AI New Era: Cloud Vendors Expand Production Intensively, Domestic Demand Reigns

With the end of the scaling law for pre-training on public data, the global AI industry landscape is undergoing profound changes.

Cloud Service Providers (CSPs) are stepping into the spotlight, becoming key players in driving the full-link development of AI from data to models, from training to inference, and from cloud to edge.

On the 29th, Minsheng Securities' Fang Jing team released a deep report on the electronic industry titled "New Paradigm of AI: Cloud Vendors Lead + Domestic Demand is King," pointing out that whether it is overseas giants like Google and Amazon or domestic companies like ByteDance and Tencent, the cloud vendor giants have taken up the baton for the development of the AI industry. In this process, the industrial chain advantages of Chinese local enterprises are particularly prominent. For example, in fields such as PCB (Printed Circuit Board), copper cables, temperature control, and power supply, Chinese companies have accumulated significant advantages through years of deep cultivation.

With the vigorous expansion of local cloud vendors, the era of domestic demand being king is also approaching. Minsheng Securities believes that compared to the past, AI industry investment in the next 25 years will place greater emphasis on collaborative establishment, order fulfillment, and performance realization, with investment returns being more robust.

CSPs Lead a New Round of Computing Power Competition

The growth of large AI models has long relied on the "Scaling Law" of public data, which improves model performance by increasing the amount of data, computing power, and parameter scale.

However, as public data gradually becomes exhausted, this model is reaching its end. The concept of Scaling Law 2.0 has emerged—emphasizing the quality of data rather than quantity, with private domain data and high-precision manually labeled data becoming the focus of competition.

Cloud vendors that possess a large amount of user behavior data have a unique advantage necessary for training large models and are playing an increasingly important role in the AI ecosystem:

"As AI applications gradually land, AI will shift from the training end to the inference end. Self-developed ASICs by cloud vendors will be more cost-effective than general-purpose GPUs, and they will have high-quality traffic interfaces to empower the inference side; hardware is the most important carrier for the implementation of large models, and CSPs will serve as important providers of large models, fully empowering AIPC, AI phones, AI glasses, robots, etc. The business model of CSP + electronic brands (Apple, Huawei, Lenovo, etc.) will become clearer."

From public data to private domain data, from training to inference, from cloud to edge, the three major tools are all in the hands of CSPs, forming a perfect closed loop of training - inference - commercialization:

We believe that the discourse power of CSPs will further increase in the future, and the discourse power of AI development will be passed from computing power companies represented by NVIDIA and large model companies represented by OpenAI to major CSP giants.

The demand for computing power from AI large models is still growing exponentially. Minsheng Securities believes that the core contradiction in the current overseas computing power industry chain is the growth of cloud vendors' capital expenditures and the demand for computing power.

In the past, GPU companies represented by NVIDIA dominated the AI chip market, but in recent years, cloud vendors have accelerated the development of self-researched chips, moving from "outsourcing" to "self-research," and significantly expanding computing power infrastructure relying on large cloud computing platforms.

Minsheng Securities believes that the accelerated self-research of cloud vendors' acceleration cards will become the most core source of incremental AI chips in the future. On one hand, cloud vendors' self-researched acceleration cards are significantly better in cost compared to outsourcing from commercial companies like NVIDIA. In Q3 2024, NVIDIA's gross margin has reached 74.4%, and adopting self-researched acceleration cards will help cloud vendors obtain more AI computing power with limited capital expenditures.

On the other hand, cloud vendors' self-researched ASICs are more flexible. Cloud vendors can design AI chips and server architectures according to their own model training and inference needs, thereby achieving better training and inference results. As the self-researched ASIC products of cloud vendors gradually mature, the proportion of self-research in AI computing power layout is expected to gradually increase in the future.

AI Terminal Products Open a New Situation

Compared to traditional hardware manufacturers, internet giants are directly participating in terminal hardware development by leveraging their control over large models. For example, ByteDance's "Doubao" has quickly accumulated hundreds of millions of users, demonstrating enormous market potential.

"From May to July this year, the daily new users of the Doubao App skyrocketed from 200,000 to 900,000, and in September, it became the first AI application in China to surpass 100 million users. According to data from Quantum Bit Think Tank, by the end of November, Doubao's cumulative user base for 2024 had exceeded 160 million; in November, an average of 800,000 new users downloaded Doubao daily, with nearly 9 million daily active users, ranking second globally among AI applications and first domestically."

Minsheng Securities believes that 2025 is expected to witness a resonance between the cloud and the terminal, with the intersection of the two being the ByteDance industrial chain. The abundant computing resources and the willingness to invest heavily are the main reasons for the rapid growth of ByteDance's Doubao.

