The Battle for the Trillion-Dollar AI Chip Market: Who Will Prevail?

[Introduction] The Chinese AI chip market is evolving into a “hundred flowers blooming” era reminiscent of the “Warring States period“.

By Yin Zhenmao, China Fund Reporter

The rapidly growing Chinese artificial intelligence (AI) chip market is attracting numerous chip companies, both domestic and international, including Nvidia, to compete for opportunities.

However, by 2025, the global AI chip market is showing a trend of “GPU leading, ASIC rising,” indicating the arrival of the ASIC architecture era for AI chips. The Chinese AI chip market is also transitioning from Nvidia’s previous dominance to a “hundred flowers blooming” era reminiscent of the “Warring States period“.

Industry experts believe that both GPU and ASIC architectures will develop rapidly in the domestic AI chip market; however, among domestic AI chip manufacturers, ASIC architecture currently holds a dominant position, and products based on ASIC architecture are expected to perform better than their foreign counterparts.

Competing for the Trillion-Dollar Market

On July 15, Nvidia founder and CEO Jensen Huang announced during his visit to China that the supply of the H20 chip, specifically for the Chinese market, will resume. Meanwhile, another AI chip giant, AMD, also expressed hopes to restore sales to China.

This marks Huang’s third visit to China this year, highlighting Nvidia’s significant focus on the Chinese market. Behind this is Nvidia’s relative decline in the Chinese market in recent years.

In May 2025, Huang stated that four years ago, Nvidia’s market share in China was as high as 95%, but now it has dropped to 50%. According to the prospectus of Muxi Integrated, Nvidia’s revenue share in China fell from 26.42% in 2021 to 13.11% in 2024.

On the other hand, the market share of domestic AI chip manufacturers is rapidly increasing. IDC data shows that by 2024, the penetration rate of domestic AI chip brands in China is expected to reach about 30%, with shipments reaching 820,000 units, a significant increase from the 15% penetration rate of domestic brands in the same period last year.

On June 30, the Shanghai Stock Exchange accepted the IPO applications of domestic GPU company Moore Threads and Muxi Integrated for the Sci-Tech Innovation Board.

The day after the acceptance, on July 1, shares of AI chip leader Cambricon fell sharply after opening, closing down 6.4%, while Haiguang Information also dropped 3.39%.

Industry insiders generally believe that if Moore Threads and Muxi Integrated successfully go public, the scarcity of Cambricon, as the only AI chip company in A-shares, will be somewhat impacted.

Baidu’s Kunlun Core recently completed a new round of financing, with Baidu’s shareholding ratio diluted from an early 76.17% to the current 67.5%. Industry insiders believe that Kunlun Core’s financing aims directly at an IPO.

Currently, mainstream AI chips include general-purpose architectures like GPUs, as well as dedicated architectures like ASICs and FPGAs. Among these, GPUs dominate the AI market, but dedicated chips represented by ASICs and FPGAs have also been commercialized and occupy a certain market share.

Participants in the domestic AI chip market include both domestic and international GPU chip companies such as Nvidia, AMD, Haiguang Information, Moore Threads, and Muxi Integrated, as well as ASIC chip companies like Huawei HiSilicon, Cambricon, and Kunlun Core.

According to Frost & Sullivan, the scale of China’s AI chip market is expected to surge from 142.537 billion yuan in 2024 to 1.34 trillion yuan in 2029, with an average annual growth rate of 53.7% from 2025 to 2029.

Globally, the development history of the GPU industry, which currently holds a major position in AI chips, can be divided into the initial stage, evolution stage, and explosive growth stage. The rise of AI technology has propelled GPUs to dominate the AI computing field. Currently, the industry has entered the explosive growth stage.

However, China’s GPU industry is still in the initial stage.

ASIC’s Strong Rise

Globally, the AI chip market has basically formed an oligopoly dominated by Nvidia and AMD, with Nvidia as the industry leader, holding over 80% of the global market share. However, Nvidia and its representative GPU architecture are facing challenges from ASIC architecture.

Wang Haiwei, co-chief analyst of the electronic industry at Founder Securities, told China Fund reporters that by 2025, the global AI chip market is showing a trend of “GPU leading, ASIC rising.”

A recent report from Nomura Securities indicates that Nvidia’s GPUs currently account for over 80% of the AI server market, while ASICs only account for 8% to 11%. By 2025, the combined shipment of ASIC chips from Google and Amazon AWS is expected to be about 40% to 60% of Nvidia’s GPU shipments. By 2026, with large-scale deployments by Meta and Microsoft, ASIC shipments are expected to surpass Nvidia’s GPUs, marking the official arrival of the ASIC era.

According to professionals from Haiguang Information, GPUs use a large-scale parallel computing unit design, with hardware design highly dependent on programmability, enabling flexible task scheduling through a CUDA-like ecosystem; ASICs, on the other hand, are fully customized chips generally optimized for specific algorithms, providing computational efficiency and energy efficiency in that domain.

