55% of China’s tokens are burned in short dramas, and 58% in the United States are burned in codes: the flow of tokens exposes two “survival methods” of AI

Skye has something to say

2026.11.11 14:35

· From Beijing

Full text 5664 words

Text | Skye has something to say

Regarding the difference in AI computing power consumption between China and the United States, there is a frequently quoted statement: about 55% of China’s tokens go to AI short dramas and videos, and about 58% of the United States’ tokens are used for programming and research and development.

The sources of both figures are unreliable.

The original source of China’s 55% is the June 2026 investor relations activity record sheet of an A-share listed company that heavily invested in AI short dramas, which states that “among all token consumption shares in the country, AI short dramas and video generation top the list, accounting for 55%.” After verification, the Beijing News Shell Finance pointed out that this is neither a national statistical caliber, nor the sample coverage and conversion method have been made public; the statement that "China uses 55% for short dramas, and the United States uses 58% for codes" itself has "no authoritative source and has not been carefully verified."

The figure of 58% in the United States also cannot withstand questioning. Some analysis and traceability found that it may come from the "State of AI" report jointly released by OpenRouter and a16z: among samples of more than 100 trillion Tokens, by the end of 2025, the OpenAI model accounted for 29% each in programming and technology-related tasks, totaling exactly 58%. But OpenRouter is just an API aggregation platform, and its statistics are only large language models, not video generation; its Token distribution is not equal to the U.S. Token consumption distribution. Many other media outlets have checked and concluded that the figure of 58% in the United States cannot be compared with authoritative statistics.

Both of these numbers deserve question marks. But the reason why they can trigger such a large-scale discussion is because they point to a real structural difference: China's AI applications are heavily focused on content generation, while American AI applications are heavily focused on coding and R&D. The differences are real, the numbers are rough. What is worth dismantling is the industrial logic behind the differences.

But before breaking down the logic, let’s break down a concept first

The comparison between 55% and 58% is based on an implicit premise: treating Token as computing power. This premise is wrong.

Token is the unit of measurement when the model processes information. It measures throughput - how many words and frames are in and out in one call. Computing power is the computing power of the GPU per unit time, and it measures how much work the hardware does. For the same Token, the corresponding computing power consumption in different tasks is not equal.

How big is this imbalance?

According to the "Token Economic Development Research Report (2026)" of the China Academy of Information and Communications Technology, a single episode of 1.5-minute AI short drama can consume 5 million to 10 million tokens. According to industry estimates, approximately 50,000 Tokens are consumed to produce 1 second of 1080p video, which is 3 million Tokens per minute. According to industry-wide estimates cited by Shell Finance in the Beijing News, a technical R&D team of 8 to 10 people consumes only about 3 million to 5 million tokens in a week of intensive R&D.

In other words: writing code for a week is not worth doing a skit for a day.

This report also has a more critical comparison: the value density of consumer-level scenarios is about US$0.01/million words, while the value of industrial-level scenarios such as drug research and development, financial quantification, and precision code generation can reach US$1,000/million words—a difference of 100,000 times.

Therefore, the comparison of "55% vs. 58%" is not valid in terms of methodology: it compares high-density consumption scenarios to low-density, high-output scenarios. China's Token consumption is large. First of all, it is because the video shooting itself consumes too many Tokens, not because it has misconfigured the computing power. The comment on China.com put it directly: video generation is a high-density consumption scenario, and code generation is a low-density and high-efficiency scenario. Comparing the two together is not methodologically valid.

Only after clarifying this level can there be room for discussion. The real question is not "who consumes more", but where the value stays after it is consumed.

Chinese short drama: Token is burned on "trial and error"

The reason why the token consumption of the AI ​​short drama is so amazing is that it is a chain-triggered pipeline. Script generation, storyboard design, character modeling, video generation, dubbing, editing, each process requires independent use of large models. The computing power cost of a 12-episode AI short drama is estimated to be about 200,000 yuan; the computing power cost of a 120-minute AI live-action short drama alone reaches 30,000 to 40,000 yuan on the model side, accounting for 50% of the total cost in mass production. In actual production, it is normal to draw the same shot more than ten times or even dozens of times. A large amount of tokens are consumed in the cycle of generation, dissatisfaction, and re-drawing, rather than the final output.