The implementation of AI goes beyond the cloud; terminal products are becoming the focus of the next round of transformation. From smartphones to wearable devices, to smart home and automotive electronics, AI is driving the comprehensive intelligence of hardware.

“We are optimistic about the trend of AI + smart terminals. AI will reconstruct the growth of the electronics industry, inject new vitality into smart hardware, bring about a deep transformation of product logic, and accelerate the trend of hardware intelligence and companionship. Whether it is smartphones, PCs, AIOT, wearable devices, or automotive electronics, there is potential for revaluation.

With vast potential market demand, we are optimistic about the continued deepening of cooperation between leading large model companies like ByteDance and hardware manufacturers in various fields, continuously launching diversified innovative applications such as intelligent dialogue and early childhood education, and new smart terminals are about to achieve large-scale implementation.”

Minsheng Securities states that the important "partners" of core cloud/hardware manufacturers will become the main line of investment. As the AI industry begins to move towards the application layer/inference side, the discourse power of the AI industry may tilt towards cloud manufacturers/electronic brand manufacturers. However, regardless of how the discourse power is passed on, the implementation of AI applications requires the support of hardware, making it crucial to become an important partner of industry leaders in this process. Key cloud manufacturers (Google, Microsoft, Meta, ByteDance, Baidu) + electronic brand manufacturers (Apple, Huawei, Xiaomi, Tesla, etc.) are worth paying attention to.

The Era of Domestic Demand is Coming

Currently, whether it is overseas giants like Google and Amazon, or domestic companies like ByteDance and Tencent, cloud manufacturing giants have taken over the baton of AI industry development. In this process, the industrial chain advantages of Chinese local enterprises are particularly prominent. For example, in fields such as PCB (printed circuit boards), copper cables, temperature control, and power supplies, Chinese companies have accumulated significant advantages through years of deep cultivation. Unlike the previous stage that relied heavily on capital investment, future AI investments will focus more on collaborative establishment, order implementation, and performance realization. With the vigorous expansion of cloud vendors, domestic demand will become an important engine for China's AI industry, and investment returns will also be more robust.

Overview of Related Industry Chain Companies

In the AI industry chain, companies at various stages are participating in this transformation through technological innovation and resource advantages. Minsheng Securities has systematically sorted out the participation of these companies:

ASIC

ZTE Corporation: The company has been deeply engaged in chip research and development for nearly 30 years, continuously strengthening its investment in advanced process design, architectural innovation, packaging technology, and core intellectual property, establishing a digital and efficient development platform, forming industry-leading full-process chip design capabilities, focusing on the research and development of ASIC chip underlying technologies, and closely following the trend of integrated computing and networking, creating an efficient, environmentally friendly, and intelligent full-stack computing and networking foundation around "data, computing power, and network."

AEC

Ruikeda: The company focuses on the design, development, and manufacturing of connection system products and is a well-known connector manufacturer in China, possessing a complete connector supporting industrial chain. The company is closely monitoring the AI and data center industries and will focus on the layout and development of AEC high-speed component products in the future.

Zhaolong Interconnect: The company is a leading high-tech enterprise in China specializing in the design and manufacturing of data cables, specialized cables, and connection products. High-speed cable components are the company's key development direction, and it has a complete set of independent design capabilities for high-speed connection products inside and outside servers, including PCBA, wire-end connectors, etc., with deep technical accumulation in signal simulation, structural design, and mold design.

Bochuang Technology: The company's products mainly target telecommunications and data centers, as well as consumer and industrial interconnect fields, and it has supplied high-speed copper cables in bulk to many domestic and foreign internet customers. Its high-performance 800G AEC series products support 400G/800G transmission rates, covering OSFP/QSFP-DD/QSFP112 packaging, with a maximum transmission distance of 7 meters, greatly enriching application scenarios compared to traditional DAC.

Luxshare Precision: Copper connections are the core products of the company's business. In the field of high-speed interconnection for data centers, the company has built a complete solution for cabinet interconnection, inter-cabinet interconnection, servers, and switches, and has proactively collaborated with leading chip manufacturers to jointly formulate next-generation high-speed connection standards such as 800G and 1.6T for mainstream data centers and cloud service providers worldwide.

Copper Connection

Wolfe Nuclear Materials: The company's high-speed copper cable business is operated by Letin Intelligent Connection, which has rich product development experience and process control experience. Through years of technological accumulation, Letin Intelligent Connection has mastered the core technologies of all key products, with the main products being 400G/800G high-speed communication cables. Currently, a series of products such as SFP, QSFP, QSFP-DD, SAS, and Mini-SAS have gained recognition from many well-known customers, and 224G high-speed communication cable products have been stably delivered in bulk to direct customers Jingda Co., Ltd.: The company is one of the world's leading manufacturers of electromagnetic wire, with main products including copper-based electromagnetic wire. The company's holding subsidiary, Hengfeng Special Conductors, primarily produces silver-plated wire, nickel-plated wire, tin-plated wire, pure silver wire, and pure nickel wire, possessing core technology for the preparation and processing of high-performance copper-based wire materials. It has developed a complete set of preparation and processing technology with independent intellectual property rights, successfully applied to a series of products such as copper and precious metal wire materials.

