分析师：张刚 登记编码：S0730511010001 zhanggang@ccnew.com 021-50586990 晨会聚焦 证券研究报告-晨会聚焦 发布日期：2025 年 12 月 09 日 资料来源：聚源，中原证券研究所 -13% -7% -1% 5% 10% 16% 22% 28% 2024.12 2025.04 2025.08 2025.12 上证指数 深证成指 | 国内市场表现 | | | | | | --- | --- | --- | --- | --- | | 指数名称 | | 昨日收盘价 | 涨跌幅(%) | | | 上证指数 | | 3,924.08 | | 0.54 | | 深证成指 | | 13,329.99 | | 1.39 | | 创业板指 | | 2,022.77 | | -0.47 | | 沪深 | 300 | 4,621.75 | | 0.81 | | 上证 | 50 | 2,443.97 | | -0.52 | | 科创 | 50 | 891.46 | | 0.14 | | 创业板 | 50 | 1,924.26 | | -0.67 | | 中证 | 100 | 4,49 ...

Core Insights - UBS highlights the recent advancements in AI with the launch of new large language models (LLMs) by companies like Google, Anthropic, and DeepSeek, intensifying competition in the industry [1] - The report emphasizes the continued relevance of the "scaling laws" in model performance, indicating that computational power will remain a critical factor in determining the AI competitive landscape [1] Model Performance - The latest generation of models has shown significant breakthroughs, with Gemini 3 Deep Think and Claude Opus 4.5 achieving multi-step reasoning task scores of 45% and 38%, respectively, surpassing previous models that scored between 10%-20% [2] - This performance aligns with the effectiveness of the AI model pre-training scaling laws, where increased computational investment leads to non-linear improvements in model capabilities [2] Chip Technology Competition - Google’s Gemini 3 Pro is trained entirely on self-developed TPU chips, sparking discussions about the competition between GPUs and AI-specific ASIC chips [2] - ASIC chips are noted for their higher efficiency in specific AI tasks, while GPUs maintain a 90% market share in data center chips due to their flexible architecture and extensive software ecosystem [2] - The collaboration between OpenAI and Broadcom, as well as Anthropic and Google, is expected to enhance the focus on ASIC chips, with both chip types anticipated to coexist in the future [2] Market Trends - The introduction of next-generation chips like NVIDIA's Blackwell and Rubin is expected to sustain the competition for computational expansion, leading to an upward revision of AI capital expenditure forecasts by UBS [3] - The advancements from Google, Anthropic, and DeepSeek are increasing competitive pressure on companies like OpenAI, driving the AI industry towards a multi-model and multi-vendor landscape, a trend expected to persist at least until 2026 [3]

Chinese entrepreneur Liang Wenfeng, the founder and CEO of DeepSeek, appears to have made a big score from the initial public offering of chip designer Moore Threads Technology, whose shares surged more than fivefold in its trading debut in Shanghai. High-Flyer Quantitative Fund, a hedge fund co-founded by Liang, had subscribed to a total of 82,244 shares of Moore Threads, at 114.28 yuan per share, via two entities: Zhejiang High-Flyer Asset Management and Ningbo High-Flyer Quant Investment Management Par ...

Core Insights - China's open-source artificial intelligence models accounted for nearly 30% of total global usage, with Chinese-language prompts ranking second in token volume behind English [1][4]. Group 1: Market Share and Growth - Chinese open-source large language models (LLMs) have seen significant growth, rising from a low base of 1.2% in late 2024 to nearly 30% in 2025 [4]. - In 2025, Chinese open-source LLMs averaged 13% of weekly token volume, closely matching the 13.7% average recorded by AI models from the rest of the world [5]. Group 2: Competitive Landscape - Proprietary Western models, such as OpenAI's GPT-4o and GPT-5, maintain a dominant global share of 70% [3]. - The rise of Chinese open-source AI models reflects competitive quality, rapid iteration, and dense release cycles [7]. Group 3: Key Players - Major contributors to the surge in open-source LLM usage include Alibaba Group Holding's Qwen family of models, DeepSeek's V3, and Moonshot AI's Kimi K2 [2]. - The aggressive release schedules of Alibaba Cloud's Qwen and DeepSeek have enabled users to adapt quickly to increased development workloads [8]. Group 4: Industry Position - The report indicates that China has emerged as a major force in AI model development, becoming a close peer to the United States despite restrictions on access to advanced graphics processing units [6].

