China Releases 1,519 AI Short Dramas a Day, Industry Data Shows

Chinese business outlet Caixin reported that Shenzhen-listed Zhongwen Online disclosed 55% of China's AI token usage goes to short-form drama and video generation, a figure corroborated by the China Academy of Information and Communications Technology. Tokens are the basic units AI models process, serving as a measure of AI service usage and cost. Video generation inherently consumes significantly more tokens than text or images; for example, generating a five-second silent video at 1080p resolution with Kuaishou's Kling Video V3 model costs roughly 80,000 tokens per second, compared to about 20,000 for a single still image from a similar model. The iterative nature of video production, requiring clips to be regenerated until desired, further drives this high consumption.

This statistic led to self-deprecation among some Chinese internet users, suggesting a focus on entertainment over productivity. However, ByteDance's AI service, Doubao, countered that its video generation models use only about 10% of its total computing resources for training and inference. This discrepancy arises because Zhongwen Online's figure measures token processing, while ByteDance's reflects overall computing resource allocation. Training an AI model initially consumes vast computing resources but generates few user-facing tokens, meaning heavy token use does not always equate to a proportionally heavy draw on total computing power.

Despite the debate over measurement, the production of AI short-form dramas is experiencing rapid growth. The China Netcasting Services Association reported that over 95% of the 128,000 short-form dramas released in the first quarter of this year were AI-made. An additional 153,000 AI short-form dramas were added in the second quarter, bringing the first-half total to 275,000. Chinese media noted that 430,000 short-form dramas were newly released from January through August this year, a 13-fold increase over the previous year, a trend analysts attribute partly to ByteDance's Seedance 2.0 video generation model unveiled in February.