Nvidia's grip on China's AI chip market has slipped to 34%, a figure that lands like a verdict on three years of U.S. export controls.

That single number, reported by Yahoo Finance this week, captures how thoroughly American policy has redrawn the competitive map. The 34% slice, attributed to ongoing Commerce Department restrictions on advanced AI silicon, marks a steep contraction in the world's second-largest AI market. For nearly a decade, Nvidia defined this space.

How did Nvidia lose two-thirds of China?

The road runs through three control packages. October 2022 rules capped interconnect bandwidth. October 2023 tightened performance thresholds and added licensing requirements. The December 2024 update pushed further, restricting HBM configurations and advanced packaging.

Each round shrank the legal product set Nvidia could ship into China without an export license. The company's H100, A100, A800, and H800 parts anchored the AI training stack across Chinese hyperscalers, state-owned clouds, and university research clusters. The 34% line tells a different story now.

Who picked up the slack?

Huawei. The Ascend 910B and 910C now anchor Chinese AI training clusters, with deployments reported across state-owned operators and major cloud platforms. Cambricon, Biren, and Iluvatar CoreX all saw order books swell through 2024.

Alibaba redirected inference workloads to its Hanguang silicon. Baidu leaned harder on its 昆仑芯 design. Smaller players including MetaX and Enflame have found niches in specific inference and training segments. The domestic stack is real. It is scaling.

What is Nvidia doing about it?

The company built the H20, a China-specific SKU designed to sit beneath the October 2023 performance thresholds. It worked, until December 2024, when the new rules effectively sidelined even the H20. A successor part, reportedly in development, faces the same gating question: what can clear U.S. export review?

Nvidia's leadership has framed the situation as a long-term engagement, emphasizing compliance and continued investment in unrestricted markets. The math is unforgiving. A 66-point share swing in a market this size reshapes R&D allocation, product roadmaps, and competitive positioning for years.

The financial impact is visible. China revenue has dropped from a meaningful slice of Nvidia's data center sales to a fraction of its former weight, a shift that has drawn multiple sell-side estimate cuts since December. Wall Street is recalibrating.

What changes the 34% line?

Three variables matter. Whether the current control framework survives administration turnover in Washington. Whether product-specific carve-outs emerge for H20-class silicon. How quickly Chinese foundries can scale 7nm and below capacity with acceptable yields.

The CUDA moat still holds. Chinese AI developers have spent years building on Nvidia's software stack, and that switching cost remains real. But domestic frameworks — Huawei's CANN, Biren's BiRT, Alibaba's open inference platforms — are narrowing the gap through aggressive optimization. Government procurement preferences accelerate the shift. State-backed subsidies lower the price barrier for domestic accelerators.

Foundry capacity remains the wild card. SMIC's 7nm yield reportedly sits well behind TSMC's, and advanced packaging capacity at Chinese OSATs remains constrained. These bottlenecks buy Nvidia time, but they don't reset the trajectory.

The 34% figure is a snapshot of a market still finding its post-control equilibrium. It will not be the last snapshot.