China's leading artificial intelligence laboratories are continuing to rely on Nvidia hardware for training their largest models, according to a South China Morning Post report, as the economics of switching to domestic Chinese chips remain stubbornly unfavorable for cutting-edge AI development. The finding underscores how years of US export controls and massive state investment in indigenous chipmaking have yet to break Western technology dominance in the most compute-intensive AI applications.

Why Domestic Chips Face an Uphill Battle

The core issue isn't that Chinese semiconductor alternatives don't exist—Huawei's Ascend lineup and other domestic processors have made genuine progress—but rather the total cost of ownership when it comes to training frontier-scale models. Industry sources indicate that performance-per-dollar ratios for local chips lag behind Nvidia's H20 and previous-generation Hopper architecture in scenarios requiring massive parallel compute clusters. For companies optimizing training efficiency at scale, the price premium for sticking with proven infrastructure often beats waiting for domestic alternatives to mature.

Export Controls Create a Complicated Landscape

US restrictions have progressively tightened access to advanced American chips, pushing Chinese firms toward gray market channels and existing stockpiles of previously purchased hardware. The H20 chip—Nvidia's China-specific offering designed to comply with earlier export rules—has become particularly significant in this context, representing a middle ground between cutting-edge Western technology and the constraints imposed by trade policy. Companies have been reluctant to abandon these systems given the uncertainty around domestic replacements.

Infrastructure Implications for AI Builders Worldwide

The situation highlights uncomfortable realities about semiconductor supply chains that extend beyond geopolitics into practical engineering decisions. When the world's largest AI development market still gravitates toward a single vendor's architecture, it reinforces ecosystem lock-in effects that make switching costs even higher over time. For infrastructure teams planning hardware investments, the Chinese case demonstrates how deeply embedded Nvidia's software stack and optimization patterns have become across global AI development practices.

Key Takeaways

  • China's top AI labs continue using Nvidia H20 and older Hopper chips despite export restrictions
  • Domestic alternatives like Huawei Ascend face performance and cost gaps at frontier model scale
  • Total cost of ownership, not just chip pricing, drives infrastructure decisions in AI training
  • The transition to local hardware faces years of friction even with strong government backing

The Bottom Line

This isn't really a story about export controls failing—it's a story about how hard it is to engineer your way out of an economic reality. Until domestic Chinese chips can match Nvidia's training efficiency at scale, expect the dependency to persist regardless of policy pressure.