There is another number in the same investor relations record: core productivity scenarios such as technology research and development, code development, and model iteration only account for 15% of domestic Token consumption. 55% vs. 15%, this contrast is more illustrative than the comparison of "China vs. the United States" - China's Token consumption has been highly tilted towards the content side in the country.

But the prosperity of the content side has not left the value on the content side. DataEye's "AI Drama and Comic Data Report for the First Half of 2026" shows that Douyin added 221,900 AI dramas and comics in the first half of the year, and only 1,055 of them exceeded 100 million in playback volume, with a break-even rate of 0.48%. Using the industry-recognized break-even line of 50 million views, only 1.3% of the projects met the standard, that is, 98.7% of the projects were unable to recoup their costs within half a year of being launched, and on average only 1 of every 77 movies was profitable. The investment expenditure accounts for about 70% of the total revenue, and the final account share of the producer is only 4% to 7%. The saved labor costs are transferred to the platform to purchase volume. Tokens are burned in the generation process, money is made at both ends of the flow and computing power, and content companies bear all the uncertainty in the middle.

American Programming: Token burned in “workflow”

Although the source of the 58% figure in the United States is unreliable, the token consumption in the American programming scene does have a verifiable volume.

Microsoft's FY2026 Q4 earnings conference call disclosed that the number of GitHub Copilot users has reached 50 million, and Copilot-related revenue has accelerated by more than 60% quarter-on-quarter in a single quarter. Note that 50 million is the number of users, and the scale of paid subscriptions is much smaller. The number of paid subscriptions previously disclosed by Microsoft is in the order of 4.7 million. The difference between the two is about ten times and cannot be confused. Cursor's annualized revenue run rate has climbed from approximately US$100 million in early 2025 to exceeding US$4 billion in June 2026 - it should be noted that this is not audited financial report data, but a run rate estimate by the media citing people familiar with the matter.

The token consumption logic of programming scenarios is fundamentally different from that of short plays. The skit is a "one-pass" chain generation. If you make a mistake, it will be drawn again. The conversion rate between token consumption and effective output is extremely low. Programming assistance is a single point embedded in the workflow: generate a piece of code, interpret an error, and reconstruct a function. The amount of tokens each call is much smaller than video generation, but the call directly corresponds to a deliverable output. Developers rarely redraw the code calculated more than a dozen times for a function; the compiler and tests will tell developers whether the code is correct or not. The waste rate of Token is naturally lower than that of video generation.

What’s more important is value retention. Cursor's payment logic is "charge based on seat or usage", and enterprises pay for the developer time saved. The money comes out of the enterprise's efficiency budget and falls directly into the hands of the tool provider. The payment logic of short dramas is to "buy streams and buy traffic". The money comes out of the content company's marketing budget and falls into the hands of the platform. Both models consume Tokens, but the value created by Tokens remains at the tool layer and leaks to the traffic layer.

"China does entertainment and the United States does productivity" is an overly neat narrative

After only seeing the short drama about programming, it is easy to draw a conclusion: China is using AI to create nipple fun, and the United States is using AI to improve productivity. This narrative has a partial basis in fact, but it obscures three more complex realities.

First, a considerable part of the Token consumption in the United States is contributed by Chinese programmers. The Beijing News Shell Finance quoted industry insiders as saying that many software engineers at domestic technology companies use American models such as Claude to write code, and these consumptions are recorded in the U.S. Token ledger; OpenRouter's report also shows that Asia's share of global call expenditures has increased from about 13% initially to about 31%. The "nationality" of a Token is not equal to the nationality of the user.

Second, AI penetration in China’s productivity scenarios is not as stagnant as 55% versus 15% suggests. A commentary in the Economic Daily on October 26 pointed out that the proportion of large models and intelligent agents used by Chinese industrial enterprises has risen to 47.5%, a total of 35,000 basic-level smart factories have been built, and the localization rate of computing power chips has increased to 37%; TCL CSOT uses AI to schedule production, and the production plan is compressed from 7 days to 30 minutes; CATL has deployed more than 7,000 detection points on a production line, and the cell failure rate has dropped from one in a million to one in a billion. Stanford University's "2026 AI Index Report" shows that as of March 2026, the performance gap between China and the United States' top models has narrowed to 2.7% (Economic Daily's statement is "less than 3%", the caliber of the two is similar but the source is different). Short plays are the largest single stream, but not the only stream.

