计算机 | 证券研究报告 — 行业周报 2026 年 2 月 28 日 黄仁勋称代理式 AI 拐点已至； 朱雀三号有望年内复用飞行 计算机行业"一周解码" 英伟达 2026 财年第四季度财报营收与指引均超预期，黄仁勋称代理式 AI 拐点已至， 并看好云厂商资本支出转化收益；蓝箭航天计划于 2026 年第二季度再次开展朱雀三 号火箭回收试验，若成功，将在第四季度尝试首次重复使用飞行。 支撑评级的要点 投资建议 ◼ 建议关注 AI Agent、云与基础设施及商业航天板块企业，包括汉得信息、用友网 络、赛意信息、鼎捷数智、润泽科技、博睿数据、星环科技、中科星图、星图测 控、佳缘科技、航天宏图、上海瀚讯、盛邦安全等。 评级面临的主要风险 ◼ 技术创新不及预期；政策推行不到位；下游需求景气度不稳定的风险。 相关研究报告 《科技巨头加速 AI 军备竞赛，资本整合与 战略重组塑造行业新格局》20260130 《持续把握商业航天机会》20260125 《AI 商业化加速落地，核心科技自主可控 需求再燃》20260119 中银国际证券股份有限公司 具备证券投资咨询业务资格 计算机 证券分析师：杨思睿 (8610)6622932 ...

Investment Rating - The report does not explicitly state an investment rating for the industry Core Insights - The report highlights a significant increase in demand for computing power driven by advancements in AI models, particularly in video generation and multi-modal applications, indicating a structural shift in the industry towards higher quality and more complex models [6][11][17] - The introduction of GLM-5 and its "Interleaved Thinking" mechanism represents a new paradigm where computing power is exchanged for intelligence, leading to increased computational demands for inference tasks [13][27] - The supply side is expected to transition from a state of scarcity to structural balance by 2026, with improvements in domestic chip performance and the approval of NVIDIA's H200 chips for the Chinese market [42][43] Summary by Sections Section 1: Supply and Demand Dynamics - Seedance 2.0 has shown a strong user demand with long wait times for video generation, indicating a critical need for computing resources [11][12] - The demand for computing power is expected to escalate as major internet companies continue to develop large-scale AI models, with a notable focus on multi-modal capabilities [17][18] Section 2: Rapid Release of Computing Demand - The competition among leading internet firms is intensifying, with significant investments in AI infrastructure and model development, leading to a sharp increase in inference computing demand [32][34] - The report predicts that by 2026, the AI application landscape will expand significantly, driving real-time inference computing consumption [6][33] Section 3: Supply Side Improvements - The approval of NVIDIA's H200 chips is expected to alleviate computing power shortages for major internet companies, enhancing model iteration speeds [42] - Domestic chip manufacturers are making strides in performance and ecosystem development, with several companies achieving significant advancements in their products [43][44] Section 4: Full-Chain Inflation in Domestic Computing Power - The report anticipates a "full-chain inflation" cycle in the computing power industry by 2026, with growth expected across various segments including AI data centers and cloud services [49] - Major tech companies are projected to increase their capital expenditures significantly, further driving demand for computing resources [51] Section 5: Related Companies - The report identifies several companies as relevant to the industry, including Dongyangguang, Cambrian, Haiguang Information, Wangsu Technology, and others [4][55]

AI Agent 从概念走向落地，OpenClaw 引爆自主执行需求， Tokens 从"对 话消耗"走向"任务消耗"，调用量高速增长。此前，问答式 AI 一轮对话消 耗几百到几千 Token，用户不再提问，Token 就不会再消耗。Agent 模式下， AI 做任务时会持续在后台跑流程。拿 OpenClaw 举例，其 Token 消耗粗略 看有三种模式：1）多轮自我修正。一次编程任务可能经历"写代码→运行 →报错→改→再跑"几十个回合，每个回合都是一次完整的模型调用。2） 上下文无限膨胀。为了让 Agent"记住"之前的操作，每次调用都要携带完 整对话历史。有用户实测，一个活跃会话的上下文很快膨胀到 23 万 Token 以上。3）工具链级联。Agent 处理一个"帮我整理邮件并创建待办"的任 务，可能触发 5-10 次 API 调用，每次都背着完整上下文。较直观的数字是： 7×24 小时跑 OpenClaw，全量使用 Claude API，月成本在 800 到 1500 美 元之间。再看全球头部 AI 模型 API 聚合平台 OpenRouter 数据：2 月 9 日当 周，平台处理了 13 万亿 Tok ...

