The release of DeepSeek V4, a multimodal open-source AI model, has been delayed, sparking discussions on China's AI ecosystem's shift from NVIDIA's CUDA to Huawei's Ascend chips. The model, expected to feature a trillion-parameter scale, is undergoing optimization for Huawei's platform, marking a significant move towards reducing reliance on CUDA. This transition, however, presents challenges due to differences in hardware and software capabilities between NVIDIA and Huawei. DeepSeek's delay underscores the complexities of adapting AI models to non-CUDA platforms, highlighting the trade-offs between compatibility and independence. While Huawei's CANN framework aims to emulate CUDA's capabilities, the process reveals systemic challenges in achieving performance parity. The situation reflects broader strategic decisions facing China's AI industry as it balances short-term compatibility with long-term self-reliance.