Third, the token consumption of the short drama itself is also a "stress test" of China's AI capabilities. The engineering requirements for high concurrency, low latency, and large-scale video generation have forced the advancement of computing power scheduling and inference optimization. The problem is not that the short drama uses tokens, but whether the spillover of this engineering capability is enough to offset the diminishing marginal content at the content level.

The real difference is not in “what to build” but in “who pays for the token”

If you look down the comparison between China and the United States from the level of "content vs. code", you will see a more fundamental difference: in the productivity scene in the United States, there is a mature payment market to bear the token cost; in the entertainment scene in China, the token cost is absorbed by the in-stream bidding, and a sustainable value cycle is not formed.

The software payment habits of American enterprises are mature. Software subscriptions are as rigid expenses as utility bills. Senior engineers in Silicon Valley earn hundreds of thousands of dollars per year. The manpower saved by writing AI codes is real money. Therefore, enterprise-level applications are the first to be mass-produced. The payers and users are the same people, and the value feedback chain is short and clear. Cursor can reach a run rate of US$4 billion because a group of companies are really using it to replace part of the development workload.

For Chinese short dramas, the payer is the streaming platform, the user is the audience, and the payer is the content company. Most of the money spent by content companies becomes advertising revenue for the platform, rather than returning to the content production process. Tokens burned out production, but the production did not turn into profits for the content company, nor did it turn into reusable technical assets.

This is not “China AI has gone the wrong way.” China has the world's largest content consumption market, and the explosion of AI in short drama scenes is a reasonable choice determined by the market structure. But a reasonable choice does not mean the best choice. When 55% of tokens flow to a link where "more than 90% of the project cannot recover its cost and the producer only gets a single-digit percentage", the long-term return on this investment needs to be recalculated.

Conclusion

The comparison of "55% vs. 58%" spreads because it uses a simple number to express a real anxiety: Is China's AI spending too much on "making people watch more enjoyable" instead of "making production more efficient?"

The numbers themselves cannot withstand verification, and Tokens do not equal computing power. But the problems the anxiety points to are real. The paid closed loop of the American programming scene proves that Token can create measurable and retainable value. The closed loop of investment in Chinese short dramas proves that tokens can create traffic, but traffic does not equal value retention.

The flow of tokens will eventually become the direction of the industry. China does not need to deny short dramas because of the 55% figure - it is reasonable to use AI to reduce the cost of short dramas, and the number of calls generated by large-scale scenes does dilute the cost of reasoning. The pricing of domestic video models has dropped to less than 1 yuan per second. But what needs to be asked is: outside of short plays, those scenes that are more difficult and slower, but can change production efficiency once completed, have sufficient computing power been obtained, and whether the same clear payment closed loop has been formed. If the answer is no, then 55% is not just a question of statistical caliber, but a question of configuration choice.

The consumption flow of Token is essentially the voting result of capital and market. The short drama consumes a large amount of tokens in exchange for traffic scale, and the code scenes accurately consume tokens to accumulate productivity value. Traffic can be achieved quickly, but the closed loop of industrial value requires long-term and intensive cultivation. In the follow-up, I will dismantle in depth: Domestic industrial AI and enterprise-level intelligence, which tracks have passed the real payment closed loop and got rid of the inefficient model of "burning tokens in exchange for water". Follow me and continue to understand the true profit logic of the AI ​​industry.

