据悉，华为推出的CloudMatrix 384超节点集群，通过新型高速互联总线实现384张卡互联成为一个超级云服务器，单集群算力规模最高可达300P（超业界 67%），训练效率提升至41%以上，并显著降低模型训练成本。 记者 李奕璇 一是整个大模型的发展趋势，从小模型到大模型，从短序列到长序列，单模态到多模态，而且还是从快应答到深度思考。首先是需求的驱动力作为驱动去开 发一个新的形态。另外，依托国内半导体的支撑，华为从2019年就进入了实体清单，华为很多的芯片只能是设计出来，但是生产加工不出来。所以这也导致 华为现在芯片单颗的算力和国外先进的厂商还有一定的差距，这个差距在短时间内还无法突破。为了弥补当前的短板，华为推出了384超节点的架构，通过 系统化的功能传递，来满足当前国内国外的需求。 另外，华为384节点有16个机柜，其中12个机柜是算力节点，4个是网络机柜，通过4个网络机柜把12个机柜里面进行全光互联，实现了384颗芯片成为了一体 机，真正成为一个服务器，保证了可以实现卡和卡之间的时延达到了150纳秒左右。而为什么要做384颗，是因为英伟达提出72个节点，华为一定要超越对 手，推出384节点这个集群是 ...

《人民日报》采访原文中，任正非原话的确是"芯片问题其实没必要担心，用叠加和集群等方法，计算结果上与最先进水平是相当的。"不过，这一个回 答， 实际续接的是"您对人工智能的未来背景怎么看"。也就是说，任正非的本意并没有那么"高调"，而是说中国在AI领域有芯片可用，正因为在单芯片 性能上落后美国芯片不少，所以昇腾才要剑走偏锋，通过堆叠（istack）和集群（Cluster）去力求达到英伟达那种性能。 其实只需要把原来的标题改成"AI芯片的问题没必要担心"，仅仅一词之差，整体的语义就回归了。实际原文中任正非整体态度很谦虚，但看到这样的报 道，或许有些人就会认为任正非很"狂"。此类的断章取义其实还有很多…… "被误解"的任正非 实际上，如果经常关注国外报道，也许你会看到这样的字眼——华为创始人认为中国无需担心美国芯片制裁。 但如果细细品味一下《人民日报》的原文报道，就会发现，这种说法完全是境外媒体的捏造和臆测。任正非这一篇采访非常振奋人心，不仅提振了国内芯 片行业的士气，也非常谦逊，接受批评，正视不足，是一篇值得细细品味的报道。 中共中央办公厅国务院办公厅关于进 保障和改善民生 着力解决群众急难愁盼的意见 (2025 ...

Investment Rating - The report suggests focusing on domestic AI development-driven sectors such as servers, IDC, switches, and connectors, as well as overseas AI development-driven sectors like servers and optical modules [4] Core Insights - The U.S. Department of Commerce upgraded chip controls against China, banning Huawei's Ascend chips and limiting U.S. chips for AI training in China, which is expected to benefit domestic chip suppliers [1][6] - NVIDIA's AI chip shipments fell short of expectations, leading to a projected slight decline in capital expenditures from domestic internet companies in the short term, but this will accelerate the domestic replacement of AI chips [1][5] - The establishment of an AI chip R&D center in Shanghai by NVIDIA and a significant order for 18,000 GB300 servers from Saudi Arabia highlight the global strategic layout and increasing demand for high-speed optical modules and liquid cooling technology [1][6] Summary by Sections Communication Sector Overview - The telecom operators achieved a total revenue of 446.9 billion yuan in the first quarter of 2025, a year-on-year increase of 0.7%, with a total telecom business volume growth of 7.7% [3][12] - The optical module exports saw a growth of 19.6% year-on-year in March, with a cumulative growth of 3.18% from January to March 2025 [28][30] Server Sector - The server index experienced a slight decline of 2.51% this week, primarily due to delays in NVIDIA's GB200 deliveries and lower-than-expected capital expenditures from companies like Alibaba [5][6] - The report anticipates that the domestic AI chip supply will benefit from the U.S. chip control upgrades, which may lead to a long-term increase in AI-driven business growth [5][6] Optical Module Sector - The optical module industry is showing signs of recovery, with leading companies like NewEase and Zhongji Xuchuang reporting significant revenue growth of 179% and 264% respectively in Q1 2025 [7] - The overall revenue of the optical module sector increased by 48% year-on-year, driven by AI computing demand and cost reduction measures [7] IDC Sector - The IDC industry is expected to grow driven by increasing AI computing demand and supportive policies, with resources concentrating on companies that meet energy consumption standards and have sufficient project reserves [2]

