Fact-Check: Viral China-US AI Token Comparison Uses Incompatible Data
A viral post claimed that 55% of China’s AI token consumption was for short dramas and entertainment, while 58% of U.S. usage supported programming, R&D and enterprise automation. NoNoise traced the Chinese figures to Chinese Online’s investor-relations disclosure, which cited undisclosed industry-call data. The closest source for the U.S. figure was an OpenRouter and a16z report showing that programming and technical traffic each represented 29% of tokens from OpenAI models on OpenRouter at the end of 2025. That dataset covered only OpenRouter traffic for OpenAI models and large-language-model interactions, not overall U.S. AI use or video generation. No unified China-U.S. comparison covering token use across text and video models was identified.
The figures do not establish either country’s national AI-use patterns. They also illustrate that token consumption is not equivalent to computing-resource consumption, because models can require different amounts of computation to process the same number of tokens.
两年前，知乎上有个话题：为什么说「美国把AI用来娱乐，中国把AI用于工业生产」？彼时，美国的文生视频大模型Sora横空出世，其生成的视频几乎以假乱真，在全球范围内引发惊叹与焦虑。而中国的DeepSeek还没出圈，豆包尚未起势，文生视频大模型Seedance更是「查无此人」。两年后，Sora关停，Seedanc......
本文来自微信公众号： 降噪NoNoise ，编辑：小维，作者：刘诗雨
两年前，知乎上有个话题：为什么说「美国把AI用来娱乐，中国把AI用于工业生产」？
彼时，美国的文生视频大模型Sora横空出世，其生成的视频几乎以假乱真，在全球范围内引发惊叹与焦虑。而中国的DeepSeek还没出圈，豆包尚未起势，文生视频大模型Seedance更是「查无此人」。
#Chinese Online#OpenRouter#a16z#ByteDance#Sora#Seedance#Kling AI#OpenAI#Boston Consulting Group