[Huang Renxun's Worst Fear is Happening: DeepSeek Begins Moving Away from U.S.-Dominated AI Tech Stack, Adapting Core Components to Huawei Ascend] Beating AI Newsflash: DeepSeek has partnered with Huawei to migrate an entire suite of AI training and inference core components onto Ascend. The newly open-sourced tools cover operator development, matrix computation, cross-card communication, Attention mechanisms, and data filtering, essentially corresponding to the foundational tools DeepSeek previously built for NVIDIA GPUs. Reuters reports this as the latest development in Chinese tech companies seeking alternatives to NVIDIA's ecosystem. The most critical aspect here is TileLang. Many operators in DeepSeek V4 training have already been implemented using TileLang. It addresses a fundamental issue: how developers can efficiently execute model computations directly on chips. Previously, this capability was heavily built around CUDA, but now DeepSeek and Huawei are creating a corresponding version for Ascend. This directly aligns with Huang Renxun's concerns expressed months ago. In April, during Dwarkesh Patel's podcast, he stated that if DeepSeek were to release on Huawei chips first, it would be a "bad outcome" for the U.S. His worry has never been solely about Huawei chips themselves but about models being optimized for Huawei's architecture. Once global developers find that the same model runs better by default on non-U.S. hardware, the competitive edge of U.S. chips will gradually erode. DeepSeek's subsequent actions have almost entirely followed this trajectory. V4 has already begun adapting to Ascend 950, and Huawei has claimed its chips were involved in parts of V4 training. Last week, *The Information* revealed that Liang Wenfeng has prioritized increasing training on domestic chips as a key focus for DeepSeek, with new Huawei training chips expected as early as Q4. Now, the software layer is being addressed. From model adaptation to training chips, matrix computation, MoE communication, Attention mechanisms, and operator development, DeepSeek is progressively migrating foundational capabilities that were heavily reliant on NVIDIA and CUDA to Huawei Ascend. The U.S.-China AI rivalry has shifted from competing over stronger models and more GPUs to building a comprehensive tech stack spanning chips, operators, communication, and training frameworks. CUDA's hardest-to-replace aspect has never been just the GPU itself but the entire software ecosystem built around it. DeepSeek and Huawei are now working to fill in this layer. [Original Link]