Chinese AI labs shift model training to domestic chips amid US export curbs

Chinese AI labs are increasingly moving earlier stages of model training onto domestically produced chips, driven by US export controls and Beijing's push for self-sufficiency. While Chinese AI models have become competitive with US counterparts, hardware still lags. The three stages of AI development are pre-training (most computationally intensive), post-training (fine-tuning), and inference (running the model). Domestic chips are widely used for inference, but no top Chinese model has been pre-trained on homegrown silicon. Labs are now experimenting with shifting pre-training and post-training to local hardware. Economist Gary Ng of Natixis noted that while this may slow development relative to US labs, China is building a rare domestic AI supply chain. The article highlights five models using domestic chips across these stages.