"The development of artificial intelligence technology is not limited by the human baseline as the highest ceiling, but far exceeds the human level." Wei Kai, Director of the Artificial Intelligence Research Institute of China Academy of Information and Communications Technology, made this judgment at the "To the Future · C-Talk" Technology Speech Conference, revealing the key turning point where AI technology is moving from technological breakthroughs to industrial penetration. With the implementation of the State Council's "Opinions on Further Implementing the 'Artificial Intelligence +' Action" and the more than 800 application scenarios built by central enterprises' "AI +" special actions in 16 industries, AI technology has become the core engine driving the development of new quality productive forces. In this process, as the hardware cornerstone of "Artificial Intelligence +", the independent controllability and large-scale manufacturing capacity of AI chips directly determine China's strategic initiative in global AI competition.
Market Changes: From Capital Frenzy to Value Return
The AI industry is undergoing a profound "qualitative change". Gamma Data shows that the global AI industry financing amount reached 482.9 billion yuan in 2024, a year-on-year increase of 77.2%, but the number of investments decreased by 52.4%. Capital is shifting from extensive layout of casting a wide net to areas of in-depth technological cultivation. This transformation is particularly obvious in the AI chip field - Gartner data shows that the global AI chip revenue reached 71.252 billion US dollars in 2024, a year-on-year increase of 33%, of which the server AI accelerator market size has reached 21 billion US dollars, and is expected to increase to 33 billion US dollars by 2028. In the Chinese market, Frost & Sullivan statistics show that the AI chip market size reached 142.537 billion yuan in 2024, of which customized chips such as ASIC accounted for 30.1%, becoming the fastest-growing segment.
Changes in market structure reflect the upgrading of industrial demand. As the parameter scale of large models exceeds one trillion, a single chip can no longer carry training and inference needs. The "computing power interconnection network" pointed out by researchers at the Institute of Computing Technology of the Chinese Academy of Sciences has become an industry consensus - building an ultra-high bandwidth and low-latency computing power collaboration system through full-link optimization from on-chip interconnection, chiplet interconnection to system-level interconnection. In this context, the independent controllability of AI chips is no longer a technological sentiment, but a necessary condition for industrial security. "Artificial Intelligence + is not a simple technological superposition, but an all-round paradigm shift". In key fields such as energy and manufacturing, independently controllable AI chips have become the core element to ensure industrial security.
Independent Controll ability: From Single-Point Breakthrough to System Capability Construction
Domestic AI chips are achieving a leap from "usable" to "easy to use". At the 2025 World Artificial Intelligence Conference, Suiyuan Technology and Muxi simultaneously released the latest generation of training-inference integrated chips, marking the breakthrough of domestic chips in the high-end field. Suiyuan Technology's L600 chip is equipped with 144GB storage capacity and supports FP8 low-precision computing. Its previous generation product S60 has shipped 70,000 units, supporting the construction of five major intelligent computing centers including the Gansu Qingyang 10,000-card inference cluster. Muxi's Xiyun C600 adopts advanced HBM3e memory to achieve 144GB storage capacity. More importantly, it achieves "domestic independent controllability of the entire technical process of chip R&D and manufacturing", breaking the dependence of high-end AI chips on overseas supply chains.
This breakthrough is by no means accidental. In the design link, Verisilicon, known as "China's first semiconductor IP stock", saw its AI computing power-related revenue account for 52% in the first half of 2025. Its semiconductor IP licensing business ranks eighth in the world, providing key technical support for domestic chips. In the manufacturing link, enterprises such as SMIC and Huahong Semiconductor have mass production capabilities for advanced processes of 14nm and below, providing stable production capacity guarantee for AI chips. As a key link in the industrial chain, leading enterprises such as Changjiang Electronics Technology and Tongfu Microelectronics are accelerating breakthroughs in 2.5D/3D packaging technology. Among them, Tongfu Microelectronics has achieved mass production of CoWoS for domestic AI chips such as Huawei Ascend 910B and CambriconMLU590 (Cambricon MLU590).
Equipment localization has become the key to breaking the monopoly. The domestically developed first CoWoS-level TCB thermocompression bonding machine launched by Plaision has an accuracy of ±1μm and a delivery cycle of only 30-45 days, much shorter than the 60-120 days of international manufacturers. It has entered the supply chains of Foxconn, Huawei and other enterprises. This equipment breaks the monopoly of international giants such as ASM Pacific on thermocompression bonding equipment, increasing the yield of domestic CoWoS packaging from 60% in the initial stage to more than 85%, clearing the equipment obstacles for large-scale production of AI chips.