#这是一场势必不亚于10年前移动支付的大变革——最确定的方向：算力云服务！ 云厂提价背后的需求逻辑正在 强化，国内AI算力基础设施的主升或许才刚开始。 投资关键看含GPU量，国产的有边际变化，含nv的更稀缺，继续首推【协创、宏景】 #Agent使用走向平权但不普惠（都要花钱了！）。 智谱全网寻找"算力合伙人"并因运营和算力紧缺致歉。 延续周末观点，继续强CALL算力云服务 新边际变化：D.S.近期发布预期；D.S. V4首次破例优先向国产芯片开放提前访问权（REUTERS）；中国AI模型 的API调用量首次超过美国，四款大模型霸榜全球前五（OpenRouter）；华为云码道agent；再一轮的算力紧缺以及 预期的涨价潮；预期XXClaw批量出现。 #这是一场势必不亚于10年前移动支付的大变革——最确定的方向：算力云服务 延续周末观点，继续强CALL算力云服务 新边际变化：D.S.近期发布预期；D.S. V4首次破例优先向国产芯片开放提前访问权（REUTERS）；中国AI模型 的API调用量首次超过美国，四款大模型霸榜全球前五（OpenRouter）；华为云码道agent；再一轮的算力紧缺以及 预期的涨价潮；预 ...

Market Performance - The public REITs secondary market experienced a downward trend this week, with overall trading activity declining [1] - The CSI REITs total return index fell by 1.08% this week, underperforming the CSI 300 index [1][2] - As of February 27, the index closed at 1035.78 points, with a cumulative increase of 2.57% year-to-date [2] REITs Performance - Out of 79 listed public REITs, 17 saw an increase, with the top performers being data center, highway, consumer, and logistics REITs [2] - The best-performing REIT this week was the Southern Runze Technology Data Center REIT, which rose by 3.48% [2] - Conversely, 62 REITs declined, with 41 experiencing a drop of over 1%, the largest being Huatai Nanjing Jianye REIT, which fell by 6.34% [3] New REITs Filings - Four new public REITs were filed this week, indicating potential market expansion [4] - The Guangfa New Town Wuyue Commercial Real Estate REIT was officially filed, with a project scale of 616 million yuan [4] - The Jiashi Shoukai Commercial Real Estate REIT was also filed, aiming to revitalize existing assets and lower debt ratios [4] Additional Filings - The Guojin Jiazhe New Energy REIT was filed, focusing on wind power projects as underlying assets [5] - The Jianxin Tianjin Lingang Port REIT was filed, marking the first port real estate REIT in the country [6] - The Shanghai and Shenzhen stock exchanges issued the first round of review inquiries for five commercial real estate REIT projects this week [6]

润泽科技涨1.99%，报104.2元/股，股价再创新高，总市值突破1702.98亿元，成交额达64.26亿元。（AI 生成） ...

Core Insights - The Chinese large model market has seen a significant surge, with token usage reaching 41.2 trillion, surpassing the U.S. models for the first time [1][2] - Major Chinese models dominate the top five in usage, indicating a shift in the competitive landscape [1][2] - The market is bifurcating, with established players like BAT focusing on integrating models into existing services, while new entrants like Kimi and MiniMax are expanding their developer ecosystems [1][2] Token Usage and Developer Insights - The 41.2 trillion tokens are primarily driven by global developers, with U.S. developers accounting for 47.17% of usage compared to only 6.01% from China [2] - The surge in token usage reflects real demand from developers who are willing to invest in models that deliver performance and cost efficiency [2][6] - MiniMax M2.5 and Kimi K2.5 are highlighted for their competitive performance and cost advantages, with MiniMax achieving the highest usage in coding and search tasks [2][3] Cost Efficiency and Performance - Chinese models are significantly cheaper, costing only 1/10 to 1/20 of their U.S. counterparts, which is reshaping the economic calculations for developers [3][4] - The cost structure of models like MiniMax and Kimi allows for substantial savings in computational expenses, making them attractive options for developers [3][4] - The introduction of the "Mixture of Experts" (MoE) architecture has optimized engineering efficiency, contributing to lower costs [3] Demand Dynamics and Token Consumption - The emergence of agent-based scenarios has changed the token consumption logic, leading to exponential increases in token usage for complex tasks [5][6] - Over 70% of token consumption comes from large internet companies and professional developers, indicating a strong demand for these models in production environments [6] Business Model Evolution - The industry is shifting towards a "Results-as-a-Service" (RaaS) model, where payment is based on outcomes rather than token usage [8][9] - This transition requires a rethinking of pricing structures, moving from token-based to results-based billing, which aligns better with client expectations [9][10] - The challenge remains in accurately attributing results to AI contributions, complicating the implementation of this new model [18][19] Market Trends and Future Outlook - The current landscape shows a growing willingness among businesses to pay for quantifiable results, driven by changes in procurement processes [16][20] - The financial sustainability of new players is under scrutiny, as they face high computational costs that can exceed revenue [8][26] - The ability to successfully implement results-based pricing will be crucial for the survival and growth of these new entrants in the market [26][27]

