2025 年，AI ASIC（专用集成电路）在整个芯片行业中扮演着越来越重要的角 色。尤其是在 AI 芯片市场，ASIC 因其在特定场景和任务需求上的产品匹配优 势，表现出显著的业绩增长。ASIC 芯片通常为特定目的或算法专门定制，能够 在特定任务上发挥强大的算力。例如，在深度学习算法中，ASIC 可以更高效地 进行矩阵运算和数据处理。 相比之下，GPU 虽然综合能力较强，但功耗和成本 也较高。因此，在满足不同下游客户需求时，ASIC 和 GPU 各有优劣。随着 AI 整体向推理侧发展及应用功能的丰富化，相对单一任务集中解决方案的 ASIC 将成为重要选择。此外，在设计路径上，不仅要提升算力，还需兼顾高速通信 能力，这是影响集群整体性能的重要因素。 海外互联网厂商自研 ASIC 进度加快，与博通、Marvell 等紧密合作，普 遍采用 sky up 或机柜连接方式，光模块需求随网络升级而提升，对国 内市场有重要启示。 博通 AI 业务收入占比快速提升，核心竞争力在于领先的 IP 储备、深耕 网络层多年以及超大规模芯片落地能力，Marvell 在 AWS 产品中逐步成 为主要对接厂商，数据中心业务市场空间预期上 ...

[Table_Rank] 投资评级：增持（维持） 报告日期： 2025-08-27 | [Table_BaseData] 收盘价（元） | 17.15 | | --- | --- | | 近 12 个月最高/最低（元） | 22.27/12.57 | | 总股本（百万股） | 428 | | 流通股本（百万股） | 239 | | 流通股比例（%） | 55.95 | | 总市值（亿元） | 73 | | 流通市值（亿元） | 41 | [Table_Chart] 公司价格与沪深 300 走势比较 -20% 0% 20% 40% 60% 80% 美瑞新材 沪深300 [Table_StockNameRptType] 美瑞新材（300848） 公司点评 [Table_Author] 分析师：王强峰 执业证书号：S0010522110002 电话：13621792701 邮箱：wangqf@hazq.com 分析师：潘宁馨 执业证书号：S0010524070002 电话：13816562460 邮箱：pannx@hazq.com [Table_CompanyReport] 相关报告 1.业绩符合预期，聚氨酯一体化 ...

点击蓝字，关注我们 小摩核心观点： 市场对 AI 资本支出见顶的担忧无需过度， 2026-2027 年增长确定性强 近一周，市场再度担忧 2026 年 AI 资本支出（以下简称 "AI capex"）或因当前支出规模高企而见顶，主要顾虑包 括：一是 AI 部署缺乏投资回报率（ROI），二是企业持续为 AI capex 提供资金的能力存疑。叠加英伟达 （NVDA）、超威半导体（AMD）对华 GPU 出货短期缺乏增长空间，投资者近期对 AI 板块态度趋于谨慎。 但摩根大通持不同观点，核心依据包括四点： 生成式 AI（Gen-AI）未现过热迹象： 头部 4 家云服务提供商（CSP）仍可通过增长的经营现金流支撑资本支 出，且有充足能力将增长的资本支出预算延续至 2027 年以后； 新增投资主体持续入场： 大型私人 AI 实验室（如 OpenAI、xAI、Anthropic）及主权基金纷纷加入，进一步推动 AI 投资； AI 应用场景大幅扩容： 消费级 AI 应用增长显著，而企业级 AI 虽进展较慢（推理模型落地仅 6 个月，尚处早期 阶段），但代币消耗量已实现强劲增长； （头部 4 家 CSP 2027 年资本支出情 ...

智通财经APP获悉，小摩发布研报称，市场担忧2026年AI资本支出(capex)可能见顶，投资者对AI板块趋于谨慎。但该行 通过四大核心论证不同观点：生成式AI未现过热现象、新增投资主体持续入场、AI应用场景大幅扩容，以及中国市场的 潜在需求释放。摩根大通预测，2026年AI capex增速至少达20%，若推理模型渗透率提升，2027年有望进一步增长。 小摩主要观点如下： 近一周，市场再度担忧 2026 年 AI 资本支出(以下简称 "AI capex")或因当前支出规模高企而见顶，主要顾虑包括：一是 AI 部署缺乏投资回报率(ROI)，二是企业持续为 AI capex 提供资金的能力存疑。叠加英伟达(NVDA)、超威半导体(AMD) 对华 GPU 出货短期缺乏增长空间，投资者近期对 AI 板块态度趋于谨慎。 但摩根大通持不同观点，核心依据包括四点： 生成式 AI(Gen-AI)未现过热迹象：头部 4 家云服务提供商(CSP)仍可通过增长的经营现金流支撑资本支出，且有充足能 力将增长的资本支出预算延续至 2027 年以后; 新增投资主体持续入场：大型私人 AI 实验室(如 OpenAI、xAI、Anthrop ...

