A percentage that changed units on the way round\n\nThe claim travelled well because it has a clean shape: China is spending its AI capacity on disposable entertainment while others spend theirs on research. In its circulating form, more than half of China's computing power goes to making short dramas.\n\nTrace it back and the shape changes. The underlying figures come from an investor-relations activity record disclosed by Chinese Online in June. In that reported breakdown, 55% of domestic AI token consumption was attributed to short dramas and video generation, 24% to e-commerce marketing, and 15% to software development. Those records were recently resurfaced by self-media accounts. The circulating compute claim resembles that June statistic about token consumption, but it substitutes a different measurement — and the available reporting does not establish the complete path by which the statistic became the viral claim.\n\nThe two measurements are not interchangeable. Token consumption is a measure of model input and output volume under a particular accounting method; it is not inherently a measure of spending, and it is not a measure of the computing resources consumed. Different model architectures and workloads can require very different amounts of computation per token: two workloads can post the same token count while one ties up far more GPU time than the other, because what each token costs to produce depends on the model and the task. Video systems may not even count in tokens at all — some are accounted per frame, per second or per generation. And the June disclosure does not explain how it made video-related token counts comparable with other applications in the first place.\n\nLiu Xing's response, reported on 19 October, makes the unit point from inside the company most associated with the claim. ByteDance runs Doubao and the Seedance video models, and distributes short dramas at scale. If any single firm's allocation were going to resemble the viral figure, it would be ByteDance's. Liu put training and inference for video-generation models, combined, at about 10% of the company's total compute. According to the Jiemian News reporting carried by TechFlow, that share also accounts for frontier research and parameter tuning of underlying vision foundation models — broader than finished dramas alone, though ByteDance has published no detailed workload breakdown.\n\n## What the rebuttal does not prove\n\nIt would be easy to swap one clean number for another, and that would repeat the original error. Liu's 10% is a statement by a public relations head about his own employer. It is not audited, and ByteDance has not published what counts as its total compute, over what period, or how training and inference were separated. A company defending its reputation is not a neutral source for its own allocation.\n\nThe scopes also differ in a way that cuts both directions. Chinese Online was describing, however loosely, a national pattern; Liu was describing one firm. ByteDance's 10% statement does not independently disprove Chinese Online's 55% token estimate, and a national token share cannot describe ByteDance's data centres. The two figures differ in measurement, population, workload and provenance; they cannot be subtracted, compared as a ratio, or used to invalidate each other directly. Both can be roughly right about their own narrow claims and still leave the actual question — how much of China's compute serves video generation — unanswered by either side. What the 10% does show is why the 55% estimate cannot be repackaged as a claim about national computing power.\n\nThere is a second gap in the June record worth keeping in view. In the reporting available, Chinese Online did not identify the underlying data source behind the percentages, the sample coverage or period they were drawn from, or the calculation method used. A figure with no stated method can be repeated indefinitely without ever being checkable. That, more than the entertainment angle, is why it survived from June to October and acquired a stronger claim along the way.\n\n## The practical reading\n\nVideo generation is a large and growing workload in China. Nothing in Liu's statement denies that, and in Chinese Online's reported breakdown, short dramas and video generation ranked first among the categories listed, at 55% of token consumption — the company's reported estimate, to be treated as such given the undisclosed methodology behind it, not as a verified measurement of China's AI usage. Short dramas are a real, high-volume use of the technology, as this site reported in October on the production economics behind them.\n\nWhat readers should retire is the compute version of the claim. If a statistic arrives as a share, ask what it is a share of, whose population it covers, and when it was measured. Here those three questions turn "half of China's compute makes dramas" into something smaller and more honest: a June investor-relations disclosure attributed 55% of domestic token consumption to video, without identifying its data source, coverage or method. That is a reported share of token consumption, not a measured share of China's computing capacity or GPU workload. A genuine national compute breakdown would need consistent workload definitions, a specified period, representative coverage and a comparable measure of computing resources. Neither side published one here.

Sources

  1. [1] TechFlow (English), "Doubao Responds to 'Half of Compute for Binge-Worthy Dramas': ByteDance Video Generation Compute Is Only About 10%", 19 October 2026, citing Jiemian News — Liu Xing's statement that the claim "lacks factual basis", the ~10% of ByteDance compute for video-generation model training and inference, and that the share also accounts for frontier research and parameter tuning of underlying vision foundation modelsRead source
  2. [2] PANews (English), "ByteDance responds to 'over half of China's computing power used to make short dramas'", 19 October 2026, citing IT Home — corroborating version of Liu Xing's statementRead source
  3. [3] ChainCatcher (English), "Doubao refutes that half of the computing power...", 19 October 2026 — corroborating version; dates the Chinese Online investor-relations record to June 2026Read source
  4. [4] BigGo Finance (English), "Doubao Clarifies 'Half of China's Computing Power Used for Short Dramas' Claim", 19 October 2026 — corroborating version; notes the two figures use different statistical scopesRead source
  5. [5] Related alfw.ai coverage (different event — production economics): "China's AI Drama Boom Has a 98.7% Casualty Rate", 5 October 2026Read source