Huawei Rotating Chairman: Ascend Has Surpassed Nvidia in China Market Share - Geopolitechs

The "share reversal" claim is less about technical parity and more about structural bifurcation. Ascend's CANN stack still trails CUDA by two to three generations in operator coverage and developer tooling, but under export-control constraints, "good enough" has become functionally equivalent to "viable." This mirrors the 2019 ARM-licensing disruption that forced Chinese mobile SoC designers to build parallel ecosystems—not through superiority, but through necessity. The real cost sits deeper in the supply chain. SMIC's 7nm yield and cost structure mean Ascend's per-card TCO runs three to five times that of comparable nodes on TSMC's 4nm process. This margin compression will persistently squeeze Chinese LLM developers' training economics. Nvidia's rational response is not price competition; it is accelerating H20/B30 inference-optimized SKUs as compliance vehicles, effectively downgrading the China market from a profit center to a regulatory cost center. The 12-to-24-month tail effect is the critical variable: two mutually incompatible AI compute stacks will harden into permanent infrastructure. Developer migration cost shifts from a one-time expense to a perpetual tax. That is the structural wound to Nvidia that no single-quarter revenue figure can capture.