OpenRouter data shows that the proportion of tokens used by U.S. companies from Chinese models has risen from less than 10% a year ago to nearly 60% by early July, peaking at 63%, with DeepSeek being a major driver. This indicates that domestic models are no longer just being tested by developers but have entered the actual workflows of U.S. companies by leveraging low costs, open weights, and sufficiently strong programming and agent capabilities. However, this does not mean that domestic models have captured 60% of the U.S. enterprise AI market. The statistics in the chart reflect the number of tokens on the OpenRouter platform and do not include traffic directly invoking providers like OpenAI or Anthropic. Additionally, lower-cost models are more likely to handle large-scale, long-chain tasks, which naturally amplifies their token share. This also suggests that companies are reserving high-end models for complex tasks while assigning a large volume of standardized work to domestic models that are 'good enough and more cost-effective.' AI competition is shifting from single-model rankings to multi-model orchestration and per-task cost efficiency.