小摩核心观点： 市场对 AI 资本支出见顶的担忧无需过度，2026-2027 年增长确定性强 AI 应用场景大幅扩容：消费级 AI 应用增长显著，而企业级 AI 虽进展较慢（推理模型落地仅 6 个月，尚处早期阶 段），但代币消耗量已实现强劲增长； 但摩根大通持不同观点，核心依据包括四点： 生成式 AI（Gen-AI）未现过热迹象：头部 4 家云服务提供商（CSP）仍可通过增长的经营现金流支撑资本支出，且有 充足能力将增长的资本支出预算延续至 2027 年以后； 新增投资主体持续入场：大型私人 AI 实验室（如 OpenAI、xAI、Anthropic）及主权基金纷纷加入，进一步推动 AI 投 资； 此外，非 AI 领域的涨价讨论已逐步扩散（如 BT 基板、T 型玻璃、部分 ABF 基板、电源管理器件、传统 DRAM、高 端封测等），这或成为科技板块每股收益（EPS）下一轮上调的核心驱动因素。 头部 4 家 CSP 资金充足，2027 年资本支出仍将增长 中国市场潜力待释放：中国 AI capex 尚未真正启动，将成为需求增长的另一重要驱动力。 近一周，市场再度担忧 2026 年 AI 资本支出（以下简称 "A ...

Core Viewpoint - The recent advancements in large model technology have renewed interest in AI-specific chips, particularly Google's TPU, which has evolved significantly since its deployment in 2015, now reaching its 7th generation [1][9]. Group 1: TPU Overview - TPU is a specialized chip designed by Google to enhance the speed of machine learning model inference and training, focusing on executing mathematical operations efficiently [9]. - The architecture of TPU allows it to perform matrix multiplication efficiently, which constitutes a significant portion of computations in deep learning models [14][31]. Group 2: TinyTPU Project - The TinyTPU project was initiated by engineers from Western University in Canada to create an open-source ML inference and training chip, motivated by the lack of a complete open-source codebase for such accelerators [5][7]. - The project emphasizes a hands-on approach to learning hardware design and deep learning principles, avoiding reliance on AI tools for coding [6]. Group 3: Hardware Design Insights - The project team established a design philosophy of exploring unconventional ideas before consulting external resources, leading to the re-invention of many key mechanisms used in TPU [6]. - The hardware design process involves understanding clock cycles, using Verilog for hardware description, and implementing a systolic array architecture for efficient matrix multiplication [10][12][26]. Group 4: Training and Inference Mechanisms - The TinyTPU architecture allows for continuous inference by utilizing a double buffering mechanism, which enables the loading of new weights while processing current computations [61][64]. - The training process leverages the same architecture as inference, with additional modules for gradient calculation and weight updates, allowing for efficient training of neural networks [71][118]. Group 5: Control and Instruction Set - The control unit of TinyTPU employs a custom instruction set architecture (ISA) to manage control signals and data flow, enhancing the efficiency of operations [68][117]. - The ISA has evolved to include 94 bits, ensuring that all necessary control flags and data fields are accounted for without compromising performance [117].

智通财经APP获悉，中信建投证券发布研究报告称，8月22日，上证指数涨1.45%站上3800点，算力、芯片股集体爆发。近期国产算力板块迎来密集催化， 8月13日，腾讯业绩会表示公司推理芯片供应渠道侧具备多种选择；8月21日，DeepSeek更新模型版本至DeepSeek-V3.1，预计将支持FP8精度以及国产芯 片；近期，基于华为昇腾芯片的服务器产品陆续在政府、金融、运营商等行业落地大单。国产算力芯片迎来国产替代窗口期。考虑到英伟达新品迎来大幅 性能升级，并面向中国市场禁售，国产算力芯片发展刻不容缓。当前已经涌现出一大批国产算力芯片厂商，昇腾、寒武纪相继推出自研AI芯片，海光信 息的DCU也逐渐打出知名度，其他配套环节的国产化进程也正在加速推进。 国产人工智能芯片发展趋势及展望 海外龙头占据垄断地位，AI加速芯片市场呈现"一超多强"态势。 数据中心CPU市场上，英特尔份额有所下降但仍保持较大领先优势，AMD持续抢占份额势头正盛。AI加速计算芯片市场上，英伟达凭借硬件优势和软件 生态一家独大，在训练、推理端均占据领先地位。根据IDC数据，2024年国内AI加速计算芯片市场中，英伟达出货份额达70%，华为昇腾出货 ...

