American AI is expensive. Some startups are turning to cheap Chinese models

American AI is expensive, and some startups are shifting to lower-cost Chinese models to control rising inference costs. In San Francisco, Lindy.ai said Anthropic’s top-tier models became its No. 1 expense, exceeding payroll and other operating costs across more than two dozen employees. The company reported moving 100% of its traffic last month to DeepSeek-V4, citing a “10x cheaper” cost that saved it millions. The broader backdrop is that U.S. firms are still leading “frontier” AI development, but experts say Chinese models lag by roughly six to 12 months in capabilities. Still, China dominates open-source options that can be freely downloaded and adapted. Companies such as Airbnb, Uber, and Perplexity have discussed using models like Alibaba’s Qwen, while intermediaries such as Featherless provide access to tens of thousands of models.