At least 10 AI chip companies in China have shipped products, with Nvidia's market share in the country potentially dropping from 40% to 8% this year

According to Beat Intelligence monitoring, on April 19, The Information outlined the landscape of AI chips in China. While Washington's politicians and lobbyists mention Huawei the most, during the same period, at least 10 Chinese companies were involved in designing and shipping AI chips. Huawei is just the most well-known among them. Others like Cambricon, Horizon Robotics, Moore Threads, Horizon Robotics, Muqi, Tiansu Zhixin (Iluvatar CoreX), Bitmain, Horizon Robotics, and Kunlun Core are also synchronously supplying these chips, with over half of them already being listed on the stock market. Companies such as Bitmain and Kunlun Core are also preparing for an IPO. Customers have expanded from early state-owned enterprises to Alibaba, ByteDance, and Tencent. These tech giants even deploy their own engineers to assist in chip software development, providing design companies with more stable revenue and closer feedback to real business needs.

Bernstein estimates that by 2025, Nvidia and Huawei will each hold about 40% of the AI chip market in China, with Nvidia's market share possibly dropping to around 8% this year. The revenue of domestic AI chip companies is expected to increase from around $2 billion in 2023 to about $82 billion in 2028.

In terms of strategy, these companies do not intend to directly compete with Nvidia in training chips but are focusing on inference. While their single-card performance still lags, they compensate by using larger chip clusters and network technology to efficiently handle millions of inference requests in a stable and cost-effective manner.

There are two bottlenecks: production capacity and software. Cambricon and Horizon Robotics secured domestic foundry capacity earlier, with the latter still catching up. Tiansu Zhixin uses TSMC wafers, while Baidu Kunlun Core previously used Samsung. However, both of these sources contain U.S. technology, and export controls would limit the performance ceiling of Chinese chips. As a result, Kunlun Core has already begun discussions with domestic foundries to switch orders. In terms of software, Cambricon and Tiansu Zhixin prioritize code compatibility with CUDA to minimize developers' changes during the switch.

Alibaba Bitmain and Baidu Kunlun Core are taking a different approach, optimizing hardware, compilers, and software together. Since early 2025, both companies have been using self-developed chips to train some of their in-house models. Huawei's Ascend 950PR, launched this year, has various versions tailored for inference workloads. DeepSeek directly coordinates with Huawei to enable new models to run on Ascend hardware on the same day they are launched.

The supply chain is also catching up. After the U.S. severed the supply of high-end HBM to China in 2024, Huawei and Tiansu Zhixin are now testing domestically produced HBM3 coupled with their AI processors. Most of China's leading chips have matched or exceeded Nvidia's H20, which was specially modified for the Chinese market. Tiansu Zhixin and Bitmain are developing the next-generation products that compete with Nvidia's H200, the most powerful model Nvidia is currently allowed to sell to China.