时令 发自 凹非寺 量子位 | 公众号 QbitAI 大模型耗电惊人，舆论一浪高过一浪。 现在，谷歌用硬核数据强势还击。 一次Gemini查询仅能耗0.24wh， 相当于微波炉运行1秒 ，排放0.03 g CO₂e，甚至 比人放一次屁还少 ，同时消耗约5滴水。 （文中数据均表示中位数） 谷歌首席科学家Jeff Dean公开表示： 这个数值明显低于公众普遍预期的水平。 若Gemini只考虑TPU和GPU的消耗，一次查询只需消耗0.10wh，排放0.02 g CO₂e，耗水量约0.12ml。 但这种计算方式仅考虑了机器在运行中的能耗，而忽略了一些关键因素。 并且像谷歌这样的大模型厂商，一直专注研究如何降低AI消耗。 一年间，通过模型优化和额外清洁能源的使用，Gemini的能耗降至原来的1/33，碳排放降至1/44，同时提供了更高质量的响应。 下面让我们一起看看谷歌是如何衡量和减少AI能源消耗的？ 如何衡量AI消耗能源 谷歌首先指出，目前许多关于AI能源消耗的计算反映的都是理论效率，而非大规模运行下的实际效率。 | | Active Al Accelerators | | | | | | --- | --- | - ...

Core Viewpoint - Nvidia (NASDAQ: NVDA) is facing significant challenges ahead, including high valuation and unprecedented competition, despite its current success in the AI and robotics sectors [1][2]. Valuation Concerns - Nvidia's stock price has doubled since its low in April, indicating it is currently overbought and may be due for a correction [2]. - The stock is trading well above its 50-day moving average (approximately $163) and 200-day moving average (approximately $136) [2]. - Analysts express skepticism about the stock's future performance given its high price-to-earnings ratio exceeding 40 [9]. Competitive Landscape - Major tech companies are developing their own AI chips, posing a significant threat to Nvidia: - Google has developed TPU and plans to release new AI chips in 2024 and 2025 [5]. - Microsoft is working on Azure Maia AI Accelerator and Azure Cobalt CPU for its data centers [6]. - Amazon has its own AI chips, "Trainium" and "Inferentia," for training and inference tasks [7]. - Emerging companies like Cerebras Systems, Tenstorrent, and Graphcore are also entering the AI chip market, potentially disrupting Nvidia's dominance [8]. Market Influence - Nvidia currently accounts for approximately 8% of the S&P 500 index, which raises concerns about its influence on the overall market [8]. - The historical context shows that no chip company has maintained a leading position in the S&P 500 for an extended period, indicating potential volatility for Nvidia [15]. Historical Performance and Future Outlook - Historical trends suggest that past performance does not guarantee future results, and Nvidia's current valuation may not be sustainable [9][10]. - The company is attempting to diversify its business, which could stabilize its position in the long term, similar to Microsoft [14]. - Analysts suggest that the current high stock price may present an ideal selling opportunity before the upcoming earnings report [14].

Investment Rating - The report maintains a "Buy" investment rating for the company [1] Core Views - The industry is under pressure, and the company is waiting for downstream demand recovery [5] - The company's revenue for the first half of 2025 was 12.137 billion yuan, a year-on-year decrease of 11.7%, with a net profit attributable to shareholders of 983 million yuan, down 35.23% year-on-year [4][5] - The chemical fiber segment showed a gross margin increase of 3.68 percentage points despite a revenue decline [5] - Future growth is expected as the economy recovers and downstream demand is released, particularly with advancements in domestic technology [7][8] Summary by Sections Financial Performance - In Q2 2025, the company achieved a revenue of 5.823 billion yuan, down 17.84% year-on-year and 7.78% quarter-on-quarter, with a net profit of 479 million yuan, down 42.61% year-on-year [4][5] - The revenue breakdown shows a decline in various segments: chemical fibers down 9.43%, new chemical materials down 8.82%, and basic chemical products down 15.01% [5] Profitability and Cost Structure - The gross margin for the chemical fiber segment was 18.65%, an increase of 3.68 percentage points year-on-year, while the gross margin for basic chemical products was 4.37%, a decrease of 11.08 percentage points [5][6] - The financial expense ratio increased mainly due to a decrease in interest income [6] Future Outlook - The company forecasts net profits for 2025-2027 to be 2.101 billion, 3.057 billion, and 3.588 billion yuan respectively, with corresponding P/E ratios of 18.3, 12.6, and 10.7 [9] - The report anticipates that the integration of the polyurethane industry will lead to profit growth in the long term [9]