DeepSeek Grants Huawei Exclusive Early Access to V4 — Memory Chip War Intensifies

Reports confirmed that DeepSeek withheld its next-generation V4 model from Nvidia and AMD for optimization, granting early deployment access exclusively to Huawei and Cambricon for inference hardware. In parallel, Huawei launched its UCM (Unified Cache Manager) system designed to reduce reliance on costly High Bandwidth Memory (HBM) chips — a critical bottleneck since China is denied access to HBM from South Korean suppliers Samsung and SK Hynix under US export controls. DeepSeek's architectural innovations, including reduced memory footprint requirements, align closely with Huawei's hardware roadmap. Analysts view the partnership as a strategic pivot: China is rebuilding its AI compute ecosystem to be self-sufficient by co-designing model architecture with available domestic hardware, rather than optimizing for unobtainable Nvidia chips. Each Huawei Ascend 910C delivers roughly 60% of an Nvidia H100's inference performance — but the DeepSeek-Huawei co-optimization is expected to narrow that gap for specific inference workloads.