随着中央经济工作会议为2026年经济工作定调，"跨年行情"与"春季躁动"的预期在券商研报中升温。 机构普遍认为，"扩内需、反内卷"的政策红利将引领市场走出通缩路径，而AI算力、商业航天等科技主 线与具备全球竞争力的"出海"品种，则构成了进攻方向。同时，资金对低波动、高股息资产的偏好也值 得关注，市场风格切换或成新看点。 中信证券：配置上要寻求交集，即海外敞口为基底、内需积极变化也会产生催化的品种 国金证券：中央经济工作会议定调扩内需、反内卷，走出通缩路径明晰 本周国内对于基本面预期影响最大的事件便是中央经济工作会议的召开，我们认为在中央经济工作会议 定调扩内需和注重反内卷的背景下，未来国内走出通缩的路径已经明晰： 外需品种是今年已经被充分证实和演绎过的线索，但延续到明年只是既有事实的延续，超预期的空间有 限，赚业绩而不是估值的钱；内需品种目前景气上升的品种有限，且未来的政策以及对于内需的刺激效 果是或有事件，并且今年刚经历过从过高预期到回归现实的修正，但投资者的预期足够低、持仓足够 低，一旦超预期可以享有估值弹性。 从配置角度，最大的交集就是我们此前一直提示的资源与传统制造业领域中国在全球有份额优势的行业 对 ...

证券研究报告·策略报告·策略周评 （2）算力硬件领域，国内厂商聚焦全栈体系突破。算力硬件领域本周以国 内厂商的技术突破为核心，竞争焦点从单一算力供给升级为"芯片-服务器 -平台"的全栈体系构建。国内端，联想 12 月 10 日正式推出"AI 工厂" 解决方案，同步发布万全异构智算平台 4.0 与问天 WA8080aG5 算力服务 器——前者通过异构算力调度技术实现大模型预训练时间缩短 35%，后者 支持单 GPU 超 1000 瓦功耗，为未来更高性能模型训练预留升级空间，成 功破解国内算力异构化整合难题。同日，华为云发布"盘古算力调度系统"， 可实现跨区域算力资源的智能分配，使算力利用率提升至 85%以上。这些 布局不仅直接拉动通信设备、高规格数据中心的需求，更推动算力基础设 施从"资源型"向"生产型"转型。 （3）应用端多场景突破，AI 升级为产业核心引擎。应用端本周呈现多场 景突破态势，AI 技术从工具属性向产业核心引擎升级。内容创作领域，快 手可灵在 12 月 9 日-12 日期间保持日更节奏，先后上线数字人 2.0、图片 O1 与视频 O1 模型，通过标准化 API 接口构建起从脚本生成到视频输出的 ...

Core Insights - The AI industry is rapidly evolving, with a consensus forming that by 2026, the performance iteration of large models will stabilize, allowing companies to focus on productization of large models [1] - Lenovo has launched the "AI Factory" solution and new high-end AI servers, aiming to inject strong momentum into China's AI industry for high-quality development [3][10] Industry Developments - Lenovo's "AI Factory" solution aims to transform traditional data centers into efficient AI factories, utilizing customer data as raw materials to produce intelligent models and services [10][11] - The launch of Lenovo's "万全异构智算平台4.0" addresses the diverse computing power needs in China, enabling efficient collaboration among different brands of CPUs and GPUs [6][12] Technological Innovations - The new platform has achieved significant performance improvements, including a 35% reduction in pre-training time and a 50% reduction in post-training time, along with a 60% increase in bandwidth utilization for inference scenarios [6] - Lenovo's WA8080a G5 server features a 10U chassis design to support GPUs with power consumption exceeding 1000 watts, addressing the rapid evolution of GPU technology [9] Strategic Positioning - Lenovo's approach emphasizes a complete ecosystem, integrating hardware and software to overcome the challenges of heterogeneous computing power in China [12] - The company aims to provide a comprehensive service system that addresses various client needs, from data governance to AI production management [11]

Core Insights - The article discusses the transformative impact of AI on various industries, emphasizing the shift towards a new paradigm characterized by "large models + large computing power + large data" as the foundation for AI 2.0 development [1][9] - It highlights the rapid growth of China's AI server market, which is expected to more than double by mid-2025 compared to the same period in 2024, driven by increasing demand [3][9] - The article outlines the evolution of large models, focusing on their deepening, broadening, and lengthening capabilities, which are essential for advancing AI technology [11][12] Industry Trends - The AI infrastructure development is entering a three-phase cycle: short-term acceleration, medium-term expansion, and long-term structural upgrades, necessitating a shift from merely stacking hardware to providing systematic services [10][12] - The penetration rate of AI in core products has increased from 5% in 2018 to an estimated 60% by next year, indicating a significant market evolution [12] - The global AI competition landscape is changing, with companies needing to balance between international and domestic GPU resources amid geopolitical tensions [5][12] Company Developments - Lenovo has introduced the "AI Factory" solution, aiming to streamline AI application development and deployment processes, transforming complex tasks into a standardized, efficient production line [15][16] - The Lenovo Wanquan Heterogeneous Intelligent Computing Platform 4.0 has been upgraded to support various computing scenarios, achieving a throughput of over 12,000 Tokens/s for local model deployment [16] - Lenovo's approach emphasizes the importance of a robust foundational support system to facilitate the transition from traditional computing centers to AI factories, helping businesses reduce initial barriers and operational costs [16]

在2025年，中国AI产业正经历从狂热的"百模大战"向冷静的"应用落地"转型的关键周期。12月5日，在 北京举办的2025联想异构智算产业联盟高峰论坛上，联想集团试图通过发布"AI工厂"解决方案、核武级 服务器以及全新的行业标准。在算力不再稀缺的当下，如何解决效率与商业价值的错配？联想中国基础 设施业务群战略管理总监黄山与服务器事业部总经理周韬在接受媒体专访时，不仅详解了硬件架构的激 进演进，更直言不讳地指出了当前大模型商业化面临的挑战与机遇。 算力中心的终结与"AI工厂"的崛起 他进一步强调，这种差距将导致竞争层面的不对等："这些业务如果有生态形成了联动以后，我只做数 据中心，我就被形成降维打击了。怎么叫降维打击？当赋能用户的时候，给AI工厂只给计算中心。首 先方案的完整性不如别人。第二个，因为你没有跟其他的业务板块协同，你的效率肯定不如一个完整的 AI工厂，最终的效率一定是不行的。" 联想的商业创新在于"补足"客户的能力短板。黄山解释道："我们的客户有的有自己的数据，有自己的 设计能力，但是没有生产能力。有的客户也有生产能力，但没有生产管理能力，有的客户出来的应用如 何做迭代，他没有成周期升级维护的能力。 ...