Is "US AI boosts productivity, Chinese AI creates bread and circuses" a false premise?

By She Zongming (Shell Finance Qianfan Jingfa Creator)

The trigger was this: a listed company that has bet heavily on the AI short-drama track disclosed in its recently released "Investor Relations Activity Record" that in China's Token consumption structure, short videos and AI short-drama generation scenarios account for as much as 55%, e-commerce advertising content generation accounts for 24%, while core productivity scenarios such as technology R&D, code development, and model iteration account for only 15%.

In other words, the Token consumption of AI short dramas and video generation has reached more than three times that of software development, making it the largest core consumption scenario in China's Token economy.

In contrast, reports say that in the United States, about 58% of Tokens flow to code generation and enterprise automation.

Comparisons are what people dread most. Once the chart comparing China-US Token consumption structures came out, the effect was comparable to "America's AI godfather Altman vs. China's AI godfather Li Yizhou."

So public opinion exploded, and many people said: this is equivalent to the West using gunpowder to make guns and cannons, while China uses gunpowder to make fireworks.

Some even asked in anguish: if computing power is all burned on short dramas, what will China rely on to win the China-US technology competition?

This reminds me of what Le Bon said: what influences the imagination of the masses is not the facts themselves, but the way they occur and attract attention. If I were to state my view, I would say they must be condensed and processed before they form a startling and astonishing image. Whoever masters the art of influencing the imagination of the masses also masters the art of ruling them.

I am not saying this comparison has no warning value at all; I am only saying that before getting indignant, it is more important to sort out the facts first.

01

First, three premises: 1. The claim that "55% of China's Tokens are used for short dramas, while 58% of America's Tokens are used for code" has no authoritative source and has not been rigorously verified.

The former comes from a listed company that is heavily betting on AI short dramas, so you can judge its credibility for yourselves; for the latter, no such "argument" can be found. a16z, McKinsey, and Gartner have all never said it, so it is not rigorous industry survey data by any measure.

2. Token usage and computing power usage are two different concepts. Saying that "most of China's AI computing power is used for short dramas, while most of America's AI computing power is used for code" is hard to call scientific.

Tokens are the billing unit for large-model invocation volume, while computing power is the computing capability of GPU hardware per unit of time. The computing power consumed per Token is not equivalent across different tasks.

3. The share of Tokens for code generation on US API platforms is indeed very high, but a large number of those are contributed by Chinese programmers. As far as I know, many programmers at major domestic companies use Claude to write code. These are still recorded in America's Token consumption ledger.

So using the glaring "15% : 58%" to judge how hard Chinese and Americans are investing in writing code may be thinking too simply, and ignores the statistical misalignment that comes with cross-border use.

Saying all this is to say that "55% of China's AI computing power is used for short dramas, while 58% of America's AI computing power is used for code" may be a straw-man proposition. Although it is very good at stirring emotions, it is most likely a fictitious target.

Some may say: setting the facts aside, is the phenomenon of a large number of domestic Tokens being used to produce mindless short dramas not worth worrying about?

Of course it is. But even if we follow that line of discussion, it is still necessary to clarify two points—

First, AI short videos and short dramas becoming the core Token consumption scenario in China is a structural result caused by the excessively high computing-power consumption density of video generation scenarios, not direct evidence that the whole society's AI resources are excessively flowing into entertainment and neglecting technology R&D.

Second, public opinion praises "American AI as emphasizing R&D and productivity" and disparages "Chinese AI as emphasizing monetization and quick, short, and flat content." This objectively does exist. In essence, it is a continuation of the phenomenon that "the American internet emphasizes basic R&D, while the Chinese internet emphasizes application scenarios." But before judging from a moral high ground, one must first see that this is a differentiated path inevitably formed by the different development stages and market environments of the Chinese and American AI industries.

Sounds awkward, right? Let me explain in detail.

02

The public's misunderstanding of AI computing power allocation may stem from a common cognitive blind spot: equating Token consumption volume with the weight of computing power investment, without knowing that there is a huge gap in Token consumption density between video and code generation.

Many people think that if more Tokens are used for short dramas than for writing code, then someone must have casually allocated more AI computing power resources to those domineering-CEO short dramas.

