2026年2月26日早盘，通信设备板块涨幅居前，CPO概念活跃，截至10:20，国证通信指数(399389)强势 上涨1.93%，成分股中天科技、亨通光电、通鼎互联10cm涨停，烽火通信，长飞光纤等个股跟涨。 通信ETF嘉实(159695)紧密跟踪国证通信指数，一键布局光通信产业链。 MACD金叉信号形成，这些股涨势不错！ 2026春节期间，国产大模型token调用量井喷，截至2月22日此前一周的全球大模型token用量中，前三甲 均为国产大模型。中信证券认为，token的爆发式增长，本质上反映出AI推理需求的指数级扩容，而国 产算力凭借着成本优势及不断完善的生态，有望在基础设施层逐步占据主导。建议重点关注由超节点互 联密度提升带来的价值重估机遇，包括光通信、高速线模组、交换芯片及交换机、IDC等环节。 数据显示，截至2026年1月30日，国证通信指数前十大权重股分别为中际旭创、新易盛、中兴通讯、天 孚通信、中国移动、中国电信、中国联通、信维通信、亨通光电、中天科技，前十大权重股合计占比 54.9%。 量子通信迈入大规模网络部署新阶段，我国建成全球首个基于集成光量子芯片的大规模量子密钥分发网 络。该网络由北京 ...

随着有"地球最重要股票"称号的"AI芯片霸主"英伟达(NVDA.US)将在美东时间周三美股收盘后(北京时间周四晨间)公布季度业绩，一场关于AI算力投资主题 的"压力测试"也随之到来。聚焦于AI算力基础设施以及更广泛AI基建狂潮的全球投资者们正寻求证据，证明这家全球最高市值芯片巨头的利润正在紧密依托 高达6500亿美元至7000亿美元的美国四大科技巨头(hyperscalers)巨额AI资本支出预算大趋势而同步实现强劲增长预期。 与此同时，来自hyperscalers近期密集宣布将推出基于自研模式的性价比更高AI ASIC芯片的动向，也正显现出对英伟达在全球AI基建最核心领域——AI芯片 领域长期绝对主导地位构成风险的迹象。 占据标普500指数以及纳斯达克100指数高额权重(大约35%-40%)的所谓"七大科技巨头"，即"Magnificent Seven"(Mag 7)，它们包括：苹果、微软、谷歌、特 斯拉、英伟达、亚马逊以及Facebook母公司Meta Platforms，它们乃标普500指数屡创新高的核心推动力，也被华尔街顶级投资机构们视为在自互联网时代以 来最大技术变革背景下最有能力为投资者们带来巨额 ...

智通财经APP获悉，随着有"地球最重要股票"称号的"AI芯片霸主"英伟达(NVDA.US)将在美东时间周三美股收盘后(北 京时间周四晨间)公布季度业绩，一场关于AI算力投资主题的"压力测试"也随之到来。聚焦于AI算力基础设施以及更广 泛AI基建狂潮的全球投资者们正寻求证据，证明这家全球最高市值芯片巨头的利润正在紧密依托高达6500亿美元至 7000亿美元的美国四大科技巨头(hyperscalers)巨额AI资本支出预算大趋势而同步实现强劲增长预期。 与此同时，来自hyperscalers近期密集宣布将推出基于自研模式的性价比更高AI ASIC芯片的动向，也正显现出对英伟达 在全球AI基建最核心领域——AI芯片领域长期绝对主导地位构成风险的迹象。 在过去三年推动美国股市迈向超级牛市轨迹之后，在纳斯达克100指数以及标普500指数中占据高权重的英伟达股价在 2026年迄今仅上涨约2%，主要因Anthropic一系列AI代理产品引发的"AI末日叙事"重创软件股以及高估值的科技巨头 们，加之超大规模云厂商们加速自研性价比更高的替代AI ASIC芯片(比如TPU)并推动多供应商策略，叠加AMD等竞争 加剧。下图为英伟达 ...

英伟达此前曾宣布计划将预填充操作 ——AI 推理工作负载中内存需求相对较低的部分 —— 卸载到配备 GDDR7 的独立 GPURubinCPX 上。 这一设计策略可能基于预填充操作对内存带宽要求相对较低、无需 HBM 的评估。 根据我们的渠道核查，RubinCPX 预计将采用。 英伟达此前曾宣布计划将预填充操作 ——AI 推理工作负载中内存需求相对较低的部分 —— 卸载到配备 GDDR7 的独立 GPURubinCPX 上。 这一设计策略可能基于预填充操作对内存带宽要求相对较低、无需 HBM 的评估。 因此，RubinCPX 最终将内存规格改为 根据我们的渠道核查，RubinCPX 预计将采用。 因此，RubinCPX 最终将内存规格改为 HBM 具有重大意义。 这表明在实际生产环境中，即便是预填充操作也需要，源头信息加微WUXL7713而 GDDR7 最终无法提供足够的 性能效率。 若此潜在需求成为现实，。 由于英伟达尚未正式公布此规格变更，我们尚未将 RubinCPX 相关的 HBM 需求纳入 HBM 模型。 ...