Group 1 - The humanoid robot industry is experiencing significant positive developments, with leading companies actively engaging in technology research and commercial implementation [1][2] - Tesla's CEO Musk shared a video of the "Optimus" humanoid robot running, indicating advancements in the robot's capabilities [1] - The launch of the full-size, high-efficiency humanoid robot T800 by Zhongqing Robotics marks the beginning of its sales process [1] Group 2 - Domestic and international companies are increasingly entering the humanoid robot market, with notable advancements in commercialization by firms like Tesla and Figure AI [2] - The emergence of AI companies such as DeepSeek is facilitating the development of general-purpose humanoid robot models, leading to a diverse and competitive industry landscape [2] - The humanoid robot sector is entering industrial applications, indicating a strong trend towards commercial viability [2]

Investment Rating - The report maintains an investment rating of "Outperform the Market - A" [7] Core Insights - DeepSeek has launched the V3.2 model, enhancing its reasoning capabilities to a globally leading level, suitable for everyday use in Q&A and general agent tasks [12][27] - The V3.2 model achieved performance comparable to GPT-5 in benchmark tests, slightly below Gemini-3.0-Pro, while significantly reducing output length and computational costs [12][27] - The introduction of the DSA (DeepSeek Sparse Attention) mechanism reduces context computation costs, changing complexity from O(L²) to O(Lk), where k is a fixed value of 2048 [13][14] - The report highlights the launch of the Doubao mobile assistant, which integrates AI capabilities into mobile operating systems, allowing users to perform complex tasks with voice commands [15] Summary by Sections Industry Performance - The computer sector underperformed relative to the Shanghai Composite Index, with a 1-month relative return of -5.4% and a 3-month return of -4.5% [5][16] - The computer sector index ranked 25th among 30 industry indices, indicating weaker performance [19] Important Industry News - Google’s TPUv7 has begun to challenge NVIDIA's dominance in AI chips, marking a significant shift in the competitive landscape [25] - The 2025 World Computing Conference showcased advancements in computing systems, emphasizing the importance of system capabilities over individual card performance [26]

Core Insights - The 2026 will be the year of AI application in China, emphasizing the necessity for individuals to have their digital avatars to avoid being left behind by the times [3][5][18] - The narrative of Sun Wukong illustrates common pitfalls for middle management, including the ability trap, efficiency trap, and identity misalignment, which can hinder organizational growth [3][12][25] - The journey of Sun Wukong serves as a metaphor for entrepreneurial growth, highlighting the importance of reflection and maturity after facing challenges [3][10][24] Industry Transformations - Transition from scale-driven growth to efficiency-driven growth, indicating a shift in business priorities [5][18] - Movement from investment-oriented strategies to consumer-oriented approaches, reflecting changing market dynamics [5][18] - The rise of "one-person companies" exemplified by DeepSeek, showcasing a trend towards leaner operational models [5][18] Market Dynamics - The shift from local economies to global branding, indicating a need for companies to expand their market reach [5][18] - The evolution from traditional fixed asset economies to surrounding and virtual economies, suggesting a transformation in value creation [5][18] - The transition from following strategies to disruptive innovation, highlighting the need for companies to embrace creativity and change [5][18] Leadership and Management - The importance of recognizing new leadership dynamics and market structures, as illustrated by the characters in "Journey to the West" [19][20] - The necessity for entrepreneurs to build effective teams and strategies to navigate crises and market challenges [22][24] - The role of key figures in guiding organizations through transitions, emphasizing the need for strong leadership [25]