That really is not the case. Generally speaking, video generation is a typical "high-density Token consumption scenario," while code generation is a "low-density, high-efficiency Token consumption scenario." The difference in consumption density between the two is like heaven and earth.

How big is it? There is common industry calculation data to prove it: a 10-minute high-definition AI short drama, including the entire process of image rendering, dynamic special effects, subtitle dubbing, shot switching, and other generation work, requires 80 million to 120 million Tokens throughout.

All those pixel-level modeling, lighting iteration, dynamic correction, and the redundant computation of repeated fine-tuning, image quality optimization, and content error correction all burn Tokens. If we say AI short dramas are Token gold-devourers, who supports it and who opposes it?

By contrast, in AI-assisted R&D scenarios, an 8-to-10-person technical R&D team relying on AI tools to assist with code writing, vulnerability detection, feature iteration, and framework optimization consumes only about 3 million to 5 million Tokens in a week of high-intensity R&D.

What needs to be seen is that here, video Tokens and text Tokens are two different pricing units, and they cannot yet be directly compared as equal-weight numerical values.

Mistakenly equating the share of Token consumption with the share of computing power investment, and deriving from "55%:15%" that China's AI industry is "heavy on entertainment, light on R&D, chasing short-term profits, and lacking long-term vision," may not withstand the fact-checking and debunking of science popularization bloggers at all.

Moreover, even when compared horizontally under the existing calibers, the Token consumption used for software development in China and the United States may not necessarily be that vastly different.

At present, China's absolute volume of Token consumption already ranks among the top in the world. According to public information from the National Data Administration, as of March this year, domestic daily average Token calls exceeded 140 trillion.

Even if the software development scenario really accounts for 15%, under an extremely large overall computing power base, the corresponding scale of computing power investment is not small, and the gap in absolute volume corresponding to the United States' "58%" is not as large as some people imagine.

What's more, the share of Token consumption by Chinese models on the OpenRouter platform (about 60%) has far exceeded that of US models. The "State of AI" report jointly released by OpenRouter and a16z shows that on this platform, the share of Tokens for programming tasks rose from 11% in early 2025 to more than 50%, becoming the platform's largest single usage category... Taken together, the conclusion that "the software development scenario accounts for only 15% of China's Token consumption" is still too much for the math class representative's brain to handle.

03

I want to repeat a common-sense point: commercialization is not the opposite of productivity; it is the blood transfusion pump of productivity.

Looking at the history of the global technology industry, the steam engine was first used to pump water from coal mines, not for the power innovation of the Industrial Revolution; electricity was first used for lighting and entertainment, not for industrial production... Commercialization has always been an unavoidable stage for frontier technologies to achieve large-scale popularization and complete capital and technology accumulation.

Many people will say: how come American companies can go all in on hard tech, while you use resources to produce electronic dopamine?

It is true that the US AI industry uses a large amount of computing power for frontier model R&D and application software development, but this is supported by a mature hard-core technology investment and financing system and a special path for transforming technological barriers: relying on top-tier venture capital and long-term funding from tech giants, companies like Anthropic and OpenAI burn Tokens with a lavishness no less than the style of some Jiahao in the Hangzhou skyline or in the in11 yacht cabin.

It is also true that some Chinese companies use AI to improve productivity while others use AI to manufacture "tittytainment," but this has its inevitability: the primary problem facing most Chinese AI companies is "survival." Only when the commercialization chain runs through can the AI industry go further in a positive cycle where model vendors, cloud vendors, and application vendors all benefit.

Behind AI short dramas is a huge industrial chain connecting content parties, platform parties, model vendors, cloud vendors, and CDNs, and every link uses Tokens. Taking AI short dramas as an important implementation scenario can also be regarded as conducting an ultra-large-scale commercialization stress test: it can prove that China's model capabilities can support high-frequency, high-concurrency, high-iteration industrialized content production, and the AI capabilities honed by the multiple tests of "high frame rate, consistency, and draw convergence" in video generation can also be used in more fields such as industrial simulation, digital twins, and film and television previsualization.

