Token-share figures are being used to size China's AI demand by sector; ByteDance's 10% compute figure suggests video's share depends heavily on whether tokens or hardware are counted.
ByteDance has pushed back against one of the most widely repeated numbers in China’s AI economy: the claim that AI-generated short dramas and video generation account for 55% of the country’s AI token use. Liu Xing, head of public relations for ByteDance’s Doubao AI assistant, wrote on Weibo on October 19 that the computing figure published by an unnamed company was unreliable, and that only about 10% of ByteDance’s total computing power goes to training and running its video-generation models. Caixin Global reported on October 21 that the company was challenging the figure, which originated with COL Group, a Chinese producer of AI short dramas.
The 55% figure gained weight in September, when the China Academy of Information and Communications Technology (CAICT), a state-backed research institute, carried the same share in its report on China’s token economy. A token is the unit of text, image or video that an AI model reads or produces, and providers bill and measure usage by it. By the time Liu posted, the line "half of China’s computing power is making melodramas" had become a trending topic, according to China News Weekly in an interview republished by ifeng.
The two figures measure different things. COL’s number describes industry-wide token consumption; ByteDance’s describes the share of its own hardware devoted to video models, as the Beijing Daily pointed out. Neither COL Group nor CAICT had publicly responded to Liu in the reports reviewed through October 21, so the gap between the two has not been reconciled.
A COL Group investor record started the 55% figure
COL Group first gave the breakdown in an investor relations activity record from a June 3 roadshow in Beijing and an online call, the Beijing Daily reported. The record put AI short dramas and video generation at 55% of China’s token consumption, e-commerce and marketing at 24%, and software development at 15%. It said token use in AI dramas and video generation had reached more than double that of software development, making it the largest single consumer in the token economy.
COL explained the heavy usage by the demands of production, according to Southern Metropolis Daily. Live-action AI dramas need more frames and 1080p resolution, and use roughly two to three times the tokens of 720p AI comic-style dramas, while teams regenerate shots repeatedly to satisfy directors. The company has its own stake in the trend: it said it released more than 1,200 AI short dramas in the first half of 2026, cut per-episode costs by 80% to 90%, and expects annual capacity of about 3,000 titles.
CAICT’s Token Economy Development Research Report (2026) (《词元（Token）经济发展研究报告（2026年）》) described AI dramas and video generation as the core application driving token consumption, with a 55% share that ranked first, the Beijing Daily reported. According to China News Weekly, CAICT’s report put China’s average daily token calls at 500 trillion in June 2026, up from 100 billion at the start of 2024. China News Weekly also cited an estimate that a single 1.5-minute AI drama episode can consume 5 million to 10 million tokens.
Token counts and hardware share answer different questions
Liu’s argument rests on how tokens are distributed across model types. In a second post, he agreed with a commenter that video-generation models in fact use fewer tokens than language models, because language models take in large volumes of input, especially in multi-turn conversations with AI agents, Southern Metropolis Daily reported. Doubao’s own scale makes that point concrete: its daily token calls passed 180 trillion as of June 2026, more than ten times the level a year earlier, according to China News Weekly.
Liu Zun, head of the DataEye research institute, told China News Weekly that heavy token use by AI dramas reflects the scale of the application and does not mean the sector is absorbing scarce high-end computing power. Most of that use happens at the inference stage, when a trained model generates output, often on general-purpose, mid-range chips. Research and foundation-model training draw on a separate supply of high-end GPUs with different buyers, hardware and costs. He added that the 55% figure covers inference-side tokens, including text that accompanies video, which makes a viral comparison with a 58% share for coding in the United States misleading.
Liu Zun did see one real effect. Large-scale token use by AI dramas pushes up demand and prices for inference capacity, he said, putting it in competition with other inference-heavy applications such as AI-generated online fiction.
The dispute puts pressure on how China counts AI demand
Token statistics have become a headline measure of China’s AI adoption, and EastFrontier has tracked CAICT’s work on inference demand and token factories. A sector-by-sector share first published by a company with a stake in AI dramas, and later matched by a state research body, now faces a public challenge from ByteDance, whose Doubao handles more than 180 trillion tokens a day.
The AI drama industry is also under scrutiny on other fronts, from regulators to audiences, as the viewer-fatigue test facing microdramas shows. The clearest resolution would come from CAICT publishing the method and sample behind its 55% share, or from COL Group saying how it reached the number. A breakdown of token use by model type from ByteDance or other large providers would also show whether video’s share holds up when hardware is counted alongside tokens.