Reuters Confirms DeepSeek V4 / Huawei Shift; Analysts Debate Ascend Performance vs. Nvidia

Reuters independently confirmed on April 4 the report that DeepSeek's V4 model is being built to run on Huawei's Ascend AI chips. ChinaTechNews.com and multiple industry analysts simultaneously published assessments of the strategic and technical implications. Technical analysts noted that while Huawei's Ascend 910C chips are capable inference accelerators, they face significant software ecosystem gaps versus Nvidia's CUDA platform and roughly 40% lower performance-per-watt compared to Nvidia H100-class hardware. However, analysts noted that DeepSeek's signature architectural innovations — mixture-of-experts routing, multi-head latent attention, and aggressive quantization — could partially compensate for hardware disadvantages, just as its training efficiency innovations had made the V3 model viable on lower-spec Nvidia hardware. The geopolitical significance is independent of technical parity: a world-competitive frontier AI model running on domestic Chinese chips would validate China's self-sufficiency roadmap and accelerate adoption of the Huawei Ascend ecosystem by Chinese enterprises currently hesitant due to software compatibility concerns. V4 launch remains expected in late April 2026.