Investment Rating - The report suggests focusing on sectors driven by domestic AI development such as servers and IDC, as well as sectors like servers and optical modules driven by overseas AI development [5] Core Insights - Marvell reported revenue of $2.075 billion for the quarter, exceeding market expectations and guidance by $15 million, and announced the acquisition of Celestial AI for approximately $3.25 billion to enhance its position in the CPO field [1][50] - AWS launched the AI training chip Trainium 3 with a computing power of 2.52 PFLOPS FP8 at its annual cloud computing event, indicating strong growth in AI infrastructure [1] - Credo's Q2 FY2026 revenue reached $268 million, a year-over-year increase of 272.1%, driven by growth in its core AEC and IC businesses [1] - DeepSeek introduced two AI models, achieving performance levels close to GPT-5, indicating advancements in domestic AI capabilities [1][47] - ByteDance's Doubao team launched the "Doubao Phone Assistant," with the first device selling out quickly, showcasing the potential for AI applications in consumer electronics [1] Summary by Sections Communication Sector - The communication sector shows a steady upward trend, with significant investments in cloud and IDC businesses compensating for pressures in traditional telecom services [14] Server Sector - The server index decreased by 1.22% this week, but AWS's announcement of a significant expansion in AI/HPC data centers suggests ongoing demand growth [2][7] Optical Modules - The optical module index increased by 4.67%, supported by Marvell's strong quarterly performance and strategic acquisition [2][7] IDC Sector - The IDC index rose by 0.50%, with DeepSeek's new AI models expected to drive demand in data centers [3][8] Core Data Updates - Telecom business revenue reached 1.467 trillion yuan in the first ten months of 2025, a year-over-year increase of 0.9% [4][15] - The export value of optical modules decreased by 27.6% year-over-year in October, attributed to domestic companies establishing overseas factories [31] Market Trends - The communication sector's performance this week ranked second among all industries, with notable gains in specific companies [39][42]

3) DeepSeek Sparse Attention (DSA)DeepSeek’s new V3.2 model introduces DeepSeek Sparse Attention (DSA), which brings complexity down from O(L²) to O(Lk), where k is fixed.How it works:A lightweight Lightning Indexer scores which tokens actually matter for each query.Small number of heads, runs in FP8, computationally cheap.Then a selection mechanism retrieves only the top-k key-value entries.The key insight is that only 2048 tokens get selected per query, regardless of context length.So the expensive attent ...

Group 1 - Huawei's founder Ren Zhengfei emphasized that the future will see an excess of computing power, stating that building numerous large models is a correct exploration [2][3] - Ren highlighted that AI's true value lies in its application rather than invention, asserting that AI could contribute significantly to various industries, such as coal washing and steel production [2][3] - SoftBank's Masayoshi Son expressed regret over selling Nvidia shares, indicating that the sale was driven by the need for capital to invest in AI projects like OpenAI [3] Group 2 - Google CEO Sundar Pichai called for a national AI regulatory framework in the U.S. to avoid regulatory chaos and maintain competitive advantage [4] - Elon Musk predicted that AI could resolve the U.S. debt crisis within three years, suggesting that productivity gains from AI will soon outpace inflation [4] - The gap between China and the U.S. in AI capabilities has reportedly narrowed, with fewer doubts expressed about this disparity in the current market [4] Group 3 - Nvidia's CFO refuted claims of an AI bubble, stating that the current phase is an early stage of transitioning to AI-required data center infrastructure [5] - Nvidia's CEO Jensen Huang noted that AI will not directly eliminate jobs but will create new, unconventional roles [6] - UBS reported that the likelihood of an AI bubble in China is low, citing limited financing and a pragmatic approach to AI investments by leading internet companies [7] Group 4 - AMD's CEO Lisa Su downplayed concerns about an AI bubble, asserting that the demand for chips will continue to grow as AI technology develops [8] - Morgan Stanley projected explosive growth in Google's TPU production, raising estimates significantly for the coming years [9][10] - SemiAnalysis indicated that Google's advantage lies in its AI infrastructure rather than chip performance, suggesting a competitive threat to Nvidia [11] Group 5 - DeepSeek announced the release of two official model versions, aiming to enhance reasoning capabilities and support various applications [12] - The first domestic GPU company, Moore Threads, saw a significant stock price increase upon its debut, raising substantial funds for AI chip development [13] - HSBC partnered with Mistral AI to integrate generative AI tools across its operations, enhancing efficiency and customer service [13] Group 6 - The Ministry of Industry and Information Technology reported that China's software business revenue reached 125.1 billion yuan, reflecting a 13.2% year-on-year growth [16] - The first AI crop "genetic scientist" in China is set to launch globally next year, showcasing advancements in agricultural research [17] - Alibaba released an updated image generation model, Qwen-Image, which has been integrated into its app for user access [18]