Is half of China's AI compute being used to make AI junk content? The controversy over the 55% token figure

Is Half of China’s AI Making Slop?

The 55% token myth, and China’s guilty pleasureZilan QianOct 30, 2026

Zilan Qian is a research scholar at GovAI researching AI development and regulation in China.

English text is being prepared.

Why, in a society that places such a premium on productivity, has an industry devoted to manufacturing seemingly useless pleasure become so enormously popular? And what should we make of the shame attached to it? I argue that both answers lie in Chinese society’s obsession with usefulness, where technologies and people alike are judged by their productive value, so much so that resting can feel like a personal failing. AI slop offers an escape from that pressure, and the shame is the same pressure catching up with people even in their escape. Its shoddy production deserves criticism, but the pleasure people take in it should not have to justify itself by being useful.

Various short dramas available on Hongguo, ByteDance’s free short dramas app.“America uses gunpowder to make weapons; China uses it to make fireworks.”

English text is being prepared.

The controversy became so intense that ByteDance, maker of the top video-generation model Seedance 2.5, publicly challenged the figures, saying that video generation accounted for only around 10% of its own computing capacity.3 The People’s Daily commentary also rejected the simplistic comparison and enumerated China’s many applications of AI in scientific research and industrial production.How accurate is the 55% claim?

English text is being prepared.

Caixin also claimed that the number is consistent with the Token Economy Development Research Report (2026), published on September 30 by the China Academy of Information and Communications Technology (CAICT), China’s state-backed think tank, which states that “AI short dramas account for 55% of token consumption across the entire industry.” However, the report can no longer be found on CAICT’s website.

English text is being prepared.

Lu Xun’s essay on electricity is in fact about the “dual-use” nature of technology. In Shanghai, he observed, electric shocks were applied to prisoners again and again, until their teeth loosened and their nerves were dulled for good. As the city commemorated Thomas Edison, many praised the telegraph and the telephone without considering that the same electricity was being used to torture people: “The fortunate use electricity to cure illness and beautify themselves, while the oppressed suffer and die by it.” Only in the final paragraph does Lu Xun turn to how China put technologies to superstitious or self-destructive uses: gunpowder for firecrackers to worship the gods, the compass for feng shui, and opium consumed as food. Yet today’s AI debate has picked up only this last paragraph, reducing the perspective of the duality of technology into the anxiety over whether it is put to productive use. Uselessness, it seems, has become a greater worry than harm.

Electricity-powered lights lit up 1930s Shanghai. Source.

When Lu Xun wrote in 1933 that foreigners used gunpowder to make bullets while China used it to make firecrackers, China was in extreme crisis. Japan had seized Manchuria and attacked Shanghai, the Nationalists and the Communists were fighting each other, much of the population lived in deep rural poverty, and life expectancy was around 35 years. Today, China is far stabler and far richer, yet the sense of crisis persists. The same metaphor is now applied to the country’s AI tokens, and the same sense of obligation stigmatizes the hours ordinary people spend watching slop at the end of a long day. Perhaps one sign of a country no longer in crisis is being able to enjoy fireworks without apology. Fireworks, after all, do not have to become weapons to be worth making.

Zilan is grateful to Afra Wang and Alan Chan for their helpful feedback.

English text is being prepared.

Most Chinese articles do not disclose sources. The number is likely taken from Menlo Venture’s “2026: The State of Consumer AI” report, which actually states that AI takes over 58% of the coding projects.3

Although ByteDance claimed that video generation only accounts for 10%, that number is about computing capacity, whereas the 55% refers to the share of tokens output by models. The former measures the share of ByteDance's computing resources devoted to video generation, while the latter supposedly measures the share of tokens consumed by AI applications across China.4

The mention of Claude is interesting. Many Chinese developers do use Claude and other US models, often through so-called transfer stations, which give users an API address and key and then forward their requests to the real model upstream. Since Claude’s weights are not publicly released, genuine Claude tokens can only be generated on Anthropic’s own servers or authorised cloud providers outside mainland China. These tokens are therefore unlikely to show up in China’s own token consumption records – which may partly explain why software development looks so small. (Some transfer stations do secretly swap in cheaper domestic models while labelling them as Claude, so not all of these tokens are generated abroad.)5

English text is being prepared.

For an example, see 36Kr’s report on micro and short drama titled “I don’t understand it, is this the newest ‘tittytainment’ for young people?”; and another Chinese media report from 2021 titled “be alert to ‘tittytainment’! How short videos are ‘killing’ your life”.7

English text is being prepared.

A guest post by

Is half of China's AI compute being used to make AI junk content? The controversy over the 55% token figure

Topics

Tags

Body · AI translation