PCB

Shengyi Technology: The company has successfully developed multiple server clients, including Amazon, and the AI-supported motherboard and accelerator card projects have entered mass production. Future next-generation product projects are under continuous collaborative development. Shengyi Technology focuses on high-precision multilayer PCBs and high-frequency high-speed PCB research and development, with several clients' AI product projects having successfully achieved mass production. In the first three quarters of 2024, the company's revenue from server products saw significant year-on-year growth, with server products accounting for 42.45%, an increase of 20.87 points year-on-year.

Guanghe Technology: Guanghe Technology is deeply engaged in the high-speed PCB field, with products mainly positioned in the mid-to-high-end application market, where revenue from PCB products for servers accounts for about 70%. The shipment proportion of Guanghe Technology's AI server-related products has exceeded 25%, with a growth rate significantly higher than that of traditional servers, particularly for products like UBB and I/O boards. The company has passed certification from downstream clients for NVIDIA's Blackwell GB200 chip-related products, mainly focusing on server CPU motherboards.

Shennan Circuit: Shennan Circuit has been deeply engaged in the PCB industry for 40 years and has become a global leader in wireless base station RF power amplifier PCB supply and the largest domestic packaging substrate supplier. The company's orders in the data center sector saw significant year-on-year growth in the first three quarters, mainly due to increased demand for AI accelerator cards and EagleStream platform products. The demand for 400G and above high-speed switches and optical module products has grown driven by AI-related needs, optimizing the product structure. In terms of new product research and development, the company has collaborated with downstream clients to develop next-generation platform products.

Weilgao: The company's products cover thick copper boards, Mini LED optoelectronic boards, planar transformer boards, optical modules, etc., applied in industrial control, display, automotive electronics, new energy, communication equipment, AI servers, and other fields. The company's industrial control business accounts for about 40%, and it is currently expanding from primary power supplies to secondary power supplies. In the AI field, the company mainly includes DC-DC related power supply product PCBs, with related PCB layers reaching up to 30 layers.

Cooling

Shenling Environment: The company offers numerous liquid cooling products for data centers, including end-to-end full-chain solutions, core CDU equipment, secondary side systems, manifold pipelines, primary side, and external cooling sources. A new data center product base will soon be put into production, and the quality, technology, and delivery capabilities of related businesses, especially liquid cooling products, will significantly improve in the future The company participated early in the liquid cooling product field, has many project cases, and has high technical stickiness with important clients. The growth trend of related businesses in the future is promising.

InvoTech: InvoTech is an industry-leading provider of precision temperature control energy-saving solutions and products. With its key independent technologies, it has established an end-to-end layout for liquid cooling technology, launching the Coolinside liquid cooling cabinet and full-chain liquid cooling solutions for computing power devices and data centers. As a leading enterprise in data center liquid cooling projects, it has delivered a cumulative 900MW of liquid cooling projects as of April this year, with solid technical strength and market position, and rich project execution experience.

GaoLan Co., Ltd.: As one of the earliest companies in China focusing on the innovation and industrial application of power electronics thermal management technology, it possesses industry-leading technology. The main products of its thermal management business have reached domestic leading or internationally advanced levels, with some products achieving international leading standards. The company's liquid cooling solutions are primarily based on cold plate technology, and its related clients include ByteDance, Alibaba, Tencent, GDS, Inspur, and others. The company is actively expanding into overseas markets, and its thermal management products have been deployed in multiple countries and regions worldwide, with application scenarios involving high-power power electronics, data centers, and more. The company's products have obtained certifications such as CE, ETL, ASMEU, and ROHS, and have been directly or indirectly applied in overseas projects in Europe, the United States, and other regions.

Power Supply

MeigeMite: The company is an industry-leading provider of power supply solutions, possessing industry-leading technology levels and product R&D and supply capabilities for high-power, high-efficiency network power supplies. On October 17, 2024, the company announced a collaboration with NVIDIA to provide advanced power supply solutions for the NVIDIA MGX™ platform and GB200 system.

Oulutong: The company has been deeply engaged in the power supply field for many years, equipped with a comprehensive R&D and product laboratory that is fully functional and versatile. The technical parameters of its products can be independently designed, tested, and experimented, ensuring R&D speed and quality standards. Its main products include power adapters, data center power supplies, and other power supplies.

Risk Warning and Disclaimer

The market has risks, and investment should be cautious. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investing based on this is at one's own risk