国产 AI 的风口来了。 12 月 5 日，「国产 GPU 第一股」摩尔线程上市，五天内暴涨 700%，刷新了市场对国产算力的认 识，17 日，沐曦股份上市，最高涨幅744.64%。热潮背后，市场对「国产 AI」概念既缺乏足够了 解，又怀有灼灼期待。 仅仅几年前，国产算力还只是断供背景下，不得已要做的一个「备胎」选项。但现在，随着越来越多 的力量汇入这股潮流，它正在成为 AI 领域最不可忽视的趋势之一。 国产从最初的「替代方案」， 成长为一套从硬件、模型到应用的完整体系，种下了未来生态化的种子。 12 月 15 日，商汤发布了行业首个多剧集生成智能体 Seko 2.0，并于发布当天寒武纪即完成了与商 汤"日日新"的适配，成为寒武纪官宣「Day 0」适配的又一大模型体系，此前是DeepSeek。 国产 AI 的技术生态构建，已经从最初的模型兼容适配，发展到了效用提升、应用建构、生态协同的 新阶段，实现了从「可用」到「好用」、「易用」的蜕变。 AI 国产化的基石：模型适配 题图｜视觉中国 AI 行业尝试用国产算力作为替代方案的初期，悬在头顶的最大问题在于算力和模型适配。 在这个「从无到有」的阶段，国产芯片的核心目 ...

Core Insights - The company achieved a revenue of 2.36 billion yuan in H1 2025, representing a year-on-year growth of 35.6%, driven by the increasing contribution of generative AI to overall revenue [1] - The overall gross margin decreased to 38.5% in H1 2025 from 44.1% in H1 2024 due to the rising share of generative AI revenue [1] - Adjusted net loss narrowed significantly to 1.16 billion yuan, a 50% year-on-year reduction, while adjusted EBITDA loss decreased by 72.5% to 520 million yuan [1] Revenue Breakdown - Generative AI revenue reached 1.82 billion yuan in H1 2025, up 60.4% year-on-year, accounting for 77.0% of total revenue, an increase from 60.4% in H1 2024 [1] - Visual AI revenue, which now includes smart cockpit-related business from the original smart automotive segment, was 436 million yuan, growing 29.4% year-on-year [1] - The X Innovation business, encompassing smart driving, smart healthcare, home robots, and smart retail, generated 107 million yuan, reflecting a 10.2% year-on-year increase [1] Operational Efficiency - Sales expense ratio, administrative expense ratio, and R&D expense ratio were 11.6%, 25.3%, and 89.8%, respectively, showing declines of 7.4 percentage points, 16.9 percentage points, and 18.9 percentage points year-on-year [1] - Operating cash outflow improved significantly, narrowing by 81.9% to 640 million yuan [1] Business Highlights - The company’s large model, "日日新V6.5," has performance metrics that are competitive with top international models, achieving comprehensive testing results on par with Claude 4 Opus [2] - Total computing power reached 25,000 PetaFLOPS as of August 2025, with a 20% year-on-year improvement in computing power per watt [2] - The productivity tool "小浣熊" surpassed 3 million users, processing approximately 10 billion tokens daily [2] Future Outlook - Revenue forecasts for 2025-2027 have been adjusted to 4.86 billion, 6.21 billion, and 7.85 billion yuan, reflecting slight changes from previous estimates [3] - Non-GAAP net profit forecasts for the same period have been revised to -2.14 billion, -230 million, and 170 million yuan, indicating improved operational efficiency [3] - The synergy between large models, computing power, and AI applications is expected to create a robust ecosystem, enhancing the company's competitive edge [3]

