英伟达正迎来越来越多AI芯片"挑战者"。 当地时间12月2日，亚马逊发布新一代AI芯片Trainium3，并透露了其正在研发中的 Trainium4。 不久前，谷歌官宣了Gemini 3大模型，因为采用的是自研芯片，加之市场评测其模型完胜OpenAI的GPT 系列，一度造成英伟达股价大跌，甚至被视为其王者地位被动摇。 在A股市场，炒作的重点则落到了算力互联。由于谷歌已经官宣第七代自研TPU芯片Ironwood将在未来 几周内正式发布，而该芯片采用的是光电路交换机（OCS）互联架构，OCS作为光网络下一代战场得到 TPU冲击波 确认，相关产业链获暴涨。 此轮谷歌TPU芯片对英伟达GPU芯片的冲击，要从Gemini 3正式发布说起。 当地时间11月18日，谷歌母公司Alphabet推出新一代多模态大模型Gemini 3。Gemini 3 相比前代在长上 下文、多模态理解与复杂推理上的能力跃升，把谷歌的TPUs推到了聚光灯下。 谷歌称，Ironwood在训练和推理方面的性能比上一代Trillium芯片提升了4倍多，是该公司旗下迄今为止 性能最强、能效最高的定制芯片。 具体而言，在运算能力上，Ironwood芯片的峰 ...

英伟达的挑战者众多。 英伟达正迎来越来越多AI芯片"挑战者"。 当地时间12月2日，亚马逊发布新一代AI芯片Trainium3，并透露了其正在研发中的Trainium4。 不久前，谷歌官宣了Gemini3大模型，因为采用的是自研芯片，加之市场评测其模型完胜OpenAI的GPT系列，一度造成英伟达股价大跌，甚至 被视为其王者地位被动摇。 在A股市场，炒作的重点则落到了算力互联。由于谷歌已经官宣第七代自研TPU芯片Ironwood将在未来几周内正式发布，而该芯片采用的是光 电路交换机(OCS)互联架构，OCS作为光网络下一代战场得到确认，相关产业链获暴涨。 TPU冲击波 在他看来，TPU和GPU是串行路线，而非并行路线：GPU的发展未来是模型稳定，模型稳定后一定会走向TPU。"也许未来，GPU的通用市场 会逐渐收窄至高校、研究院等科研场景，更多生产场景可能会收敛至垂类模型，彼时将是TPU的天下。但这一未来并不会来得太快——模型迭 代非常快，只有像谷歌一样又做模型又做运营的企业才适合做TPU。" 随着模型的收敛、TPU的迭代，谷歌TPU也从自用走向了外销。目前，Anthropic宣布将部署多达100万个谷歌的TPU ...

Core Viewpoint - The report from Zhongyuan Securities indicates that a series of AI industry catalytic events will occur in 2026, strengthening the leading position of top optical module manufacturers due to their technological, customer, and scale advantages. The current valuation of the communication industry index is below the ten-year average, and the industry maintains a "stronger than the market" investment rating based on performance growth expectations and valuation levels [1][2]. Summary by Sections Review of 2025 - In early 2025, the DeepSeek large model boosted market sentiment, and the three major operators completed the deployment of DeepSeek computing power private networks, enhancing their cloud service capabilities. Domestic cloud manufacturers provided positive capital expenditure guidance, leading to an increase in industry valuations. However, from February to April, the industry index experienced significant fluctuations due to concerns over U.S. tariff policies and future demand for optical modules. By mid-April, the easing of tariff policies and validation of AI computing power demand led to a gradual recovery in the industry index and valuations. In late July, North American cloud manufacturers raised their capital expenditure guidance, further catalyzing the industry. Since September, leading manufacturers faced short-term performance fatigue due to product iterations and customer structure adjustments, raising concerns about unclear downstream business models [2]. Outlook for 2026 - A series of AI industry catalytic events are expected, including the mass production of NVIDIA's next-generation Rubin GPU, the release of Google's new large model Gemini, and clear capital expenditure guidance from cloud manufacturers. AI smartphones equipped with large models are anticipated to become personalized smart assistants, potentially driving the next wave of smartphone upgrades. The development of key 6G technologies by telecom operators is expected to accelerate revenue growth from AI computing power. The report is optimistic about the high industry prosperity and strong growth potential of optical modules, optical devices, optical chips, and the increasing penetration of AI smartphones, as well as the stable operations of quality dividend assets in telecom operators [3][4]. Capital Expenditure Outlook for Leading Cloud Manufacturers - The demand for 800G is increasing, and the industry is transitioning from 800G to 1.6T technology. Leading optical module manufacturers are expected to further highlight their advantages due to technological leadership, stable customer relationships, and scalable delivery capabilities. The development of AI is driving the construction of large data centers, benefiting optical device manufacturers. The long R&D and expansion cycles for optical chips create high barriers in technology, talent, customer validation, and capital, leading to a persistent supply-demand gap for certain optical chips. The increasing demand for domestic controllable solutions is expected to translate into performance for domestic computing power. Recommended companies to watch include: NewEase, Huagong Technology, Guangxun Technology, Yuanjie Technology, Shijia Photon, and Taicheng Light [4]. AI Smartphones and Market Trends - Generative AI smartphones are set to provide users with new interactive experiences, multimodal content generation capabilities, personalized services, and innovative application ecosystems. The continuous improvement of edge AI computing power and large model capabilities is expected to further increase the market penetration of AI smartphones. Innovations and upgrades in AI smartphones are likely to lead to higher average selling prices and improved profit margins. The growth in edge AI shipments will drive sustained growth in core product lines of consumer electronics components [5]. Telecom Operators' Performance - The three major telecom operators are considered quality dividend assets with high dividend yield potential, offering cash dividends twice a year. The quality of traditional business revenue is improving, and a decrease in capital expenditure is expected to lower future depreciation and amortization costs, maintaining stable operations. Additionally, telecom operators are likely to leverage their advantages in data centers, big data, and network infrastructure to reconstruct business models with the help of AI. Investment recommendations include focusing on the optical module, optical device, and optical chip sectors, as well as AI smartphone and telecom operator sectors [6].

