Representative End User Clients
Representative Automation Clients
Representative Software Clients
December 2025 witnessed a dramatic turnaround in the global AI computing power market. On December 9, US President Trump approved the sale of NVIDIA H200 chips to China, a policy relaxation that brought short-term relief to an increasingly tight computing power supply landscape. Meanwhile, in China’s capital markets, domestic GPU companies such as Moore Threads Technology and MetaX Integrated Circuits moved rapidly toward public listings. Moore Threads Technology’s stock price surged more than sevenfold in just five trading days, making it the third most expensive stock on the Chinese A-share market.
This contrast between overseas policy easing and domestic market enthusiasm highlights the current reality of China’s AI chip industry. While supplementary external supply can temporarily address urgent needs, independently developed and controllable computing power remains the long-term strategic direction for domestic AI chips.
Short-Term Effects of Policy Relaxation: Easing the Gap Without Altering the Trend
The NVIDIA H200 chips approved for sale to China are based on the company’s earlier Hopper architecture. While they outperform mainstream domestic alternatives such as Huawei Ascend 910C, they remain a full generation behind the latest Blackwell-based B300 chips, which are still subject to US export restrictions. For Chinese technology enterprises, short-term procurement of H200 chips can help ease shortages in large-model training. However, reliance on external supply continues to expose these companies to persistent supply chain risks.
Market data shows that in 2024, NVIDIA and AMD together accounted for seventy-one percent of China’s AI chip market. Among domestic players, Huawei HiSilicon held a twenty-three percent share, while GPU enterprises such as Moore Threads Technology and MetaX Integrated Circuits together accounted for less than one percent.
From an application standpoint, AI computing power demand shows a clear structural divide: a smaller number of highly sophisticated training tasks versus large-scale, high-volume inference workloads. While the H200 improves efficiency in training scenarios, domestic chips led by Huawei Ascend 910C have already demonstrated strong substitution advantages in inference workloads, driven by competitive performance and superior cost effectiveness.
Industry forecasts widely suggest that the share of domestic AI chips in Chinese enterprises’ computing power procurement will rise sharply in 2026, with full substitution in inference scenarios emerging as the primary breakthrough point.
Accelerated Technological Iteration: The Core Foundation of Domestic Chips
The competitiveness of domestic chips is increasingly supported by faster technological iteration and sustained research and development investment. Since the second half of 2025, leading domestic enterprises have released frequent signals of progress. In November, Baidu launched its new-generation Kunlun Core M100 and M300, strengthening its AI computing power portfolio. In September, Huawei announced a forward-looking roadmap for its Ascend chip line, outlining new products such as the 950PR and 950DT to be introduced over the next three years.
Cambricon raised capital through private placements in October, directing funds toward the development of full-scenario large-model chips. Its research and development spending reached 300 million yuan in the third quarter of 2025, representing a year-on-year increase of forty-two point two percent and a quarter-on-quarter increase of eleven point seven percent. This sustained investment has provided a solid foundation for faster product iteration.
Domestic chipmakers are also making progress in overcoming manufacturing process constraints. Some manufacturers have achieved computing power levels approaching those of NVIDIA H100 by leveraging advanced packaging technologies. MetaX Integrated Circuits’ Xiyun C600 series delivers performance between NVIDIA A100 and H100 and is scheduled to enter mass production in the first half of 2026.
Capital Support and Strategic Differentiation: The Breakthrough Path of China’s “Four GPU Giants”
Strong capital market interest has created a virtuous cycle of technological progress, capital support, and large-scale deployment across the domestic chip industry. In early December 2025, Moore Threads Technology and MetaX Integrated Circuits made consecutive debuts on the STAR Market, while Biren Technology and Enflame advanced preparations for their own IPOs. This collective move toward public listings has strengthened funding for future research, development, and ecosystem expansion.
Each of China’s “Four GPU Giants” has adopted a distinct strategic focus. Moore Threads Technology emphasizes a full-stack layout and ecosystem synergy, building an integrated environment spanning hardware, software, and solutions, and has already commercialized GPU intelligent computing clusters at the thousand-card scale. MetaX Integrated Circuits focuses on high-performance computing power and large-model adaptation, using its Xiyun C600 series to target both training and inference workloads. Biren Technology pursues high-end breakthroughs and technology leadership, with its BR100 positioned against NVIDIA H100. Enflame prioritizes scenario-driven deployment and ecosystem co-construction, achieving large-scale implementation across sectors such as government and energy.
This differentiated positioning reduces homogeneous competition and enables domestic GPUs to cover a wide range of computing power levels and application scenarios.
Technological Synergy and Scenario Expansion: Strengthening the Path to Commercialization
Although GPUs account for approximately eighty percent of the global AI chip market, application-specific integrated circuits, particularly NPUs, are gaining momentum due to stronger scenario adaptability. The performance of Google’s Gemini 3.0, developed on TPU-based computing clusters, along with near parity in core performance indicators between TPU 7x and NVIDIA B300, highlights the growing potential of specialized chips to challenge general-purpose GPUs. Reports that Meta plans to place multi-billion-dollar TPU procurement orders further reinforce market confidence in the ASIC approach.
For domestic enterprises such as Cambricon, NPU-centered ASIC technologies are expected to become increasingly prominent in 2026, forming a complementary relationship with domestic GPUs rather than replacing them outright.
From an application perspective, domestic AI chips have already established diversified industry coverage. Baidu’s Kunlunxin chips are used across manufacturing, transportation, and financial services. Hygon Information focuses on intelligent computing centers and large-scale data processing, while Cambricon concentrates on large-model algorithms, the internet, and cloud computing. Continued investment by leading internet companies in large AI models is accelerating commercialization, giving recognized domestic chips an early-mover advantage and expanding opportunities for scale growth.
Conclusion: Anchoring Independence to Enable Leapfrog Development
By the end of 2025, developments across the computing power market—from the temporary easing of NVIDIA H200 sales to China to the rapid IPO activity and technological progress of domestic chipmakers—clearly illustrate China’s strategic trajectory in the evolving global technology landscape. In the short term, supplementary external computing power can help relieve cyclical development pressures. Over the long term, faster technological iteration, more resilient supply chains, and expanding application scenarios are jointly forming the core momentum behind domestic AI chip advancement.
As these advantages continue to materialize in 2026, domestic AI chips are positioned to drive a decisive leap forward on China’s path toward computing power independence.