Investment Rating - The report assigns an "Outperform" rating for the company, indicating an expected relative return exceeding 10% over the next 12-18 months [26]. Core Insights - The company achieved a revenue of RMB 2.358 billion in H1 2025, a year-over-year increase of 35.6%, driven primarily by its Generative AI business, which saw a revenue increase of 72.7% to RMB 1.816 billion, contributing 77% to total revenue [10][14]. - The gross profit rose by 18.4% to RMB 908 million, with a gross margin of 38.5%. Adjusted net losses narrowed by 50% to RMB 1.162 billion, and adjusted EBITDA losses decreased by 72.5% to RMB 521 million, indicating significant improvement in profitability quality [10][14]. - The company's cash reserves at the end of the period stood at RMB 13.158 billion, reflecting strong financial health [10]. Summary by Sections Performance Overview - In the first half of 2025, the company reported a revenue of RMB 2.358 billion, a 35.6% increase year-over-year, exceeding market expectations. The Generative AI segment's revenue reached RMB 1.816 billion, marking a 72.7% increase and accounting for 77% of total revenue [10][14]. - Gross profit increased by 18.4% to RMB 908 million, with a gross margin of 38.5%. Adjusted net loss decreased by 50% to RMB 1.162 billion, and adjusted EBITDA loss reduced by 72.5% to RMB 521 million, showcasing improved loss management [10][14]. Strategic Infrastructure - The company has developed a "Compute-Model-Application" framework, achieving a total computing power of approximately 25,000 PetaFLOPS. The SenseCore 2.0 platform has been upgraded and certified at the highest level for large model inference capabilities [11]. - The domestic chip heterogeneous cluster operates at a scale of 5,000 cards with an 80% utilization rate and 95% training efficiency, positioning the company among the top three in China for platform strength [11]. Large Models and Applications - The company launched the "Rì Rì Xīn V6.0" model in April and upgraded to V6.5 in July, achieving significant advancements in multi-modal technologies. The user base for its "Little Raccoon" data analysis products surpassed 3 million, with a 510% increase in multi-modal interaction duration [12]. - The model's cost-effectiveness improved by approximately three times, and the application penetration in sectors like government and finance accelerated significantly [12]. Visual AI and Innovative Business - The Visual AI segment focuses on high-quality clients, with the "Ark" platform now serving nearly 200 cities and over 30,000 locations, achieving over 100 million daily API calls. The company maintains a leading position in the smart cabin sector [13]. - The X Innovative Business segment has launched various products, including a co-branded home robot with Disney and healthcare solutions in Singapore, enhancing its market presence [13]. Future Outlook - Management anticipates that Generative AI will continue to be the core growth driver, with a focus on replicable solutions for high-value industries. Key areas to watch include the large-scale deployment of the V6.5 model and advancements in computing infrastructure [14][15].

Core Viewpoint - The article emphasizes the strategic importance of computing power in national strength, highlighting breakthroughs in chip performance, explosive demand for inference, and the large-scale implementation of AI applications, which are reshaping the global technology competition landscape [1]. Group 1: AI Development Trends - The trend of large models potentially ending the "80/20 rule" in software usage is discussed, indicating a shift in product design logic towards more customizable solutions [4][5]. - The emergence of general-purpose tools like MCP and the need for adaptable agent models that can handle diverse tasks is highlighted, suggesting a move away from traditional product design towards more versatile AI applications [5][6]. - The article notes that the cost of large models is expected to decrease significantly each year, indicating a potential tipping point for AI applications to transition from pilot projects to widespread adoption [8][9]. Group 2: Market Opportunities and Challenges - The article identifies key pain points in the enterprise AI office market, particularly the challenge of "good-looking but not user-friendly" AI applications, which have high usage barriers and data quality issues [14][15]. - It discusses the importance of integrating AI capabilities with everyday tasks to enhance user experience and productivity, particularly in sectors like software development and data analysis [16][17]. - The potential for collaborative investment opportunities in the domestic AI ecosystem, focusing on the synergy between software, hardware, and services, is emphasized as a significant growth area [18][19]. Group 3: Future Outlook - The article suggests that AI's penetration and commercialization will gradually improve across various sectors, with a focus on enhancing user experience and meeting diverse user needs [11][12]. - It predicts that as more individuals and companies invest in AI, the timeline for widespread adoption will accelerate, leading to a transformative impact on the workforce [13]. - The article concludes with a positive outlook on the future of AI in China, emphasizing the importance of computing power in achieving a competitive edge globally [19].