Core Insights - The demand for AI inference is driving significant growth in the performance of domestic AI chip companies, with notable improvements in financial results for key players like Cambrian and Haiguang [1][2][11] - Cambrian has ended six consecutive years of losses, achieving profitability in Q4 2024 and continuing this trend into Q1 2025, with a substantial revenue increase [2][6] - Haiguang's performance remains stable, with a strong revenue growth driven by innovations in general computing products [11][14] Cambrian's Performance - Cambrian reported a revenue of 1.174 billion yuan in 2024, a year-on-year increase of 65.56%, and a net loss of 452 million yuan, which is a 46.69% reduction in losses compared to the previous year [2] - In Q1 2025, Cambrian's revenue surged to 1.111 billion yuan, a 40-fold increase year-on-year, although the quarter-on-quarter growth showed a decline [6] - The company achieved a net profit of 355 million yuan in Q1 2025, marking a 256.39% increase compared to a loss of 227 million yuan in the same quarter of the previous year [7] Haiguang's Stability - Haiguang achieved a revenue of 9.162 billion yuan in 2024, a 52.4% increase year-on-year, with a net profit of 1.931 billion yuan, up 52.87% [11] - In Q1 2025, Haiguang's revenue was 2.4 billion yuan, a 50.76% increase year-on-year, and a net profit of 506 million yuan, reflecting a 75.33% increase [11][14] - The growth is attributed to continuous technological innovation and an expanding market share in general computing products [11][15] Challenges in Specialized Markets - Companies focusing on specialized markets, such as Jingjiawei and Loongson Zhongke, are facing performance pressures, with Jingjiawei's revenue declining by 34.62% in 2024 [17] - Loongson Zhongke reported a slight revenue decrease of 0.28% in 2024, with a net loss of 625 million yuan, indicating challenges in building its ecosystem [20] - Both companies are attempting to transition from specialized to more general market applications to enhance growth [17][20] Industry Trends and Innovations - The rapid adaptation of AI chip manufacturers to DeepSeek's models is seen as a significant step towards internationalization and enhancing the domestic AI chip ecosystem [22][23] - The introduction of integrated computing products is gaining traction, with over a hundred models available, although there are concerns about their performance consistency [24] - Innovations in computing efficiency and transmission rates are being pursued, exemplified by Huawei's new CloudMatrix architecture, which significantly enhances resource interconnect bandwidth [26]

Investment Rating - The report suggests focusing on sectors driven by AI development, including servers, IDC, switches, switch chips, optical modules, and liquid cooling, as well as overseas markets for servers, switches, and optical modules [4]. Core Insights - China Unicom's Q1 2025 financial report shows growth driven by cloud and IDC data center businesses, with cloud revenue at 19.72 billion yuan, up 18.1% year-on-year, and IDC revenue at 7.22 billion yuan, up 8.8% year-on-year [1]. - Huawei's CloudMatrix 384 super node cluster, featuring 384 Ascend chips, is expected to accelerate domestic AI chip replacement and boost demand for optical modules [1]. - The demand for inference computing continues to grow as tech companies accelerate their investments in multimodal interaction and AI agents, exemplified by Tencent's AI assistant "Yuanbao" and OpenAI's new models [1][7]. - The three major telecom operators are steadily advancing cloud computing and computing power services, driving rapid growth in AI computing demand [3]. Summary by Sections Subsector Insights - **Servers**: 2025 is projected to be a year of significant growth in inference demand, with a shift from centralized cloud services to mass terminals requiring cost-effective dedicated chips. Domestic ASIC chip manufacturers like ZTE and Unisoc are expected to benefit [2][6]. - **Optical Modules**: The CloudMatrix system's architecture will significantly increase the demand for optical modules, with major domestic suppliers like LightSpeed Technology and Shijia Photonics poised for growth [2][10]. - **IDC**: The three major operators are increasing investments in computing power, with companies like DataPort and Runjian benefiting from the growing demand for data centers [3][13]. Key Data Updates - The telecom sector's main business revenue reached 295 billion yuan in the first two months of 2025, a year-on-year increase of 0.9% [17]. - The number of fixed broadband access users reached 675 million, with gigabit users accounting for 31.7% [18]. - The mobile internet traffic reached 55.39 billion GB, up 13.6% year-on-year [21]. Market Trends - The report indicates a steady upward trend in the telecom sector, with operators, optical modules, and servers showing robust growth [14][17]. - The IDC index has seen a decline of 2.17% this week, reflecting market fluctuations [13]. Company News - OpenAI has launched the GPT-4.1 series models, enhancing performance and cost efficiency, which is expected to drive demand for AI computing power [55]. - Major cloud providers like Microsoft, Google, and Amazon are significantly increasing their capital expenditures for AI and cloud services, indicating strong future demand for computing infrastructure [7].