Why Chinese AI Models Are 95% Cheaper — The Economics Explained

Short summary

Chinese AI models cost 90–95% less than Western counterparts, not due to temporary discounts but structural advantages: MoE architecture activates only 5% of parameters versus 100% for dense models, cutting inference costs ~46x. Training innovations (data curation, FP8 precision, multi-token prediction) reduce costs 3–5x further. Lower hardware costs (H20 vs H100) and researcher salaries (4–7x cheaper) complete the picture. DeepSeek V4 matches GPT-4o performance at one-tenth the price.

• •MoE architecture activates only 5% of parameters per token vs 100% for dense models, cutting inference costs ~46x
• •Training efficiency gains from data curation, FP8 precision, and multi-token prediction reduce costs another 3–5x
• •Lower hardware costs (H20 GPUs) and researcher salaries (4–7x cheaper) complete the economic advantage

Generated with AI, which can make mistakes.