting the strain on its computational resources [45] - Anthropic is actively seeking additional computational resources through partnerships with Amazon and Google [46][48] Group 5: Industry Trends and Future Outlook - The rise of inference cloud services and AI-driven applications is reshaping the competitive landscape, with a shift towards direct token sales rather than subscription models [51] - The article suggests that as affordable computational resources become more available, the long-tail market for AI services will continue to grow [52] - The ongoing price war among AI service providers is merely a surface-level issue; the deeper challenge lies in achieving technological advancements within resource constraints [58]

Core Insights - The article highlights the critical challenge faced by AI companies: the scarcity of computational resources, which is a fundamental constraint in the industry [1][5]. Pricing Dynamics - AI service pricing is fundamentally a trade-off among three performance metrics: latency, throughput, and context window [2][3]. - By adjusting these three parameters, service providers can achieve any price level, making simple price comparisons less meaningful [4][24]. DeepSeek's Strategy - DeepSeek adopted an extreme configuration with high latency, low throughput, and a minimal context window to offer low prices and maximize R&D resources [4][28]. - Despite DeepSeek's low pricing strategy, its official platform has seen a decline in user engagement, while third-party hosted models have surged in usage by nearly 20 times [16][20]. Competitive Landscape - Anthropic, another leading AI company, faces similar resource constraints, leading to a 30% decrease in API output speed due to increased demand [34][36]. - Both DeepSeek and Anthropic illustrate the complex trade-offs between computational resources, user experience, and technological advancement in the AI sector [5][53]. Market Trends - The rise of inference cloud services and the popularity of AI applications are reshaping the competitive landscape, emphasizing the need for a balance between technological breakthroughs and commercial success [5][45]. - The article suggests that the ongoing price war is merely a surface-level issue, with the real competition lying in how companies manage limited resources to achieve technological advancements [53].