证券研究报告 | 行业定期报告 2025 年 11 月 09 日 持续关注国产算力及 AI 应用 计算机周观察 20251109 TMT 及中小盘/计算机 英伟达对华销售芯片再遭限制；中科曙光发布全球首个单机柜级 640 卡超节点 scaleX640，保障 10 万卡超级大规模集群扩展部署。国内软件公司基本面触底 回升，机构持仓和预期低位，建议持续关注国产算力及 AI 应用。 ❑ 风险提示：AI 技术发展不及预期；行业竞争加剧。 推荐（维持） 行业规模 | | | 占比% | | --- | --- | --- | | 股票家数（只） | 285 | 5.5 | | 总市值（十亿元） | 4338.2 | 4.1 | | 流通市值（十亿元） | 3901.8 | 4.0 | 行业指数 % 1m 6m 12m 绝对表现 -1.3 16.7 20.9 相对表现 -2.1 -5.4 8.1 资料来源：公司数据、招商证券 -30 -20 -10 0 10 20 30 Nov/24 Mar/25 Jun/25 Oct/25 (%) 计算机 沪深300 相关报告 1、《海外持续加大 AI 投入，Q3 计 算机行业拐点确立 ...

scaleX640超节点采用"一拖二"高密架构设计，不仅实现了单机柜640卡超高速总线互连，构建大规模、 高带宽、低时延的超节点通信域，而且可以通过双scaleX640超节点组成千卡级计算单元。相比业界同 类产品，不仅综合算力性能实现倍增，同时单机柜算力密度提升20倍；相比传统方案，可实现MoE万亿 参数大模型训练推理场景30%-40%的性能提升。 通过30天+长稳运行可靠性测试验证，可保障10万卡级超大规模集群扩展部署。根据阿里巴巴和腾讯的 公告数据，2025Q2,阿里巴巴和腾讯资本支出合计577.8亿元，同比增长177.5%，将推动国内AI算力市场 持续高速增长。同时，海外AI算力芯片在国内市场受阻，国产算力芯片迎来发展机遇，有望加速崛 起。 面对国产化进程中高端芯片算力不足、标准缺失、软硬件割裂、协作成本高等痛点，公司于今年9月联 合20余家产业链企业，共同发布"AI计算开放架构"，并开放多项关键技术能力，旨在降低AI集群研发门 槛，避免重复投入，推动产业从"单点突围"走向"生态共进"。scaleX640超节点正是这一理念的有力实 践。它采用AI计算开放架构，在硬件层面支持多品牌加速卡，软件层面兼容主流 ...

证券研究报告 | 行业周报 gszqdatemark 2025 11 08 年 月 日 计算机 AI 投资继续加码 北美巨头资本开支全面上调，AI 落地持续加速。近期各科技巨头发布三 季度业绩，资本开支指引均上调： 1）谷歌云收入在第三季度增长了 34%，达到 152 亿美元，企业端的 Gemini Enterprise 人工智能平台构建的智能 Agent 用户已超过 200 万。第七代 TPU 即将上市，Anthropic 计划获取一百万个 TPU。Gemini App 目前拥有 超过 6.5 亿月活跃用户，查询量较第二季度增长了 3 倍。Veo 3 已生成超 过 2.3 亿个视频；今年晚些时候将发布 Gemini 3。搜索领域 AI 模式日活 跃用户已超过 7500 万。7 月谷歌宣布所有平台每月处理的 token 数量达 到 980 万亿。现在每月处理的 token 数量超过 1.3 千万亿。谷歌预计 2025 年资本支出将在 910 亿美元至 930 亿美元之间，高于之前估计的 850 亿 美元，展望 2026 年预计资本支出将大幅增长。 2）Meta2025 年三季报将本财年资本支出的预测调整至 7 ...

