Chinese research labs, despite limited access to advanced GPUs and lithography equipment, maintain pace with US labs by likely distilling American models, rather than possessing superior training efficiency. This suggests that while Chinese models show impressive performance, their underlying training method does not reduce the overall demand for AI infrastructure, as similar GPU resources are still needed for inference. Chinese developers reportedly use VPNs to access foreign models like Claude Code due to insufficient domestic computing resources, despite local ASICs like Huawei Ascend.
Chinese AI model development relies on distillation from US models, meaning the global demand for GPUs for both training and inference remains high, regardless of local hardware access.