Ant Group has reduced AI costs by 20% by utilizing Chinese-made chips from Alibaba and Huawei, alongside AMD alternatives, while still incorporating some Nvidia chips. The company’s AI models, based on the Mixture of Experts technique, reportedly match Nvidia's performance, with notable models like Ling-Plus and Ling-Lite outperforming competitors in benchmarks. Training one trillion tokens with their optimized approach costs approximately 5.1 million yuan (US$880,000), significantly less than traditional methods, though challenges in model training persist.