This can be compared to what games meant to Nvidia. Just imagine: if the revenue and large-scale engineering validation provided by the gaming scenario had not existed back then, could Nvidia's GPU "big tree" have grown as thick as it is today?

"The United States uses AI to improve productivity, and China uses AI to both improve productivity and create tittytainment"—this is inherently blameless. US enterprise-level AI started early, its software ecosystem is mature, and enterprises have a strong willingness to pay, so programming and R&D naturally scaled up first; China has the world's largest content consumption market and e-commerce market, and the short drama and livestreaming industrial chains are already mature. After AI video technology is implemented, it is the market's choice that it first achieves scale, not proof of "degeneration."

04

I think this at least reveals two increasingly common social perceptions.

First, the belief that the vast majority of short dramas are electronic garbage with no nourishment.

Data shows that in the first eight months of 2026, 430,000 micro-short dramas were launched in China, 13 times the total for all of last year, with more than 90% involving AI in production. The AI content has increased sharply, but the formulas are still the same: domineering CEOs, live-in sons-in-law, rebirth, revenge, golden fingers...

Although I have always opposed netizens who "open their mouths with a theory of entertainment to the death, and then open them again with a theory of tittytainment," labeling short dramas as spiritual opium and cyber addictive drugs, I have to say that more than 90% of short dramas today are "rich in thrill, deficient in value"—even though they have already become a pastime for Chinese people to kill time and a new calling card for cultural going global.

To borrow the words of my friend Xiao Chang, if film is adding to meaning—even boring moments must be given meaning through an accumulation of symbols—then short dramas are anti-meaning: even weighty matters must, in the removal of meaning, be reduced to nothing but thrill.

If one considers that the hit rate of AI short dramas/AI comic dramas exceeding 100 million views across the entire internet is only 0.117% (monitoring data from DataEye-ADX as of the end of February 2026), and that in 2026 the average completion rate of the AI short drama industry has fallen from 45%-60% in 2025 to less than 15%, with a large number of AI short dramas unable to make money, there are actually considerable problems with the health of the industry's business model.

English text is being prepared.

Given that domestic chip supply is constrained and high-end GPUs are scarce, if tense AI compute resources flow excessively into AI short dramas with “short payback cycles,” they will inevitably squeeze the “token quota” for hard tech. This is a worry deep in countless people’s hearts.

As an ordinary citizen, I very much understand this way of thinking, but as a market fundamentalist, I find it hard to agree with that kind of “planned thinking.”

What I would rather see is this: those who think AI compute should not be wasted on short dramas should themselves watch fewer short dramas. If you don’t watch and I don’t watch, its market will shrink, and fewer people will use AI to make short dramas. Those who believe AI compute should be used for software development should first form the habit of paying for subscriptions to use AI software—if the business model can be made to work through user payments, companies will naturally have more incentive to use more tokens for software development.

About the author: She Zongming, head of Digital Force Field and senior researcher at the Digital Economy Think Tank

1 Vitality China Research Tour | River-sea access, rail-sea coordination! Unlocking Guangxi’s “seaward economy” code 46,000

2 The 166th anniversary of the destruction of the Old Summer Palace, free opening for the 11th year 30,000

3 Dalian hit by strong winds, large trees uprooted and crashing onto cars, residents struggling to move forward in the wind 21,000

4 Double Ninth Festival arrives, time-honored brands’ longevity peaches, Double Ninth cakes, and Double Ninth pastries sell well 16,000

5 Capital University of Economics and Business School of Statistics and Data Science: 70 years of expanding boundaries, using data to answer the questions of the times 13,000

Latest

Did an agent try to bypass restrictions to access Amap information? Tencent responds 13 minutes ago

How can capital and art coexist after the arrival of AI? Listen to what this forum says 14 minutes ago

2,700 yuan ruins a store; don’t let ratings harm the lively atmosphere of dining | Beijing News column 31 minutes ago

45-year-old Beijing Marathon improves a little every year, warmly supporting every runner 34 minutes ago The Third National News Media Pickleball Invitational concludes in Hengshui, Hebei 34 minutes ago

Hot discussion

Video | 2026 Beijing Marathon starts today!

Video | 2026 Beijing Marathon starts! More than 30,000 runners set off from Tiananmen Square