DeepSeek's upcoming V4 model — a 1-trillion-parameter MoE system targeting late April 2026 — will run on Huawei's Ascend 950 PR, which delivers 1 PFLOPS (FP8) compute and 2 TB/s interconnect bandwidth using SMIC's N+3 process, according to TrendForce. DeepSeek withheld early optimization access from AMD and NVIDIA while working with Huawei and Cambricon, and Alibaba, ByteDance, and Tencent have placed large Ascend orders in anticipation of the release. If V4's software stack stabilizes on Ascend hardware, China's frontier AI development pipeline could become effectively independent of NVIDIA CUDA within one to two years.