Huawei and DeepSeek team up on AI chip design to rival Nvidia - Huawei Central

This is not a domestic-substitution story. It is a structural re-architecture of the AI compute stack. The critical insight: DeepSeek's R1 proved algorithmic sparsity can partially offset raw FLOPS deficits, but data-center-scale inference economics demand hardware-level tensor alignment. Huawei's Da Vinci architecture will now be redefined around DeepSeek's sparse-attention operators, pulling upstream demand onto SMIC's 7nm yield curve, CXMT's HBM3 ramp, and domestic EDA iteration cycles. On compliance: this sits within current BIS restrictions but will accelerate Washington's expected second-wave tightening on advanced packaging (ASE's China-Taiwan fabs), HBM supply, and EDA licensing. Nvidia's likely response is dual-track: fast-tracking a Blackwell China-specific SKU while lobbying for tighter export controls to protect its domestic install base. AMD's MI300X window in China is closing fast. 12-24 month call: the AI chip market bifurcates into two parallel ecosystems. China's stack trails by 2-3 generations in peak throughput, but the inference cost-per-token gap compresses below 30%. The CUDA moat is being eroded by full-stack convergence, not by any single silicon node.