字節跳動 PR Lead Questions Claim AI Dramas Use Half of China’s AI Compute

字節跳動’s 豆包 public relations lead 劉星 questioned whether a reported 55% share of China’s AI token use for short dramas and video generation can show the country’s share of computing capacity.

A June breakdown put short dramas and video generation at 55% of AI token use, e-commerce marketing at 24% and software development at 15%. The covered companies and platforms, sample size and calculation method are unclear, so the 55% figure should not be treated as a verified nationwide measure.

The figure has been used to support claims that short dramas consume more than half of China’s AI computing capacity. 劉星 questioned the data’s reliability and whether it could support a claim about China’s AI computing capacity, warning that token consumption is not a direct measure of computing power.

劉星 put training and inference for video-generation models at about 10% of 字節跳動’s total computing power. That estimate includes model research and development, and the share is not all used to produce short dramas.

The two percentages describe different things: one is a reported share of token consumption with an unclear scope, while the other concerns computing power at a single company. Compute use can vary with model size, computing methods, hardware efficiency and task type. A token share alone therefore does not establish how much of a country’s total AI computing capacity a particular application uses, and the company estimate does not represent the national industry.

AI tools are used in several parts of short-drama production, including scriptwriting, virtual filming, editing, tagging, dubbing, translation and promotion. Video teams may generate and screen scenes repeatedly to meet requirements for character appearance, scene continuity, movement, camera work and narrative pacing. Those repeated attempts can add to token use, but the token total does not by itself show the computing capacity used across China.