2.【现代汽车集团欧洲技术中心将启用价值1.5亿欧元新设施】11月6日，现代汽车集团宣布，将在位 于德国吕塞尔斯海姆的现代汽车欧洲技术中心（HMETC）启用一座价值1.5亿欧元的新设施，这标志 着集团在致力于欧洲创新和卓越工程方面迈出重要一步。投资项目包括全球最大的半消声室之一、气 候测试实验室及电动汽车充电实验室等。（界面新闻） 1.【中科曙光：发布全球首个640卡超节点scaleX640】 11月6日，2025世界互联网大会乌镇峰会 期间，中科曙光正式发布全球首个单机柜级640卡超节点scaleX640。该超节点采用"一拖二"高密架 构设计，实现单机柜640卡超高速总线互连，构建大规模、高带宽、低时延的超节点通信域。它还能 通过双scaleX640超节点组成千卡级计算单元，且通过30天+长稳运行可靠性测试验证，可保障10万 卡级超大规模集群扩展部署。（新浪财经） 欢迎加入 睿兽分析会员 ，解锁 AI、汽车、智能制造 等相关 行业日报、图谱和报告 等。 50 Ft 国家高新 八时一年资计 技术企业 CBJF+ 线 创投人物 投融资及收并购事件 热门产业 创新机会 TEE 详 2万+LP数据 全生命周期 10 ...

来源：北京日报客户端 据了解，超节点是AI算力时代的一种创新技术，通过高速互联技术将成千上万个算力卡连接、整合为 一个高效处理大规模任务的超级计算单元。当前人工智能正驱动全球产业格局重塑，而规模化落地的主 要瓶颈，在于算力基础设施能否跨越从"技术可行"到"商业可用"的鸿沟。面对诸多挑战，技术团队通过 系统工程创新，实现算、存、网、电、冷的一体化紧耦合系统设计，采用超高速正交架构、超高密度刀 片、浸没相变液冷、高压直流供电等创新技术，实现了软硬件协同的全局优化，打造出高性能、高效率 和高可靠性的超节点方案。 scaleX640超节点采用"一拖二"高密架构设计，不仅实现了单机柜640卡超高速总线互连，构建大规模、 高带宽、低时延的超节点通信域，而且可以通过双scaleX640超节点组成千卡级计算单元，不仅综合算 力性能实现倍增，同时单机柜算力密度提升20倍。相比传统方案，它能实现MoE（混合专家模型）万亿 参数大模型训练推理场景30%到40%的性能提升，并保障10万卡级超大规模集群扩展部署。 在全球信息技术产业生态体系加速调整的大背景下，中国智算发展除坚持系统级创新之外，还应加强生 态协同。面对国产化进程中高端芯 ...

11月6日算力板块反弹，国产算力更为强势，科创芯片ETF收涨4.73%，集成电路ETF收涨3.82%，芯片ETF收涨3.71%，半导体设备ETF收涨3.48%。 | 代码 | 名称 | 涨跌幅 | | --- | --- | --- | | 589100 | 科创芯片ETF国泰 | 4.73% | | 159546 | 集成电路ETF | 3.82% | | 512760 | 芯片ETF | 3.71% | | 159516 | 半导体设备ETF | 3.48% | 资料来源：wind 消息面上，11月6日，2025世界互联网大会乌镇峰会期间，中科曙光正式发布全球首个单机柜级640卡超节点scaleX640。该超节点采用"一拖二"高密架构 设计，实现单机柜640卡超高速总线互连，构建大规模、高带宽、低时延的超节点通信域。它还能通过双scaleX640超节点组成千卡级计算单元，且通过30 天+长稳运行可靠性测试验证，可保障10万卡级超大规模集群扩展部署。 资料来源：IT之家 存储方面，由于AI的虹吸效应，海外大厂将自身产能转向DDR5及HBM，短期内导致供需错配，DDR4等产能退出过快导致价格上涨。据Trend ...

国芯网[原:中国半导体论坛] 振兴国产半导体产业！ 值得注意的是，就在该消息发酵之时，国产超节点曝出重大突破。 11 月 6 日，中科曙光正 式发布全球首个单机柜级 640 卡超节点 scaleX640 ，该产品基于全球领先的开放系统硬件 架构打造，可以完全对标英伟达 CUDA 生态，提供可快速迁移的超强算力支持。 数据显示， scaleX640 超节点采用"一拖二"高密架构设计，实现了单机柜 640 卡超高速总 线互连，构建大规模、高带宽、低时延的超节点通信域，并通过双 scaleX640 超节点组成千 卡级计算单元，是当前算力规模最大、集成度最高的计算集群。 相比业界同类产品， scaleX640 综合算力性能实现倍增，同时单机柜算力密度提升 20 倍； 相比传统方案，可实现 MoE 万亿参数大模型训练推理场景 30%-40% 的性能提升。目前产 品已通过 30 天 + 长稳运行可靠性测试验证，可保障 10 万卡级超大规模集群扩展部署。 近日，国内数据中心及其芯片使用要求的小作文连续刷屏，甚至一度引发国产芯片板块迅速 走高。业内认为，这可能意味着国产算力市场进一步扩容，如 NV 卡等进口芯片在华市场占 比 ...

