According to monitoring by Dongcha Beating, The Information reported on April 19 that there are at least 10 Chinese companies designing and shipping AI chips, with Huawei being the most well-known among them. Other companies include Cambricon, Hygon, Moore Threads, Muxi, Tianxu Zhixin (Iluvatar CoreX), Biren, Pingtouge, and Kunlun Core, with more than half already listed and Pingtouge and Kunlun Core preparing for IPOs. Customers have expanded from early state-owned enterprises to major firms like Alibaba, ByteDance, and Tencent, which have sent their engineers to assist in modifying chip software, providing design companies with more stable revenue and feedback closer to real business needs. Bernstein estimates that by 2025, Nvidia and Huawei will each hold about 40% of the AI chip market in China, while Nvidia's share may fall to around 8% this year. Revenue for local AI chip companies is expected to increase from approximately $2 billion in 2023 to about $82 billion by 2028. These companies do not intend to directly compete with Nvidia in training chips but are focusing on inference. Although single-card performance still lags behind, they compensate by using larger chip clusters and networking technology to handle millions of inference requests stably and cost-effectively. The two bottlenecks are production capacity and software. Cambricon and Hygon secured domestic foundry capacity early on, while later entrants are still catching up. Tianxu Zhixin is currently using TSMC's latest process, and Baidu's Kunlun Core previously used Samsung, but both processes involve U.S. technology, and export controls will limit the performance ceiling of chips for China, prompting Kunlun Core to negotiate with domestic foundries for order transfers. In terms of software, Cambricon and Tianxu Zhixin prioritize compatibility with CUDA-written code to minimize changes for developers. Alibaba's Pingtouge and Baidu's Kunlun Core are taking a different approach by optimizing hardware, compilers, and software together; since early 2025, both have been using self-developed chips to train some of their own models. Huawei's Ascend 950PR, launched this year, offers multiple versions tailored to inference workloads, and DeepSeek has directly collaborated with Huawei to ensure new models can run on Ascend hardware on the day they go live. The supply chain is also keeping pace. After the U.S. cuts off high-end HBM supplies to China in 2024, Huawei and Tianxu Zhixin are testing domestically produced HBM3 to pair with their AI processors. Most leading Chinese chips have caught up with or surpassed Nvidia's H20, which was modified specifically for the Chinese market, while Tianxu Zhixin and Biren are developing next-generation products to compete with the H200, the strongest model currently allowed for Nvidia to sell to China.