China’s AI models run inference on domestic chips, but top systems still rely on Nvidia-class hardware for pretraining

Chinese AI models are increasingly competitive with U.S. peers, but the country’s hardware base remains a bottleneck, with domestic chips used widely for inference while leading models are not known to have been pretrained on homegrown silicon. The main constraint is a gap in computing power and supporting software ecosystems that makes full-scale training difficult on local hardware. Tighter U.S. export controls are pushing Chinese labs to try moving earlier training phases onto domestic chips, increasing pressure to close the training gap. The shift does not erase Nvidia’s current lead, but it underscores geopolitical supply risk and longer-term substitution pressure, with spillover to traditional assets largely limited to sentiment and valuation expectations.