Of China’s domestic AI token consumption, 55% goes to short drama and video generation, 24% to e-commerce advertising, leaving merely 15% for R&D. Token usage for AI-generated short dramas is more than double that of software development, making it the largest consumption scenario in “token economics.”

“Over in the U.S., computing power goes toward protein folding; here, ours churns out countless revenge-fantasy and domineering-ceo short dramas.”

This stark, sobering comparison has gone viral across Chinese social media recently.

The discussion stems from a filing by A-share listed firm Chinese Online during an investor relations session: 55% of domestic AI token consumption is burned on short drama and video generation, 24% on e-commerce advertising, and only 15% is allocated to research and development. Token usage for AI short dramas exceeds that of software development by more than twofold, making them the largest consumption category within “token economics.”

Yet the reality is far more nuanced than the simplistic label of “embarrassing.”

I. A set of figures illustrating the industry’s frenzy

In the first eight months of this year, 430,000 micro-short dramas were launched in China — 13 times the total for all of last year. On Douyin alone, 220,000 new AI short dramas premiered in the first half of the year.

An AI short drama series with 12 episodes carries computing costs of nearly RMB 200,000.

Why so costly? Video generation is inherently a token-hungry beast. The compute resources consumed by just a few seconds of video could power thousands of LLM text responses. AI generation is also inconsistent; a single scene often needs to be regenerated a dozen or more times. Insiders call this “card drawing”: every failed attempt adds to the final bill.

The irony is striking: supply has exploded, yet very few creators turn a profit.

Among those 220,000 AI short dramas, only 1,055 surpassed one billion views, a hit rate of merely 0.47%. Industry estimates suggest only one out of every 77 new dramas breaks even. Ninety percent of production houses operate at a loss, and 42% of firms exited the market within six months.

Meanwhile, computing prices have surged from roughly RMB 300 per minute at the start of the year to nearly RMB 1,000. At the same time, revenue per thousand views for short dramas has slumped from RMB 60 to RMB 15–30. Margins have been completely squeezed out by this dual shift.

II. Prospectors lose everything; shovel sellers rake in the profits

The players reliably making money in this value chain are cloud vendors selling compute and platforms selling foundation models.

Douyin’s own **Jimeng AI** raised its prices three times within a single month this April. Industry calculations show the cost of producing a 2-minute video jumped from RMB 5.2 to RMB 40.2.

Complaint platforms are flooded with grievances, but the platform faces little pressure. It hosts one of the leading models in the space; if one buyer walks away, another 20,000 are waiting in line.

This is textbook upstream supply lock-in.

Notably, Chinese Online also pointed out that this intense token consumption is not pure waste. It forms technical and resource moats within the AI short drama sector. Production houses delivering higher-fidelity content and burning larger token volumes gain an edge in the competition.

III. Don’t just dismiss it as intellectually hollow, nor be alarmed by the 55% figure

This debate deserves two sides of the story.

On one hand, reflection is warranted. Compute is one of the scarcest strategic resources of our era, yet large volumes are being funneled into formulaic dopamine-driven plots: underdog comebacks, sudden wealth, time travel. Hongguo Short Drama reports an average daily usage duration of 125 minutes, and medical professionals have publicly raised concerns over cognitive impacts. The erosion of public attention is tangible. Those same resources could accelerate drug discovery, materials science, and industrial software — fields that deliver slower but deeper long-term value.

On the other hand, direct comparisons of consumption shares between China and U.S. strategic priorities are misleading. Video tokens cost dozens of times more than text tokens by nature, automatically inflating that 55% share. Moreover, market-driven commercialization is not inherently negative:

Short dramas are among the few self-sustaining AI use cases, generating steady cash flow for model developers and cloud providers.

OpenRouter data shows that between October 5 and 11, global LLM inference volume hit 172 trillion tokens. China accounted for 86.74 trillion — more than four times the U.S. figure, ranking first globally for 24 consecutive weeks.

The top five models globally by token volume are all domestic Chinese models: DeepSeek V4.1 Flash leads with 41 trillion tokens, followed by Zhipu GLM 5.3 Flash, Xiaomi MiMo-V2.6, StepFun Step 5, and Tencent Hunyuan Hy4.

Without demand from these seemingly “lowbrow” use cases, inference cost reduction and ecosystem maturation might not have advanced nearly so quickly. Sustainable commercial cycles often pave the way for more ambitious research.

The market fundamentals are also substantial: the global micro-short drama market reached nearly USD 18 billion in 2025, with China capturing 77.78% of that market. Forecasts project the global market could exceed USD 100 billion by 2030.

AI is already embedded across the entire pipeline: script generation, virtual production, editing, localization dubbing, and ad campaign launch. This is not an industry that can simply be shut down.

IV. Takeaways for international students choosing majors and career paths

This battle over computing resources carries clear signals for students mapping out study-abroad and career plans:

First, AI infrastructure and AI application engineering offer more stable career tracks than pure content creation.

Ninety percent of short-drama production companies lose money, while cloud and model vendors capture most of the profit. Within the same industrial chain, wealth distribution is wildly uneven.

When selecting majors, prioritize fields that build “shovel-selling” capabilities: artificial intelligence, computer science, data science, cloud computing, rather than trendy content-production programs.

Second, hybrid “technology + domain expertise” profiles are in shortest supply.

AI is reshaping film, advertising, e-commerce, translation and every content-heavy industry. Professionals who understand both model limitations and domain workflows are the critical talent needed to move the industry from trial-and-error “card-drawing” toward refined, controlled production.

For graduate applications, consider interdisciplinary combinations such as CS + media studies, AI + design, or data science + business analytics.

Third, validate majors against labor-market reality, not online hype.

The rapid exit of 42% of short-drama firms in half a year illustrates that low-end capacity and speculative entrants are being weeded out fast.

Before committing to a major, check employment statistics in your target country: which roles are genuinely expanding? Which are just marketing fads? When the hype recedes, hard technical skills remain your foundation.

Fourth, stay grounded in China’s market while maintaining a global perspective.

China’s weekly LLM token throughput is over four times that of the U.S., and the world’s top five models by usage are all homegrown. China is arguably the world’s most vibrant testbed for AI deployment and monetization.

International students who bring advanced frameworks learned overseas back to China’s rich application landscape build highly competitive profiles.

Closing Thoughts

The real risk to watch is not how much compute goes to binge dramas, but path dependency. If everyone crowds into short-cycle content monetization and cutthroat competition, who will take on the slow, high-stakes long-horizon work: enterprise software, industry-specific agents, and foundational scientific research?

The tide is already receding. Daily ad spend for short dramas has fallen from its peak of RMB 180 million to roughly RMB 130 million, and low-quality capacity is being cleared out in batches. What matters now is whether this pool of compute and talent can pivot from creating fictional domineering CEOs toward building our future.

For individuals, the lesson is simpler: don’t be the prospector. Learn to build the shovels.