news

ByteDance puts video generation at about 10% of its compute, against a 55% token figure it calls unreliable

Doubao's PR head posted the number on Weibo on October 19, answering an investor note from Chinese Online that put short drama and video at 55% of domestic AI token consumption.

By Priya Shenoy · 19 October 2026

ByteDance has now put a public number on how much of its machine capacity goes into making video, and the number is about 10%. Liu Xing, public relations head for the company's Doubao model, posted it on Weibo on October 19 in response to a claim spreading through Chinese media that more than half of the country's AI compute is being burned on short dramas. He said the compute data behind that claim was "somewhat unreliable" and that of ByteDance's total compute, only around 10% goes to the training and inference of video generation models.

The 55% figure counted tokens, not GPUs

The claim Liu was answering traces back to an investor relations activity record disclosed in early June by Chinese Online, a Shenzhen-listed publisher. In that document the company said that of domestic AI token consumption, 55% came from short drama and video generation, 24% from e-commerce and marketing, and 15% from software development, with AI short drama and video consuming more than twice what software development did, making it the single largest category. Republished through the summer, that turned into the shorthand that half of China's compute is producing crowd-pleasing serials.

The two figures do not measure the same thing. One is a survey-style estimate of token consumption across an entire domestic market, reported secondhand in an investor meeting note. The other is one company's internal split of its own compute pool, stated by that company's communications lead in a social post. A token count is a billing unit; a compute share is a hardware allocation. Moving between them requires assumptions about price per token, model size and utilisation that neither party published.

What the 10% has to cover

Liu's figure explicitly spans training and inference together, which means the cost of building ByteDance's video models sits inside the same slice as every clip actually delivered to a user. That slice carries Seedance 2.0 and Seedance 2.5, the consumer generation apps, and whatever research runs are training the next version. Read that way, 10% of a company that buys compute at ByteDance's scale is not a small commitment to video, and the number does not tell you how much of it reaches short drama production specifically.

The retail end of that compute is at least observable. Our price radar tracks what a 10 second 1080p Seedance 2.5 generation costs across the platforms reselling it.

Neither side has shown its working

Chinese Online has not responded to the challenge to its figure, so the public record on how much of China's AI capacity goes to generated video now consists of one investor meeting note citing industry research that was not published, and one executive's social post citing an internal split that was not broken down. There is no methodology attached to either, no time window, and no third party in a position to audit them.

That matters beyond a dispute over short dramas. Volume figures are doing heavy lifting in the current argument about whether AI video is a real industry or an attention bubble, and our earlier look at the Douyin pipeline found 221,900 new AI shows launched in six months. Counting output is comparatively easy. Counting the silicon behind it, so far, is an exercise in trusting whoever spoke last.

Sources: The Paper, Sina Finance, iFeng Tech.

Priya Shenoy: Tracks what AI video actually costs across the platforms that resell access to the same handful of models. Treats a pricing page as a claim, not a fact, until someone checks it.