US export restrictions on advanced Nvidia AI chips to China are forcing global technology companies to duplicate their AI infrastructure investments.1 The restrictions prevent sale of high-performance GPUs including the H100 and A100 series to Chinese entities.
Huawei is positioning its Ascend 910B chip as the domestic alternative, with plans to ship approximately 750,000 units in 2026.1 Major Chinese tech firms ByteDance and Alibaba have become Huawei customers for AI accelerators.1 This marks a direct challenge to Nvidia's dominance in AI training hardware.
China launched a comprehensive domestic semiconductor development campaign in response to US sanctions.1 The initiative aims to reduce dependence on American chip technology across the AI hardware stack. Huawei Technologies has emerged as the primary competitor to Nvidia within China's borders.1
The parallel ecosystems create financial strain for multinational companies. Firms with operations spanning both regions must now procure, integrate, and maintain two distinct hardware platforms. This duplication extends beyond chips to encompass different software frameworks, development tools, and training pipelines.
The architecture divergence has implications beyond immediate hardware costs. Models trained on Nvidia infrastructure may require modification or complete retraining for Huawei chips due to differences in computational architecture and memory hierarchies. Companies face decisions about whether to maintain unified global AI systems or accept regional fragmentation.
The split also affects the AI middleware and tooling ecosystem. Developers are building region-specific optimization libraries, monitoring systems, and deployment frameworks tailored to each hardware platform. This creates parallel investment streams in software infrastructure that would otherwise be shared globally.
For companies committed to both markets, the question is no longer whether to invest in duplicate systems but how to minimize the resulting operational overhead and performance gaps between regions.
Sources:
1 Source hypothesis data (2026-03-30)