Market Overview - The market showed strong fluctuations throughout the day, with the Shanghai Composite Index rising above 4000 points. The total trading volume in the Shanghai and Shenzhen markets reached 2.06 trillion yuan, an increase of 182.9 billion yuan compared to the previous trading day. The Shanghai Index rose by 0.97%, the Shenzhen Component Index by 1.73%, and the ChiNext Index by 1.84% [1] Sector Performance - The computing power sector rebounded, with domestic computing power leading the gains. The Science and Technology Innovation Chip ETF rose by 4.73%, the Integrated Circuit ETF by 3.82%, the Chip ETF by 3.71%, and the Semiconductor Equipment ETF by 3.48% [2][3] AI and Storage Market Dynamics - Due to the AI demand, major overseas manufacturers are shifting their production capacity towards DDR5 and HBM, leading to a supply-demand mismatch and price increases for DDR4 and other products. From Q2 2025, mainstream storage product prices have been rising quarter-on-quarter, with NAND Flash prices also increasing. The price hikes are expected to continue due to sustained AI demand, potentially exacerbating the tight supply situation through 2026 [6][8] Electric Grid Sector Opportunities - The electric grid sector is poised for historical opportunities, driven by the increasing power demand from AI data centers. The U.S. is projected to face a power shortfall of approximately 73.2 GW from 2025 to 2030, which could rise to 201 GW if data center growth exceeds expectations. This situation may lead to new opportunities for domestic electric grid companies, particularly in the context of power exports [11][12]

Market Performance - The market showed a strong upward trend with a total trading volume of 2.06 trillion yuan, an increase of 182.9 billion yuan compared to the previous trading day. The Shanghai Composite Index rose by 0.97%, the Shenzhen Component Index increased by 1.73%, and the ChiNext Index gained 1.84% [1] - The computing power sector rebounded, with domestic computing power showing stronger performance. The Science and Technology Innovation Chip ETF rose by 4.73%, the Integrated Circuit ETF increased by 3.82%, the Chip ETF gained 3.71%, and the Semiconductor Equipment ETF rose by 3.48% [1][2] Factors Driving the Increase - Negative sentiment has partially dissipated. Following the National Day holiday, TSMC's CoWoS capacity adjustment and strong overseas computing power contributed to market recovery. Despite optimistic capital expenditure statements from cloud companies, the market's expectations were not fully met, leading to a noticeable recent pullback. The computing power sector rebounded as value became more apparent during this pullback [3] - Continuous price increases in storage products are driven by AI demand, leading to a supply-demand mismatch. The shift of major overseas manufacturers towards DDR5 and HBM has caused a rapid exit of DDR4 capacity, resulting in price hikes. The tightness in related capacities is expected to persist and may worsen through 2026. According to CFM, prices for storage products in servers, mobile devices, and PCs are expected to continue rising in Q4 [3][4] - On the news front, the launch of the world's first single-cabinet 640-card super node scaleX640 by Inspur on November 6 during the World Internet Conference indicates advancements in high-bandwidth, low-latency communication technology [3] Future Outlook - There remains growth potential in both domestic and overseas capital expenditures, with no immediate concerns regarding capital spending. Recent financial reports indicate that cloud companies maintain an optimistic outlook on capital expenditures. For instance, Google raised its 2025 capital expenditure forecast from $80 billion to $91-93 billion, while Meta increased its forecast to $70-72 billion [6] - The demand for AI and related technologies is on the rise, with major cloud companies showing strong cash flow growth driven by AI. Except for Amazon, the other three major companies have capital expenditures in 2025 that account for less than 75% of their operating cash flow, indicating robust financial health [6] Investment Recommendations - Investors are advised to consider both North American and domestic computing power sectors for potential investments. The overall pace of AI development is accelerating, with improving fundamentals and no immediate tariff risks. However, market volatility may persist due to increased profit-taking, suggesting a strategy of buying on dips [7] - For overseas computing power, it is recommended to focus on communication ETFs with high core content, while for domestic computing power, the Science and Technology Innovation Chip ETF is expected to benefit from the release of domestic GPUs and rising storage prices. The Semiconductor Equipment ETF is also highlighted for its improved fundamentals due to advanced process expansion and storage capacity increases [7]