AI short dramas may not be "compute slop": they may be creating new demand for compute

AI short dramas may not be "compute slop" APPSO October 25, 2026 10:43:10

Over the past couple of days, a PDF document from several months ago has been circulating. It is the minutes of an investor relations event held by a certain consulting firm, and it contains a set of startling figures: of all Token consumption in China, AI short dramas and video generation account for 55%, e-commerce and marketing account for 24%, and software development for only 15%.

What makes these numbers so memorable is not only that they are exaggerated, but also that they fit a kind of misplaced expectation: we all assumed that AI's mission was to write programs, do scientific research, and boost productivity, yet what has appeared first and in overwhelming volume may be one after another of those comeback, revenge, and domineering-CEO stories that you can finish watching in a few minutes.

First of all, how compute is actually consumed cannot easily be calculated. A training task usually enters the cluster as an independent job, and how many cards it calls, how long it runs, which model it belongs to, and how many GPU-hours it produces are all recorded in scheduling and billing records. Inference likewise goes through explicit model service endpoints, and cloud providers can tally infrastructure request volumes, Tokens, generation duration, and occupied resources. Deloitte once estimated that in 2026 inference would account for about two-thirds of all AI computing.

But if you keep asking what the inference results are used for, the infrastructure may not know. The same ten-second clip generated by a video model might be used for a short drama, an ad, game assets, or video and audio; and a single short drama will in turn call multiple models for text, images, video, and speech. Unless the customer proactively labels the business type, or the platform infers it from the account and workflow, the underlying logs can only show that "a certain model performed a generation"—they cannot show that "this generation belongs to an AI short drama."

So directly claiming that "half of all compute is used to generate short dramas" is naturally inaccurate, and to go further and speak of "the share of all domestic Token consumption" would require complete call data from ByteDance, Alibaba, Tencent, Baidu, and other model platforms—which is simply impossible.

Moreover, even if all Token consumption data were obtained, it would not directly yield a compute share. Text, images, and video have different Token definitions, and the same quantity can conceal different amounts of computation.

People easily imagine compute as a cake of fixed size waiting to be divided: short dramas take 55%, and programming, office work, and research then split the remaining 45%. But in actual operation, compute itself has no "production volume" that can be used for calculation; the concept mixes at least three things:

Compute capacity: how many GPUs a vendor has already purchased and deployed, and how many floating-point operations per second these pieces of hardware can theoretically complete at most, usually expressed as peak FLOPS. This part requires investment and construction first; even if there are temporarily no tasks, the machines must be there, which is what the phrase "buying cards" means.

Compute usage: how long these machines actually worked during a certain period, commonly measured in GPU-hours. This usage only occurs when tasks enter the cluster and occupy GPUs.

Computational workload: how many floating-point operations the model actually performs to complete a task, commonly expressed in FLOPs. It depends both on model scale and on video length, resolution, number of generations, and number of denoising steps.

Therefore, compute is not a fixed budget allocated in advance, but computation triggered by requests, brought forth by demand. If a programmer does not ask an Agent to design a program, that portion of compute does not happen, and the freed-up machines may be occupied by other tasks or may simply sit idle.

A previous Reuters investigation found that utilization rates at some domestic compute centers were only 20%–30%. Compute is not always in short supply, and demand from short dramas may get machines that were previously idle, or that had no clear customers yet, actually running.

Not half, but really a lot

The China Netcasting Services Association disclosed that in the first half of 2026, platforms launched about 367,000 micro short dramas in total, of which about 270,000 were AI micro short dramas, accounting for more than 74%. It was not until the second quarter that the supply of live-action short dramas recovered somewhat.

Among the Top 100 popular short dramas from January to August 2026, AI content accounted for about 75%

Even narrowing the scope to Douyin, the growth remains astonishing. Reuters, citing DataEye data, reported that in the first half of 2026 Douyin launched 221,900 AI dramas. However, only 1,055 of them exceeded 100 million views; a large amount of content was produced quickly and likewise quickly sank into the information feed, and a huge gap has opened up between generation scale and success rate.

