China’s AI models are closing the gap with overseas rivals on a different cost curve

Summary

Chinese AI model providers have reduced training costs by approximately one-tenth and foreign API prices to 10-20% of their original levels while maintaining 20-40% gross profit margins, driven by smaller MoE architectures, higher GPU utilization rates, and declining infrastructure costs. The article argues that this shift is steering enterprise procurement from raw benchmark performance toward cost-efficiency per token.

Why it matters

If enterprise customers evaluate AI based on return per token rather than peak capability, China's cost curve becomes a structural commercial moat rather than a temporary discount. The open-weight ecosystem (DeepSeek, Zhipu, Moonshot) amplifies this effect by allowing the entire industry to build upon the same efficiency gains. This could reshape global model procurement in high-volume, repetitive workflows where a 5-10x price difference is decisive.

Highlights

• MoE activation: Chinese models activate approximately 1-10% of parameters per task, whereas some US competitors are estimated to activate 15-30%.
• Leading Chinese providers have pushed GPU utilization rates above 70%, surpassing the industry norm of 40-50%, thanks to timing advantages.
• According to UBS estimates, Chinese APIs maintain gross profit margins in the 20-40% range, even when priced at 10-20% of foreign list prices.

Critical take

The article relies heavily on UBS estimates and directional claims without specifying the exact benchmarks or workloads where the cost difference is genuinely valid. The claim of 'closing the gap' is asserted rather than proven. Furthermore, the article acknowledges that the binding constraint is inference capacity; this implies that the price advantage could evaporate the moment demand exceeds compute power. The framework also conflates training cost savings with inference economics, where the bulk of enterprise spending actually occurs.