Core Insights - Taalas, a startup founded two years ago with a team of 24, has launched a new chip, HC1, which achieves a peak inference speed of 17,000 tokens per second, significantly outperforming competitors like Cerebras at 2,000 tokens per second [1][3][5] - The HC1 chip reduces costs by 20 times and power consumption by 10 times compared to existing solutions, enabling real-time response speeds for large language models (LLMs) [1][3] - Taalas's innovative approach involves embedding the model directly onto the silicon chip, which allows for a drastic increase in performance and efficiency [3][6] Company Overview - Taalas was founded by a team of former AMD executives, including Ljubiša Bajić, who has a strong background in high-performance GPU design [11][13] - The company focuses on developing a new architecture specifically for AI inference and training, emphasizing layered design and lattice networks [11][13] Technology and Performance - The HC1 chip utilizes TSMC's N6 process technology, with a compact size of 815mm² and a typical power consumption of 250W per chip [5][6] - By adopting a structured ASIC design philosophy, HC1 can quickly produce specialized AI inference chips at a lower cost, reducing the production cycle from six months to two months [6][8] - The chip's architecture allows for the storage of models and weights directly on the chip, enhancing speed and efficiency while maintaining some flexibility for model updates [8][10] Market Position and Future Plans - Taalas has raised $200 million in funding and plans to release a second-generation variant of HC1 in the spring, which will integrate a medium-sized inference model [13] - The company aims to deploy HC2 in the winter, which will feature higher density and faster operation [13] - Despite the impressive speed of HC1, there are concerns regarding its depth of inference and potential obsolescence due to rapid model iteration cycles [15][17]

Core Viewpoint - Taalas aims to revolutionize AI inference by hard-coding AI model weights directly into chip transistors, eliminating software redundancies and simplifying device architecture, which addresses the memory-computation barrier faced by traditional GPUs and AI XPUs [2][6][10]. Company Overview - Taalas, founded two and a half years ago, has raised over $200 million through three rounds of venture financing and is based in Toronto, a hub for AI research and chip talent [3][4]. - The founders, including CEO Ljubisa Bajic, have extensive backgrounds in chip design and AI, with previous experience at companies like AMD and Tenstorrent [3][5]. Technology and Architecture - Taalas combines ROM and SRAM to create a high-density architecture for AI inference, allowing for the storage of model weights and execution of computations at high speeds [6][10]. - The current generation of Taalas chips can support up to 8 billion parameters, with plans for the next generation to support up to 20 billion parameters, significantly reducing the number of chips needed for large models [10][11]. Production and Cost Efficiency - The cost of training a model is approximately 100 times higher than the cost of customizing a Taalas chip, making it economically viable for companies to order custom accelerators for their models [11]. - Taalas has developed a "foundry-optimized workflow" with TSMC, allowing customers to convert model weights into deployable PCI-Express cards within two months [12]. Performance Metrics - Initial performance tests indicate that Taalas's HC1 chips demonstrate lower costs and latency compared to traditional GPU systems, with the potential to disrupt the AI inference market [17][19]. - The HC1 chip integrates 53 billion transistors and operates at a power consumption of approximately 200 watts per card, with a dual-socket server consuming around 2500 watts [12][13]. Future Developments - Taalas plans to release a hard-coded 20 billion parameter model by summer and aims to support multiple models through clusters of HC cards by the end of the year [13][19].

为推进该战略，公司已实施一系列具备高度协同性的落地方案：在美国关键算力枢纽达拉斯设立全资运营主体EcoHash Technology LLC；引入具备超大规模 分布式系统管理经验的核心技术团队，由前Zoom全球基础设施负责人领衔；其创新的 "即插即用"标准化算力模块已通过技术及商业模式可行性验证；同 时，公司主动优化资产负债表，通过削减加密货币相关风险敞口、置换流动性，为战略转型储备了必要的财务资源。 这一系列举措，共同构成了一个从顶层战略设计、组织与能力建设、商业化路径验证，到财务资源匹配的完整闭环，展现出灿谷向AI算力服务商转型的坚 定决心与清晰可期的执行路径。 战略卡位：精准切入AI推理蓝海 构筑分布式算力网络结构性壁垒 智通财经APP获悉，灿谷（CANG.US）于2月10日发布重大战略转向公告，正式确立以构建 "AI分布式算力网络"作为公司第二增长曲线的核心战略方向。此 举标志着其业务模式将实现从周期性显著的加密货币矿业，向具备高成长性的AI算力基础设施服务的系统性战略跃迁。 面对全球AI算力产业高集中度、高资本壁垒的竞争格局，灿谷采取了差异化的战略聚焦路径——锚定AI推理市场这一高增长细分赛道。相较 ...

