China's effort to reduce dependence on Nvidia chips is encountering significant obstacles as the country pursues semiconductor self-sufficiency. While Huawei, China's leading domestic chip manufacturer, has achieved volume milestones, the shift in developer habits remains incomplete.
Huawei claims its Ascend AI chips have outsold Nvidia within China as of October 2026. The company has positioned its offering around reliable domestic supply rather than peak performance, emphasizing that an available chip outweighs a faster alternative that may be restricted. However, many of China's leading AI developers continue to rely on Nvidia hardware for training advanced large language models, the computationally intensive process of building AI systems from scratch.
DeepSeek illustrates this split approach: the company plans to deploy at least 160,000 Huawei Ascend 950DT chips for inference operations while maintaining Nvidia hardware for training workloads.
The Software Constraint
The stickiness stems from software infrastructure. Nvidia's CUDA platform is deeply entrenched across the AI industry as the toolkit developers use to write code that runs efficiently on Nvidia chips. Huawei's alternative platform, CANN, requires a transition carrying roughly 50% higher costs and time for training workloads. Inference deployments are rolling out with more flexibility, which explains how Huawei's sales can grow even as Nvidia retains dominance in performance-critical training applications.
Market Projections and Policy Context
Projections indicate domestic AI chipmakers could capture approximately 50% of China's AI accelerator market by 2026, while Nvidia's share may fall to around 8%, down from about 40% in 2025. These figures coexist with Nvidia potentially powering a disproportionate share of frontier model training.
US export controls enforce this environment. Washington has restricted advanced chip exports to China and arrested individuals involved in smuggling Nvidia hardware, with cases ongoing as of October 2026. The government has granted limited approvals for certain Nvidia models, creating a patchwork where some hardware moves legally, some illegally, and Chinese firms plan around both possibilities.
Strategic Implications
For Nvidia, the concern extends beyond short-term revenue to long-term competitive positioning in a historically lucrative market. For Huawei, inference deployment represents a meaningful foothold, but winning in training—where switching costs are highest and performance most critical—presents a steeper challenge and remains a source of Nvidia's leverage.
