Why NVIDIA Left China and How Domestic AI Chips Are Rising to Lead

“NVIDIA’s market share in China has dropped from 95% to 0, and we have completely exited the Chinese market,” said CEO Jensen Huang.

During a Citadel Securities event, Huang also noted that U.S. policies have caused America to lose one of the world’s largest markets.

Before the exit, NVIDIA’s leading GPU architecture and mature CUDA ecosystem gave it near‑100% dominance in China’s large‑model training chip market.

Huang previously predicted that China’s AI market would reach roughly $50 billion within two to three years.

With NVIDIA gone, a variety of domestic AI chip companies are stepping up, including general‑purpose GPU designers such as Muxi, Moore Thread, Hai Guang Information, Tian Shu Zhi Xin, and Bi Ren Technology, as well as ASIC/DSA specialists like Huawei Ascend, Cambricon, Kunlun Chip, and Suiyuan Technology.

Huawei Ascend, leveraging years of technology and ecosystem advantages, has become the main high‑end alternative; its Ascend 910 chip is already used in large‑scale AI model training at domestic internet firms.

Listed companies such as Cambricon and Hai Guang Information have made notable breakthroughs in niche areas. Cambricon’s Siyuan 590 chip supports mainstream deep‑learning frameworks and delivers about 80% of the performance of NVIDIA’s A100 in vision and large‑language‑model workloads, partnering with leading algorithm firms like Zhixiang Future and Baichuan AI.

Moore Thread has developed four generations of GPU architectures and derived more than ten chip models, covering AI computing, professional graphics, and desktop graphics acceleration.

Muxi focuses on “training‑inference integration” GPU chips; by Q1 2025 it had sold over 25,000 units and delivered products to major smart‑computing clusters in Beijing, Shanghai, Hangzhou, and other cities.

Recently, Muxi Integrated Circuit (Nanjing) Co. launched the first fully domestic general‑purpose GPU, the Xiyun C600, marking a historic breakthrough for high‑performance domestic GPUs.

The Xiyun C600 is built on Muxi’s proprietary GPU IP and completes a full domestic supply chain from design to packaging, offering large memory capacity and mixed‑precision compute to provide secure, controllable compute power for the digital economy.

IDC data shows that in 2024, domestically developed AI chips accounted for 30% of the Chinese market, and this share is expected to exceed 50% by 2025, indicating rapid growth of domestic influence.

Industry insiders predict that soon domestic compute chips will surpass foreign ones, with current usage already roughly split 50‑50, reflecting rising acceptance.

However, the biggest remaining limitation is advanced wafer manufacturing technology; while design gaps are narrowing and software ecosystems still lag behind CUDA, sufficient investment and ecosystem development can gradually close these gaps.