Dispute Over China's AI Computing Power Consumption Structure: Conflicting Data on Short Drama's Share
nashnova research
Chinese Online claims 55% of China's AI compute goes to short dramas and video generation; ByteDance fires back that video accounts for only about 10% of its resources — the gap exposes a fundamental measurement clash and a deeper question about AI monetization.
55% versus 10% — how can the same thing look so different?
Chinese Online (中文在线) told investors that 55% of China's AI compute goes to short dramas and video generation, 24% to marketing, and just 15% to software development. Caixin, South Korea's *Korea Economic Daily*, and CAICT all cited the same figures.
ByteDance's AI service Doubao (豆包) pushed back on Weibo: training and inference for video-generation models together account for only about 10% of its total compute resources.
This means → neither number is "wrong" — they measure entirely different things, and comparing them directly is like comparing weight to volume.
Why do the numbers clash?
Chinese Online counted tokens processed — the discrete units AI uses to handle information — weighted toward the output side. Whoever generates more content looks bigger.
ByteDance cited its share of total compute resources — every GPU-hour spent on training and inference, added up.
In plain terms = one measures "how many finished goods rolled off the line"; the other measures "how many production lines the factory ran." Pre-training consumes vast compute but produces very few visible tokens, so a token-share metric systematically overstates video applications.
Chinese Online's data also came from a single AI short-drama operator's Q1 2026 internals plus external estimates — whether it represents the national picture is itself in question.
How much compute does short video actually consume?
Measurement disputes aside, nobody denies that short-video production burns through massive compute.
Per Tencent Token Hub data: Kuaishou's Kling Video V3 uses roughly 80,000 tokens to generate a 5-second, 1080p silent clip — and creators typically need multiple takes. A single image costs about 20,000 tokens.
ByteDance's Volcano Engine disclosed that Doubao's daily token usage hit 180 trillion in June 2026, up more than 10× year-on-year. This reflects a simple fact: whether the share is 55% or 10%, the absolute consumption growth rate is enormous on its own.
All these AI short dramas — is anyone making money?
Per Yicai and DataEye, roughly 221,900 AI and animated short dramas launched on Douyin in H1 2026. Only 1,055 (0.48%) crossed 100 million cumulative views; about 80% stayed below 1 million.
Fewer than 1.3% broke the 50-million-view mark — widely regarded as the break-even line.
This means → the vast majority of AI short dramas cannot recoup their costs. Roughly 90% of producers are losing money; some have already started laying off staff and exiting the market.
Where did the money AI saved actually go?
AI did cut filming and editing costs — but most of the savings shifted to traffic acquisition and marketing, which now eat up 80–90% of revenue.
In plain terms = production got cheaper, but getting an audience to see the work got more expensive — the savings just moved from one pocket to another.
This reflects the deeper issue beneath the measurement debate: when 90% of participants in AI's largest compute-consumption scenario are losing money, the path to monetizing AI compute remains an unresolved structural question.
市场有风险，内容仅供研究参考，不构成投资建议。