This production capacity has even begun to create its own actors. On October 12, Mango TV officially announced five "native AI artists" all at once: Cheng Du, Fang Yu, Lin Enna, Qi Xu, and Zhan Yan. Their faces, personas, character three-view sheets, and styling systems were generated from scratch, each opened a real-name-verified Weibo account, and each was backed by real human screenwriters and managers. On October 14, the six-episode AI simulated-human drama "What a Friend" starring the five launched on Mango Lingchuang Theater, with each episode running about ten minutes.

A single video generation requires far more computation than text, and it is also accompanied by a large amount of repeated rerolling. A test by NVIDIA showed that a 5-second, 720P video requires repeating calculations about 40 times over roughly 72,000 spatiotemporal Tokens.

So producing short dramas has indeed brought new inference demand. This year, a "Token Factory" aimed mainly at short-drama film and television companies appeared in Zhengzhou, with core capabilities such as high-concurrency compute scheduling and low-latency inference responses customized for AI short-drama generation scenarios, and it plans to provide dedicated compute support for more than 3,000 short-drama-related enterprises in the local area.

Similar changes have also occurred in traditional short-drama hubs. Xi'an Daily's all-media outlet, citing public data, reported that the comprehensive annual output value of Xi'an's short-drama industry exceeds 6 billion yuan, of which the increment contributed by AI short dramas is about 3.5 billion yuan.

Whether or not one subjectively likes AI-made content, objectively this is an extremely popular market that has created a batch of computation that would not otherwise have happened. If no one made short dramas, those videos would not be generated, repeated "rerolling" would not occur, and the corresponding inference tasks, Token consumption, and GPU runtime would not exist.

In this sense, short dramas may not be taking an existing slice of compute away from scientific research and programming, but rather creating new demand: only after there are content orders are there model calls; as calls continue to grow, platforms expand inference resources, build dedicated scheduling systems, and even build "Token Factories" for short dramas.

It sounds a bit counterintuitive, but it is very real: what first sustains a foundational technology is not necessarily the lofty uses one imagines, but the uses that most easily form large-scale consumption.

This is not even limited to China. In the first half of this year, micro-short dramas on YouTube also exceeded 6.5 billion views, no small number of them involving AI generation, and they have spread widely on social media.

The more short dramas users consume, the more incentive enterprises have to purchase servers, optimize video models, and reduce the cost of each generation. They may also, through high-frequency generation, repeated trial and error, and clear commercial returns, provide video models with orders, revenue, and training feedback.

Whether short-drama content is nourishing or highbrow is not the same question as whether it can support technological expansion—yet it remains an important question.

Musk leads tech titans in a debut dance.

So brainwashing—who could resist watching to the end? From X user @garylau_ai

Tracing it back, the copy for this video that went viral overseas is precisely "these men who made RAM sell at sky-high prices," which has the same flavor as "half of NVIDIA's compute is spent in Tieling."

Generative video may be the first time ordinary people consume AI on a large scale and continuously; compute is directly exchanged for images, emotions, and a few minutes of entertainment. Users may not know they are consuming GPU time, just as they do not think about the bandwidth behind streaming, but every generation, redo, and selection is forming real demand.

If these tasks use idle resources and the costs are paid by viewers, advertisers, and production companies, it is closer to a new demand. But when scarce chips, electricity, and public subsidies are continuously invested in it and squeeze out other uses, resource allocation becomes a question that must be answered.

However, the market does not allocate computation according to knowledge value; it responds first to the demand that most easily forms orders. Short dramas can be generated at high frequency, allow repeated trial and error, and can quickly verify results with views and payments, so they will naturally push model companies to continue optimizing character consistency, shot stimulation, and batch production. Whether short-drama content is nourishing and whether it can support the expansion of video models are indeed two different questions; but the answer to the latter will sooner or later change the world the former faces.