Data from OpenRouter shows that the percentage of tokens used by US companies employing Chinese models has risen from less than 10% a year ago to nearly 60% in early July, reaching as high as 63% at one point, with DeepSeek being a major driver. This means that domestically developed models are no longer just for developers to try out, but are entering the actual workflows of US companies due to their low price, open weighting, and sufficiently strong programming and intelligent agent capabilities.
However, this does not mean that domestically developed models have captured 60% of the US enterprise AI market. The chart shows the number of tokens on the OpenRouter platform, excluding traffic from companies directly calling vendors like OpenAI and Anthropic; at the same time, lower-priced models are more likely to handle large volumes of long-chain tasks, naturally amplifying the token share.
This also means that companies are reserving high-end models for complex tasks, while entrusting a large amount of standardized work to domestically developed models that are "capable enough and lower cost." AI competition is shifting from single-model rankings to multi-model scheduling and unit task cost.