Group 1 - ByteDance has released the AI video generation model Seedance 2.0, which is referred to as "the strongest video generation model currently available" [1] - There are significant challenges regarding computing power, with queue times reaching several hours and renewal prices nearly doubling [1] - CITIC Securities predicts a surge in domestic large model token usage during the 2026 Spring Festival, with domestic models dominating the top three in global token usage [1] - The explosive growth in token usage reflects an exponential increase in AI inference demand, with domestic computing power expected to gradually dominate the infrastructure layer due to cost advantages and an improving ecosystem [1] - Investment opportunities are suggested in areas benefiting from enhanced supernode interconnectivity, including optical communication, high-speed line modules, switching chips, switches, and IDC [1] - As of February 27, 2026, the CSI Cloud Computing and Big Data Theme Index (930851) rose by 2.30%, with notable increases in constituent stocks such as Yuntian Lifei (up 20.00%) and Yunsai Zhili (up 10.01%) [1] Group 2 - The CSI Cloud Computing and Big Data Theme Index closely tracks the performance of 50 listed companies involved in cloud computing services, big data services, and related hardware [2] - As of January 30, 2026, the top ten weighted stocks in the index include iFlytek, Kingsoft Office, and others, accounting for a total of 48.47% of the index [2]

Core Viewpoint - The rise of domestic computing power chains is significantly impacting the cloud computing sector, with the cloud computing ETF (159890) experiencing notable gains and increased investor interest [1]. Group 1: Market Performance - The cloud computing ETF (159890) opened with a rise of over 1% and is currently up by 0.74%, indicating strong performance among its constituent stocks [1]. - Notable stocks include Tuowei Information, which hit the daily limit, and Yuntian Lifeng, which surged by 13%, while several others like Wangsu Science and Technology and Runhe Software rose by over 6% [1]. - The fund has seen a net inflow of over 66 million yuan in a single day, with a cumulative net subscription of approximately 58.9 million yuan over the past five days, breaking previous records [1]. Group 2: Industry Developments - DeepSeek is set to release its V4 "Sea Lion Lite" version, which supports 1 million tokens and features native multimodal processing capabilities, prioritizing domestic chip suppliers for optimization [3]. - This shift marks a significant transition for the domestic computing power industry from being merely "usable" to achieving "scale commercialization," reinforcing the logic of domestic computing power [3]. Group 3: Demand Insights - According to a report from CITIC Securities, the token usage of domestic large models has surged, with the top three models being domestic, reflecting an exponential growth in AI inference demand [4]. - The report indicates that the domestic computing power sector, benefiting from cost advantages and an improving ecosystem, is poised to dominate the infrastructure layer [4]. - China International Bank also noted significant changes in the large model industry around the Spring Festival, suggesting a new wave of growth potential for the computing power supply chain [4]. Group 4: ETF Composition - The cloud computing ETF (159890) tracks the CSI Cloud Computing and Big Data Theme Index, with holdings that balance AI computing power (41%) and AI applications (32%) [5]. - The top ten holdings include leading software and hardware companies such as iFlytek, Kingsoft Office, and Inspur Information, which are expected to benefit from the AI infrastructure wave [5][6]. - The current trend led by DeepSeek in reshaping the domestic computing power ecosystem is likely to drive the industry from "passive adaptation" to "active definition," presenting collaborative development opportunities in computing power and AI applications [6].

| 北方华创 | -4.19 | -5.23亿元 | 电子 | | --- | --- | --- | --- | | 赛微电子 | -6.19 | -4.95亿元 | 电子 | | 通富微电 | -2.14 | -4.93亿元 | 电子 | | 特变电工 | -1.5 | -4.78亿元 | 电力设备 | | 海光信息 | -3.79 | -4.18亿元 | 电子 | | 高德红外 | -4.43 | -3.98亿元 | 国 防军_ | | 长飞光纤 | -2.25 | -3.97亿元 | 通信 | | 方正科技 | -7.29 | -3.95亿元 | 电子 | | 英维克 | -4.39 | -3.68亿元 | 机械设备 | *数据仅供参考，不构成投资建议 | 股票名称 | 涨跌幅 (%) | 主力资金流向 | 所属行 | | --- | --- | --- | --- | | 新易盛 | -5.95 | -22.38亿元 | 通信 | | 工业富联 | -4.68 | -17.03亿元 | 电子 | | 中际旭创 | -5.51 | -16.59亿元 | 通信 | | 胜宏科技 | -4.71 | -9 ...