Beyond the Claim That “55% of AI Compute Goes to Short Dramas”

# Beyond the Claim That “55% of AI Compute Goes to Short Dramas”

I came across an image online claiming that 55% of AI token consumption in China goes to short dramas and video generation, while nearly 60% of AI compute in the United States is used for coding, research and enterprise automation. It creates an immediate impression: China is using AI for short dramas and entertainment, while the United States is using it for programming and research.

But once we look more closely, the comparison raises several problems.

The Chinese figure is not entirely without a source. COL Digital Publishing stated in an investor-relations document that AI short dramas and video generation accounted for 55% of domestic token consumption, e-commerce and marketing for 24%, and software development for 15%. The document, however, does not adequately explain how a nationwide figure was calculated, which models and platforms were covered, or how tokens from different systems were made comparable. So the number has a source, but it should not be treated as an independently verified national statistic.

The claim that nearly 60% of US AI compute is devoted to coding, research and enterprise automation is even harder to substantiate. There is substantial American research on AI usage, including Anthropic’s studies of Claude, OpenAI’s research on how ChatGPT is used at work, and Stanford’s AI Index. These studies do support a broad conclusion that programming and enterprise activities are important uses of generative AI in the United States. They do not, however, establish the specific figure of 58%.

More important than whether those two numbers are correct is the way the comparison itself is constructed.

Short dramas, video, e-commerce, programming and research do not belong to a single consistent classification. A short drama is a content form. Programming is a technical activity. E-commerce is a commercial sector. Research is a form of knowledge production. Putting all of them into one table makes the categories look comparable when they are not.

A more meaningful distinction may be to ask what kind of social activity AI enters. Some AI use enters production: improving work efficiency, developing software, supporting research, designing products and managing businesses. Other uses enter consumption and entertainment: generating recreational content, chatting, gaming and everyday leisure.

Even that distinction has limits. A generated product demonstration or training video is part of production. A short drama is itself a cultural product and therefore also the result of productive activity. Programming is usually highly instrumental and productive, but software can also be built specifically for entertainment.

This leads to a second distinction. Entertainment creation is still creation and has economic value, but it differs from technical activities such as software development and scientific research. The latter can become tools that enter further production processes and affect productivity and technical capability across other industries. Entertainment primarily addresses consumption needs. There is no need to rank one as inherently superior to the other, but the difference matters when we are trying to understand AI’s effect on productivity and economic structure.

There is another practical difficulty: measuring AI use by country is inherently difficult.

Major American AI services are global. ChatGPT and Claude are used around the world. The number of tokens generated or the amount of compute consumed by an American company cannot simply be equated with AI use by Americans. A model may run in a US data centre while the user is in India, Europe, Australia or elsewhere.

A serious comparison between China and the United States therefore needs to specify what is being measured: the location of users, companies, computing infrastructure, or the legal domicile of AI providers. Without that distinction, “Chinese AI” and “American AI” may not even refer to comparable statistical objects.

Still, there is a real question worth studying.

Different societies have different distributions of knowledge, occupations and industries, and these differences may influence how AI is used. If only a limited share of a population works in scientific research, software development and other knowledge-intensive fields, it would not be surprising for a large proportion of overall AI use to fall into conversation, entertainment and everyday consumption. The weaknesses of the original statistics do not eliminate that possibility.

But it would also be too simple to say that users merely prefer entertainment. Users do not determine AI usage patterns on their own.

Products shape behaviour.

If an AI system is particularly good at conversation, responds naturally and provides an enjoyable emotional experience, people will naturally use it more for conversation. If video generation becomes extremely easy and platforms continually lower the barrier to creating short videos, entertainment production will expand.

The opposite is also true. If an AI coding tool repeatedly makes mistakes, requires constant correction and eventually costs more time than writing the code directly, many users will stop relying on it. The same applies to research, engineering, law, finance and other serious domains. The higher the cost of error, the greater the demand for reliability. If the model cannot meet that threshold, users will be reluctant to entrust important work to it.

What appears on the surface as “user preference” may therefore be the result of several forces operating together. Population and occupational structures influence who uses AI. Product design determines which functions are easiest to use. Model capability determines which tasks can actually be completed. Platform business models reinforce some patterns of use more strongly than others.

Seen this way, the interesting question behind the online claim is no longer whether the two headline percentages are correct.

The more useful question is which parts of social activity AI is entering. Is it mainly expanding entertainment and consumption, or is it becoming embedded in research, engineering, software development and enterprise production? And to what extent are those differences produced by individual choice, industrial structure, product design and the capabilities of AI itself?

That gets us closer to the real issue than simply saying that “China uses AI for entertainment while America uses AI for work.”

The analysis above follows a structural approach drawn from Sustenesis Theory. It does not infer a conclusion from one isolated statistic. Instead, it places AI use back into the relations that make the observed pattern possible. The knowledge and occupational structure of users, the industrial structure of society, product design, commercial incentives, and the capability and reliability of AI constrain one another and together sustain a particular distribution of use. Even if figures such as 55% or 58% were accurate, they would still describe an outcome rather than explain why that outcome emerged. Sustenesis does not provide a ready-made empirical answer here. It provides a way of analysing the problem through difference, constraint and the structure that those interacting conditions sustain.