Core Insights - Huawei's Vice Chairman Xu Zhijun presented a clear development roadmap for its Ascend AI chips, targeting a million-card cluster by 2028, indicating a significant leap in computing power [1][2] - The introduction of the "Lingqu" all-optical interconnection protocol is expected to revolutionize the field of computing interconnectivity, enhancing data transfer speeds and overall efficiency [2] Group 1: Huawei's AI Chip Development - Huawei plans to rapidly iterate its Ascend AI chips, with the Ascend 950 series launching in 2026, followed by the Ascend 960 in 2027, and the Ascend 970 in 2028, aiming for a tenfold increase in overall computing power [1][2] - The Ascend 960 chip will achieve a comprehensive doubling of key metrics such as computing power and memory bandwidth, while the Ascend 970 will further enhance interconnect bandwidth and memory bandwidth [1] Group 2: Super Node Technology - The Atlas 950 super node, built on the new "Lingqu" protocol, is anticipated to maintain a leading position in global computing power for the coming years [2] - The establishment of large-scale computing clusters, potentially reaching 500,000 to 1,000,000 cards, will significantly reduce data transfer delays and enhance computational efficiency [2] Group 3: Implications for AI Development - With robust computing infrastructure, China's AI models are expected to experience explosive growth, positioning them among the global leaders in AI technology innovation and application [2]

Group 1 - Changfei Fiber Optic Cable (06869) saw its stock price rise over 20%, reaching a historical high of 44.86 HKD during trading, with a current price of 43.96 HKD and a trading volume of 1.22 billion HKD [1] - NVIDIA announced the launch of NVIDIA® Spectrum-XGS Ethernet, a technology that enables the combination of multiple distributed data centers into a one-gigawatt AI super factory, addressing the limitations of existing commercial Ethernet infrastructure [1] - Open-source securities highlighted that the scale-across technology could become the "third pillar" of AI computing, following scale-up and scale-out, and emphasizes the importance of optical interconnect and related industry opportunities [1] Group 2 - Zhongtai Securities reported that hollow-core fiber, which uses air as a transmission medium, has lower latency and loss, making it a key component for future AI data center interconnects [2] - The first commercial use of hollow-core fiber by China Mobile is expected in July 2025, with Changfei Fiber being the sole bidder at an actual bid price of approximately 36,000 RMB per core kilometer [2] - The report emphasizes the leading capacity level of Changfei Fiber in hollow-core technology and its successful bids in related projects [2]

Group 1 - The core idea emphasizes the importance of data elements and computing power in building a unified national market, highlighting their inherent advantages such as liquidity, high reusability, intangibility, and non-exclusivity [1][2] - The "East Data West Computing" initiative aims to guide computing power demand from the eastern regions to the western regions, facilitating cross-regional allocation of computing resources and enhancing the digital economy [3] - The construction of a unified data and computing market is essential to address the mismatch between supply and demand, as evidenced by the varying utilization rates of computing resources in different regions [1][3] Group 2 - The integration of computing power aims to break down regional and industry barriers, requiring collaboration among government departments, telecom operators, and internet companies to create a cohesive network [2] - The development of digital twin factories and smart agricultural applications demonstrates the practical benefits of computing power integration, leading to reduced operational costs and improved efficiency [4] - The establishment of a unified market is just the first step; addressing challenges such as data security and user privacy during cross-regional data sharing is crucial for long-term success [4][5]

