AI-generated short dramas are drawing large volumes of tokens in China, but token use does not tell us how much computing hardware they consume. A BigGo Finance report brings together ByteDance’s rebuttal, industry figures and a less comfortable question: whether the booming production business is making money.
ByteDance AI Watch analysis
What happened
ByteDance public-relations chief Liu Xing said the figures behind claims that short dramas use more than half of China’s computing power were unreliable. He said video-generation models account for about 10% of ByteDance’s own computing power. Those are different measures: the report says a Chinese Online investor-relations record attributed 55% of domestic AI token consumption to short dramas and video generation. Token consumption is not a national tally of compute capacity.
BigGo Finance’s report also describes the economics behind the industry’s growth. It says AI short dramas can cost about 100,000 yuan to produce, compared with 400,000 to 800,000 yuan for a live-action title, while only 1.3% of new AI drama and comic-drama titles on Douyin in the first half reportedly reached the stated break-even threshold.
Why it matters
AI video can generate a hefty token bill without using the same kind of hardware, or serving the same customers, as high-end model training. Treating those numbers as interchangeable makes for a dramatic headline and a muddled picture of where compute is actually scarce.
The commercial question is just as important. Lower production costs may help explain the surge in output, but the report’s figures suggest that more titles do not automatically mean more profitable titles. Compute demand and entertainment success are not synonyms, however much the spreadsheets might like them to be.
Our read
The useful story is not that short dramas are either stealing all the compute or using none of it. It is that a big token-consumption figure cannot, by itself, settle who is using scarce hardware or whether the resulting content pays its way. BigGo’s figures offer a reported snapshot, not a complete accounting of China’s AI infrastructure or a settled measure of industry profitability.
What to watch
- Whether the source data behind the 55% token figure becomes available with its sampling and methodology.
- Whether AI drama producers can improve the share of titles that recoup their costs.
- Whether future reporting separates inference demand from high-end training compute more clearly.
Discussion spark: Should AI compute debates focus on token consumption, hardware use, or whether the work being generated earns enough to justify the resources?
Sources and evidence
not affiliated with or endorsed by ByteDance AI