Group 1 - Nvidia and Meta have announced a multi-year, multi-billion dollar chip procurement agreement, where Meta will purchase millions of Nvidia's latest AI acceleration chips, including the upcoming "Vera Rubin" series [1][3] - Meta's CEO Mark Zuckerberg stated that the company plans to invest up to $135 billion in AI infrastructure by 2026, nearly double its spending in 2025 [3] - Despite challenges in its own AI chip development, Meta continues to rely on Nvidia's established solutions due to technical challenges and deployment delays in its self-developed AI chip projects [3] Group 2 - The transaction highlights a shift in AI computing focus from "model training" to the "inference" phase, with Nvidia's CEO Jensen Huang adjusting the product strategy to sell CPUs as standalone products rather than integrated with GPUs [3] - The growing demand for efficient, low-latency inference capabilities is driving this change in hardware architecture, as noted by industry analysts [3]

Core Viewpoint - The article provides an overview of recent IPO activities in Hong Kong, the US, and China, highlighting key companies and their market positions in the AI and technology sectors. Group 1: Aixin Yuan Zhi - Aixin Yuan Zhi Semiconductor Co., Ltd. officially listed on the Hong Kong Stock Exchange on February 10, 2026, under the stock code "0600," becoming the first Chinese edge AI chip company to go public [2] - Since its establishment in 2019, Aixin Yuan Zhi has delivered over 165 million SoCs, with significant growth in sales of its edge computing and terminal computing SoCs, which increased by approximately 69% and 400% respectively in 2024 compared to 2023 [3] - By 2024, Aixin Yuan Zhi became the fifth largest supplier of visual edge AI inference chips globally, holding a market share of 24.1% in the mid-to-high-end segment [3][4] - The company's revenue grew from 0.50 billion yuan in 2022 to 4.73 billion yuan in 2024, with a compound annual growth rate of 206.8% [4] Group 2: Haizhi Technology - Haizhi Technology Group Co., Ltd. listed on the Hong Kong Stock Exchange on February 13, 2026, under the stock code "2706," becoming the first company to focus on AI graph computing technology to eliminate large model hallucinations [6] - The company has developed the Atlas graph solution and industry-level intelligent agents, achieving a market share of approximately 50% among AI intelligent agent providers in China [7] - Revenue from Haizhi Technology increased from 3.13 billion yuan in 2022 to 5.03 billion yuan in 2024, with a significant growth of 872.2% in revenue from the Atlas intelligent agent in 2024 [8] Group 3: Qunkong Technology - Qunkong Technology received a listing application notice from the China Securities Regulatory Commission on February 14, 2026, aiming to become the first "Hangzhou Six Little Dragons" company to complete an IPO [10] - The company focuses on spatial intelligence solutions and has developed the SpatialVerse platform, which integrates core capabilities in spatial reconstruction and generation [10][11] - Qunkong Technology's revenue reached 400 million yuan in the first half of 2025, with a gross margin of 82.1%, marking a significant improvement from 72.7% in 2022 [12] Group 4: Hairou Innovation - Hairou Innovation submitted its prospectus to the Hong Kong Stock Exchange on February 13, 2026, aiming for a main board listing [13] - The company has developed the HaiPick system, which revolutionizes warehouse automation, achieving a global market share increase from 24.2% in 2023 to 31.4% in 2024 [15] - By the end of September 2025, Hairou Innovation had signed contracts with over 800 clients, including more than 70 companies listed in the Fortune Global 500 [16] Group 5: Magnesium Health - Magnesium Health updated its prospectus for a main board listing on the Hong Kong Stock Exchange, reporting a revenue increase of 33.85% to 1.873 billion yuan in the first ten months of 2025 [18] - The company aims to transform China's medical payment system by linking patients, insurers, and pharmaceutical companies, with a focus on innovative drug and insurance solutions [20] - By October 31, 2025, Magnesium Health had served approximately 2 million patients and partnered with all top 20 insurance companies in China by premium income [22]

Core Insights - Nvidia has made significant progress in AI inference with its Blackwell architecture, achieving a milestone in "token economics" [1] - The company has implemented an "extreme hardware-software co-design" strategy, optimizing hardware efficiency for complex AI inference workloads, reducing the cost of token generation to one-tenth compared to the previous Hopper architecture [1] Industry Applications - Several inference service providers, including Baseten, DeepInfra, Fireworks AI, and Together AI, are utilizing the Blackwell platform to host open-source models [2] - These companies have successfully achieved cross-industry cost reductions by combining cutting-edge open-source intelligent models, Blackwell's hardware advantages, and their own optimized inference stacks [2] - For instance, Sentient Labs, focusing on multi-agent workflows, reported a cost efficiency improvement of 25% to 50% compared to the Hopper era, while companies in the gaming sector, like Latitude, have achieved lower latency and more reliable responses [2] Technical Specifications - The core of Blackwell's efficiency lies in its flagship system, the GB200 NVL72, which features a configuration of 72 interconnected chips and up to 30TB of high-speed shared memory [6][7] - This design is well-suited for the current mainstream "Mixture of Experts (MoE)" architecture, allowing for efficient splitting and parallel processing of token batches across multiple GPUs [6][7]