A report by the Hong Kong-based South China Morning Post (SCMP) on Aug. 10 said many of China’s most advanced AI models are still trained on Nvidia chips, despite government pressure for semiconductor self-reliance, due to the cost and technical burden of switching to domestic chips. The key obstacle is the software ecosystem, as Nvidia’s CUDA is a standard for AI development, while moving to Huawei’s CANN platform requires large-scale code rewrites. James Wang, an AI developer at a Shanghai university research institute, said CUDA code cannot run directly on Ascend and requires extensive rewriting, adding that moving workflows to Huawei Ascend chips could take at least 50 percent more time and cost. An engineer in Beijing said open-source models like DeepSeek can be trained on Ascend with 2 to 3 additional engineers for about a month, while models like Moonshot AI’s Kimi K3, which only release weights, may require about 10 engineers for more than 6 months of additional work. The trend shows domestication is advancing first in inference rather than training. DeepSeek-V4 and Kimi K3 have been adapted for Huawei and Alibaba platforms. Kimi K3 was reported in July by The Information to have been trained on chips including Nvidia’s Blackwell processors. Meituan said in June, when unveiling Longcat-2.0 with 1.6 trillion parameters, that it completed training