The factual basis of this article: Chinese Online Investor Relations Activity Record Form on June 3, 2026 (original source of "55%/24%/15%", verified by Shanghai Securities News, Beijing Daily, and Beijing News Shell Finance); Beijing News Shell Finance "Is "American AI improves productivity, Chinese AI creates nipple fun" a false proposition?" (data has no authoritative source, Token ≠ computing power, a team of 8-10 people consumes 3 million-50 yuan per week) 00,000 Tokens); Shanghai Securities News "Who is creating the conflict over word consumption" (55% traceability chain: Chinese Online → July Zhongtai Media Materials → Institute of Information and Communications Technology report, this proportion is not found in the text of the brokerage report); China "Research Report on Token Economic Development (2026)" by the Academy of Information and Communications Technology (a single episode of 1.5-minute short drama has 5 million to 10 million tokens, an average daily call volume of 500 trillion, and a value density of 0.01 vs. US$1,000/million words, 70.5% of inference computing power in 2026); China Net commented "Is AI creating short dramas a waste of computing power?" (high video density vs. low code density, the methodology is not valid; domestic model pricing dropped to less than 1 yuan per second); OpenRouter × a16z "State of AI" (100 trillion Token samples, OpenAI model programming 29%, technology 29% by the end of 2025); Economic Daily "Moyin consumption of words is a blinder" (October 26, 2026, industrial AI penetration Transparency rate 47.5%, 35,000 smart factories, domestic chips 37%, the gap between top models is less than 3%); Stanford University's "2026 AI Index Report" (the performance gap between Chinese and American models is 2.7%); Microsoft FY2026 Q4 financial report conference call (July 29, 2026, GitHub Copilot users 50 million, Copilot revenue increased by more than 60% month-on-month; paid subscriptions were 4.7 million for FY26 Q2 caliber); Cursor annualized run rate (approximately US$4 billion in June 2026, the media quoted people familiar with the matter, not audited); DataEye "AI Drama and Comic Data Report for the First Half of 2026" (221,900 units, 1055 units over 100 million, 0.48%, break-even line 1.3%, 77 units and 1 unit profitable); Securities Times, China News Weekly (The investment accounted for about 70%, the account split was 4%-7%, the computing power cost of the 12-episode short drama was about 200,000 yuan, and the 120-minute model terminal was 30,000-40,000 yuan); Chao News, Huxiu, and China News Weekly reported on the verification of this data dispute; 55% and 58% are not authoritative statistics. Token consumption does not equal computing power investment. The specific flow is subject to official or third-party independent audit data.

Disclaimer: The data, industry calculations, and report narrations quoted in this article are all from public materials. Some of the data are industry sampling estimates and are not full national audit statistics. Tokens, computing power, and revenue calibers vary among different institutions and do not constitute any investment advice. The opinions in this article are only industrial logical deductions and do not represent a value judgment on any company or track.

Respect for the original work, if you are qualified, you should be appreciated.

According to the real-name requirement of the Cybersecurity Law, please bind your mobile phone number before posting a comment.

Express

2026-11-11 23:07

Wang Yi speaks on the phone with US Secretary of State Rubio

2026-11-11 23:06

Domestic commodity futures markets mostly closed higher at night, with chemical products leading the gains.

2026-11-11 22:41

Mexico's main stock index rose to 72,231 points after the opening, hitting a record high

2026-11-11 22:36

AI cloud computing service providers collectively strengthened, CoreWeave rose 14%

2026-11-11 22:34

New York gold futures fell below $4,030 per ounce, down 0.01% on the day

2026-11-11 22:33

Nasdaq China Golden Dragon Index fell more than 1%

2026-11-11 22:32

The Philadelphia Semiconductor Index opened up 4.5%, recovering all of its 12% loss

U.S. stocks opened: the three major indexes rose collectively, CoreWeave rose more than 16%

2026-11-11 22:26

Russia plans to select lunar nuclear power plant plan in 2027

JD Logistics launches Tibet train

2026-11-11 22:25

EU 2027 budget initially finalized

2026-11-11 22:24

A magnitude 3.0 earthquake occurred in Hejing County, Bayinguoleng Prefecture, Xinjiang, with a focal depth of 17 kilometers

2026-11-11 22:20

China, Turkey and Egypt launch international cooperation initiative to strengthen high-quality carbon credits

2026-11-11 22:19

Hualin Securities: CEO Dong Dong received the regulatory talk decision from the Jiangxi Securities Regulatory Bureau

Eton Electronics: The private placement investment project will start construction in December

2026-11-11 22:15

Guosheng Securities: Received decision on administrative penalties and administrative regulatory measures

2026-11-11 22:02

Xiangcai Shares: Subsidiary Xiangcai Securities received penalties and regulatory measures from the Hunan Securities Regulatory Bureau

2026-11-11 21:53

The dollar's intraday gains against the yen expanded to 0.5% and are now at 159.53

2026-11-11 21:33

OPEC maintains forecast for global oil demand growth of 2.6 million barrels per day in 2027

OPEC: Crude oil production of member countries increased by 1.97 million barrels per day in October from the previous month

Scan to download App