字节 Says Video-Generation Models Use About 10% of Its Compute Capacity

豆包 public relations chief 刘星 posted Monday that training and inference for video-generation models account for about 10% of 字节’s total computing capacity.

That company-level percentage is separate from the 55% domestic AI token-use estimate. One measures computing capacity; the other tracks token consumption. Neither indicates what share of China’s total computing resources goes to video generation.

One tally of domestic AI token use assigned 55% to short dramas and video generation, 24% to e-commerce advertising, and 15% to research and development. In that tally, AI short dramas and video generation consumed more than twice the tokens used for software development.