China's Semiconductor Self-Sufficiency Efforts Hindered by Software Ecosystem Challenges
China's leading model developers continue to rely on Nvidia chips for training advanced models, despite efforts to replace foreign semiconductors with domestic alternatives. The challenge lies not only in producing domestic chips but also in developing a competitive software ecosystem. Huawei has introduced its Compute Architecture for Neural Networks (CANN) for its Ascend chips, but developers accustomed to Nvidia's CUDA platform face significant engineering challenges in transitioning. This involves rewriting and optimizing substantial portions of existing code, which can increase time and costs by at least 50%. The transition is more straightforward for open-source models, but proprietary models like Moonshot AI's Kimi K3 require extensive work. The distinction between training and inference is crucial, as training demands significant computing resources. While some models have been adapted for domestic hardware, the transition remains complex. 