Group 1 - The core idea emphasizes the importance of data elements and computing power in building a unified national market, highlighting their inherent advantages such as liquidity, high reusability, intangibility, and non-exclusivity [1][2] - The "East Data West Computing" initiative aims to guide computing demand from the eastern regions to the western regions, facilitating cross-regional allocation of computing resources and enhancing the digital economy [3] - The construction of a unified data and computing market is essential to address the mismatch between supply and demand, as evidenced by the varying utilization rates of computing resources in different regions [1][3] Group 2 - The integration of computing power is seen as a means to promote high-quality industry development and regional coordination, with Ningxia serving as a national hub for integrated computing networks [3] - Successful examples of computing power integration include the establishment of a "5G + AI digital twin factory" in Guizhou and smart greenhouse applications in Gansu, showcasing the benefits of reduced operational costs and improved efficiency [4] - The article stresses the need for a long-term vision in building a unified market, focusing on solving issues related to data security and user privacy during cross-regional data sharing [4]

Group 1 - Data elements possess characteristics such as liquidity, high reusability, intangibility, and non-exclusivity, which provide a natural advantage in supporting the construction of a unified national market [1][2] - The "East Data West Computing" initiative aims to orderly guide computing demand from the eastern regions to the western regions, facilitating cross-regional allocation of computing resources and enhancing the digital economy [3] - The construction of a unified data market and computing market is essential to address the mismatch between supply and demand in computing resources, as evidenced by the varying utilization rates in different regions [1][3] Group 2 - The interconnection of computing resources is a key topic for high-quality industry development and regional coordinated development, with Ningxia serving as a national hub for integrated computing networks [3] - The collaboration between companies and telecommunications firms in creating "5G + AI digital twin factories" demonstrates the practical benefits of computing interconnection, such as reduced operational costs and improved efficiency [4] - The establishment of a unified national market is seen as a valuable opportunity for regions to enhance their digital economy and reshape their development advantages [3][4]

Core Insights - The article emphasizes that computing power encompasses not only computational capabilities but also supporting network transport and data storage capabilities [1] - For computing power to be effective, hardware such as servers, chips, and storage devices, along with software including operating systems, cloud computing platforms, and AI frameworks, as well as network connectivity, must all be "online" simultaneously [1] - As of the end of 2024, China's total computing power is projected to reach 280 EFLOPS (exaflops) [1] - The computing power market is experiencing a structural imbalance characterized by both "surplus" and "shortage" [1]

Core Viewpoint - The rapid development of the artificial intelligence industry is driving the demand for computing power resources, leading to increased attention on computing power interconnectivity. The "Computing Power Interconnectivity Action Plan" aims to achieve nationwide public computing power standardization and interconnectivity by 2028 [1][2]. Group 1: Computing Power and Its Importance - Computing power has become an essential infrastructure, comparable to water, electricity, and gas in the digital age. It is a core element driving the development of artificial intelligence alongside data and algorithms [1]. - High-quality computing power is crucial for various applications, including smart home devices, AI model generation, intelligent manufacturing, scientific computing, and smart healthcare [1]. Group 2: Understanding Computing Power Interconnectivity - Computing power encompasses not only computational capabilities but also the supporting network, data storage, hardware, software, and connectivity. All three aspects must be operational simultaneously for effective utilization [2]. - By the end of 2024, China's total computing power is projected to reach 280 EFLOPS (exaflops). However, there is a structural imbalance in the computing power market, with both surplus and shortage issues [2]. Group 3: Goals and Strategies for Computing Power Interconnectivity - The goal of computing power interconnectivity is to enhance the flow of computing resources across the country, addressing issues such as low utilization rates and mismatched supply and demand in the industry [3]. - Building a unified and efficient computing network involves improving foundational technologies, increasing data transmission bandwidth, establishing a unified computing power identification system, and ensuring data and privacy security during transactions [2][3]. Group 4: Economic and Social Implications - A well-connected computing network can significantly enhance the efficiency of data transmission and resource allocation, facilitating the sharing of government data and improving logistics, data flow, and capital flow [3]. - In the era of interconnected devices, accelerating interconnectivity across various sectors and elements is not optional but a necessity for economic and social development [3].