DeepSeek V4 Built Natively on Huawei Ascend — First Frontier AI Model Without Nvidia

The Information reported on April 3 (confirmed by Reuters on April 4) that DeepSeek's upcoming V4 frontier model is being developed to run natively on Huawei's Ascend AI chips, abandoning Nvidia hardware for the first time in the lab's history. DeepSeek worked directly with Huawei and Cambricon Technologies to rewrite portions of V4's training and inference code to optimize for Ascend architecture, bypassing the CUDA software ecosystem that had previously constrained Chinese AI labs to Nvidia hardware. DeepSeek withheld early V4 access from Nvidia and AMD while granting Huawei extended optimization time — a deliberate strategic pivot. The V4 model, expected to launch in the final two weeks of April 2026, represents two variants targeting different capability profiles. Major Chinese tech companies — Alibaba, ByteDance, and Tencent — placed bulk orders for hundreds of thousands of Huawei Ascend chips following the report, driving Ascend chip prices up approximately 20% within days. This marks a strategically defining milestone: for the first time, China's leading frontier AI lab is deploying a world-competitive model on entirely domestic semiconductor infrastructure, validating China's AI chip self-sufficiency strategy under the 15th Five-Year Plan.