A statistic showing that more than half of China's artificial intelligence token usage goes to producing short-form dramas has stirred debate. As self-mockery spread that the country is pouring AI into entertainment rather than productivity, rebuttals emerged arguing that the figures were measured on different bases.
According to Chinese business outlet Caixin and others on the 21st, Shenzhen-listed Zhongwen Online disclosed a breakdown of China's AI token usage by purpose in materials prepared for an investor briefing in June. Short-form drama and video generation accounted for 55%, the largest share, followed by marketing such as e-commerce and advertising at 24% and software development at 15%. The China Academy of Information and Communications Technology carried the same figures in a research report on the token economy released last month.
Tokens are the basic unit by which AI models process information, serving as a gauge of AI service usage and cost. Video consumes far more tokens than text or images. According to TokenHub, an AI model service platform, generating a five-second silent video at 1080p resolution with Kuaishou's Kling Video V3 model costs roughly 80,000 tokens per second. Producing a single still image with a model in the same family requires about 20,000. Because clips must be regenerated repeatedly until the desired scene emerges, there is little dispute that short-form drama production burns through tokens in bulk.
The controversy flared as the statistic circulated in the form of a claim that half of China's computing resources go to short-form dramas. When a comparison was added showing that the U.S. devotes 58% of its tokens to coding and work automation, Chinese internet users responded that "in the age of AI, we Chinese have turned gunpowder into fireworks" — self-deprecation that while the U.S. works on raising productivity with AI, China is making amusements.
In response, Doubao, ByteDance's AI service, countered that "the share of ByteDance's total computing resources used for training and inference of video generation models is only about 10%."
The two figures differ because they measure different things. Zhongwen Online used the number of tokens processed by AI models, while ByteDance based its figure on how total computing resources are allocated. Training an AI model for the first time consumes enormous computing resources but generates almost no tokens delivered to users as output. The point is that heavy token use does not necessarily mean a correspondingly heavy draw on computing resources.
Whether the 55% figure represents China as a whole is also unclear. Zhongwen Online is a company expanding its own AI short-form drama business, and credible third-party statistics objectively tallying token usage by purpose remain scarce. Zhongwen Online said the statistic "was calculated for the first quarter of this year by combining token usage generated in our own AI content production process with external research reports."
Setting the dispute over figures aside, AI short-form drama production is growing rapidly. According to the China Netcasting Services Association, of roughly 128,000 short-form dramas released in the first quarter of this year, 122,000 were made with AI, a share exceeding 95%. Another 153,000 AI short-form dramas were added in the second quarter, bringing the first-half total to 275,000, or an average of 1,519 a day. Chinese media reported that 430,000 short-form dramas were newly released from January through August this year, 13 times the total for all of last year. Analysts say Seedance 2.0, the video generation model ByteDance unveiled in February, accelerated the trend.
The market is also expanding. According to a report released in April by the development research center of China's National Radio and Television Administration and others, the global short-form drama market was worth about $18 billion last year, of which China accounted for more than $14 billion, or 77.78%.