Nvidia chips remain central to China's AI development

Despite Beijing's efforts to promote domestic semiconductor alternatives, China's leading AI developers continue to rely heavily on Nvidia chips to train advanced large language models (LLMs). The primary obstacle to transitioning to domestic hardware, such as Huawei's Ascend chips, is the deeply established software ecosystem surrounding Nvidia’s CUDA platform.

Switching to Huawei’s Compute Architecture for Neural Networks (CANN) requires developers to rewrite and optimize significant portions of their existing code. Industry experts note that this transition can be prohibitively expensive and time-consuming. For some research teams, moving workflows to Ascend could increase costs and time requirements by at least 50%.

The difficulty of migration varies by model type. While open-source models like DeepSeek may be migrated by a small team within a month, more complex models with restricted source code, such as Moonshot AI’s Kimi K3, could require approximately 10 engineers and over six months of additional work to transition.