Large-Scale Manufacturing: China's Advantage in Industrial Chain Collaboration
Large-scale manufacturing of AI chips is inseparable from in-depth collaboration in the industrial chain. The State-owned Assets Supervision and Administration Commission (SASAC) has promoted the establishment of 3 industry data industry consortia in transportation and logistics, green low-carbon, and smart energy, building more than 1,000 industry data sets to provide data support for scenario-based optimization of AI chips. This "application-driven - technology iteration" model enables domestic AI chips to quickly respond to market demand - Verisilicon's new orders from July to September 2025 reached 1.205 billion yuan, an increase of 85.88% over the same period last year, confirming the efficiency improvement brought by industrial chain collaboration.
To address the challenges of large-scale production and manufacturing, in the field of automated production, core component manufacturers represented by Valnk Technology play an invisible underlying supporting role. Its high-precision connectors customized for automated production lines support high-speed data transmission of more than 10Gbps per second, achieving zero signal loss in core interconnection and ensuring stable mass production of high-end products. In response to the needs of the testing link, Valnk Technology's intelligent test interface module increases the efficiency of connector harness testing required for production by 40%, greatly reducing mass production costs. This collaborative innovation between basic manufacturing and high-end industries is a concentrated embodiment of the advantages of China's complete manufacturing system.
In addition, breakthroughs in advanced packaging technology have accelerated the large-scale process. Although TSMC's CoWoS technology still dominates the high-end market, domestic manufacturers are catching up through technological innovation. Huaxinbang's breakthroughs in 2.5D/3D stacking and Chiplet heterogeneous integration can efficiently integrate chips with different functions, optimizing power consumption and costs by more than 30%. It is expected that 2025 will become a key window period for the rapid expansion of domestic 2.5D packaging production lines. The new production capacity of enterprises such as Changjiang Electronics Technology and Tongfu Microelectronics will effectively alleviate the shortage of high-end packaging supply, providing production capacity guarantee for AI chip mass production.
Technological Innovation: Reshaping the Chip Manufacturing Paradigm with AI
Faced with the industry pain points of long R&D cycles and high costs, "using AI to design AI chips" is becoming the way to break through. The programmable NPU product developed by Aijieke Core under Kunlun Wanwei adopts AI-aided design tools, shortening the chip R&D cycle to 14 months, a 40% reduction compared with the traditional model. This "algorithm-hardware" collaborative optimization model enables the chip's computing density to reach 100TOPS/W, which has completed prototype verification in scenarios such as medical image analysis. Verisilicon has also increased IP verification efficiency by 30% by introducing AI design tools, accelerating the delivery process of customized AI chips.
The layout of cutting-edge technologies demonstrates long-term competitiveness. Muxi has launched the R&D of the next-generation C700 series with an investment of 2.04 billion yuan, which is planned to enter the tape-out testing stage in 2026; Suiyuan Technology has formed a complete ecosystem from hardware to applications through the full-stack layout of "chip - all-in-one machine - cluster". This strategy of attaching equal importance to short-term mass production and long-term R&D enables domestic AI chips to maintain market competitiveness while continuously accumulating core technology patents, laying a solid foundation for independent controllability.
With the in-depth advancement of the "Artificial Intelligence +" action, the AI chip industry is ushering in triple benefits from policies, markets and technologies.
At the policy level, SASAC will continue to deepen the "AI +" special action to promote central enterprises to achieve better development in the field of artificial intelligence; at the market level, the outbreak of new scenarios such as AIPC, autonomous driving and low-altitude economy will create diversified chip demand; at the technical level, the maturity of technologies such as 2.5D/3D packaging and Chiplet will promote a leapfrog improvement in AI chip performance.
The development path of domestic AI chips is clear: take independent controllability as the bottom line, break through equipment and material bottlenecks through basic industrial chain collaboration; take large-scale manufacturing as the goal, improve production capacity through advanced packaging and automated production; take scenario innovation as the driving force, form a positive cycle of "model - data - chip". In 2025, China has entered the golden period of large models from technological maturity to application landing. Seizing this opportunity, domestic AI chips are expected to occupy a more core position in the global industrial chain, providing solid support for China's innovation in the "Artificial Intelligence +" era.
From the laboratory to the production line, from single-point breakthroughs to system capabilities, the development process of domestic AI chips has confirmed the profound philosophy of "quantitative change leading to qualitative change". When the independently controllable technological gene, the industrial capacity of large-scale manufacturing, and the rich and diverse application scenarios form a joint force, China's AI chip industry will surely achieve a leap from catching up to leading in global competition and become a key engine for the development of new quality productive forces.