This professional stated to China Fund reporters: “Due to the rapid development of AI algorithms, once the algorithms evolve, the computational efficiency of ASICs will significantly decline, making them suitable only for certain specific scenarios. Therefore, the rise of ASICs does not mean the decline of GPUs. Currently, overseas cloud service providers are actively promoting ASIC design, primarily driven by the need to optimize internal workloads and achieve better cost-performance ratios. Meanwhile, domestic AI chip manufacturers are transitioning more from ASIC architectures to GPGPU-like architectures.”

In this professional’s view, general-purpose GPUs and ASICs have different positioning and application scenarios, and both technologies will coexist for a long time, providing solutions for different demand scenarios.

Wu Xiongang, former chairman of Arm China and current chairman of Corelab Tech, believes that customization is an inevitable trend, and the era of ASIC architecture is coming. In the next ten years, both GPU and ASIC architectures in the AI chip market will grow, as the computational load is expected to increase by 10 to 100 times, leading to rapid growth for both ASIC custom architectures and GPU general architectures.

“However, because customized chips are relatively new, their advantages include lower costs and higher computational efficiency, which will inevitably lead to faster growth compared to GPU architectures as applications increase. Therefore, in the future structure of the AI chip market, the market share of ASICs will gradually increase. It is highly likely that ASIC-type AI chips will eventually share the market equally with Nvidia’s GPU-type chips,” Wu Xiongang stated.

A New Integrated Architecture Will Emerge in the Future

On July 16, Jensen Huang delivered a speech at the opening ceremony of the 3rd China Supply Chain Expo, stating that the next wave of AI will involve understanding the physical world, with robotic systems capable of performing tasks expected to emerge within ten years, coordinated by software and AI in factories.

Wu Xiongang’s Corelab Tech focuses on developing open-source robotic computing architectures and chip designs.

In Wu Xiongang’s view, as Huang mentioned, once robots start running, the total computational load will need to increase by 100,000 to 1,000,000 times, necessitating further evolution of chips. In this context, a single GPU architecture will certainly not be able to fully address these new growing applications and computational loads.

“In ten years, robots will generate a lot of data, which will be real data from the physical world, and the computational architecture required will differ from the current large models. Ultimately, the chip architecture for robots will not be purely GPU; it will definitely be an integrated architecture, as only a new integrated architecture can solve new problems,” Wu Xiongang stated. “Computational architectures are continuously iterating and evolving. In the next 5 to 10 years, GPUs will continue to play the most important role, while ASICs are rapidly rising, likely leading to a hybrid or integrated architecture.”

The China Academy of Information and Communications Technology’s “Research Report on the Development of Cloud-based AI Chip Industry (2023)” points out that in terms of AI computing architecture, generality and specificity are not entirely opposed. As AI computing architectures continue to develop, generality and specificity are increasingly merging at the architectural level to achieve a balance between generality and high performance. The evolution history of Nvidia’s GPU architecture reflects the characteristics of the mutual integration of generality and specificity.

The AI Chip Market Will Enter the “Warring States Period”

IDC China analyst Du Yunlong told China Fund reporters that currently, the top three domestic AI chip manufacturers are Huawei HiSilicon, Baidu’s Kunlun Core, and Cambricon.

According to Wang Xiaolong, director of the enterprise service department at semiconductor research firm Xinsuo Research, among the four major GPU manufacturers, Moore Threads started with graphics GPUs, and the Shanghai government has provided significant support to Suiruan Technology, while Muxi Integrated has developed the fastest this year. In the field of graphics GPUs, Haiguang Information is the strongest among domestic manufacturers.

“It is somewhat similar to the 1980s when PCs first emerged, and CPUs were in a ‘Warring States period.’ AI chips will also enter a ‘Warring States period,'” Wu Xiongang stated.

Regarding the domestic AI chip market, Moore Threads stated in its prospectus that in the segmented market for AI chips in China, the GPU market is growing the fastest, with its market share expected to rise from 69.9% in 2024 to 77.3% in 2029. This means that by 2029, the scale of China’s AI intelligent computing GPU market will reach 1.03 trillion yuan.

Wu Xiongang believes that among domestic AI chip manufacturers, ASIC architecture holds a dominant position, and products based on ASIC architecture will develop better than those abroad.

“In fact, with Nvidia’s products being restricted, the market share of GPU architecture AI chips will decline, while the market share of ASIC architecture will increase. Although Nvidia has just announced the resumption of exports of its H20 chip to China, which will significantly help meet the demand for computing chips in China in the short term, it will not have a significant impact on the long-term competitive landscape,” he said.

“Currently, both architectures have demand and will develop rapidly, especially in the absence of Nvidia. However, in the inference segment, the advantages of GPU architecture are not as pronounced as last year. Therefore, in the inference segment, at least domestically, five years from now, the market share of ASIC architecture may far exceed that of GPU architecture, but it is uncertain in the training segment,” Wu Xiongang emphasized.

Previously, Li Guojie, an academician of the Chinese Academy of Engineering and former director of the Institute of Computing Technology, Chinese Academy of Sciences, stated in an interview with China Fund reporters that in the current development of artificial intelligence, the demand for inference, which requires more devices than training, will become the primary demand.