Viral Zhao Benshan AI Videos Fuel China Computing Rumour as ByteDance Cites 10% Share

A comedy craze meets a disputed statistic

ByteDance devotes about 10% of its total computing capacity to training and running video generation models, Doubao public relations chief Liu Xing said on October 19, 2026, responding to a claim that more than half of China's computing power was being used to make short dramas.

The dispute became entangled with a different internet sensation: artificial intelligence videos placing Chinese comedian Zhao Benshan on stages styled after Michael Jackson performances. Clips circulating around October on Bilibili, Douyin and WeChat Channels prompted jokes that Zhao alone was consuming half the country's computing resources.

Behind the joke was a specific figure, 55%, attributed to an investor relations activity record from Chinese Online. The company said AI short dramas and video generation accounted for that share of domestic AI token consumption, citing industry research. As the number spread, token consumption became computing capacity in online discussion.

The two percentages describe different things. Chinese Online's 55% concerns a claimed national share of token use. Liu's approximately 10% concerns the share of one company's computing resources devoted to video models. Understanding their different units and boundaries explains how an entertainment trend became an argument about China's AI industry.

Zhao himself did not participate in making the clips. His familiar image supplied the material for creators, while the computing claims came from a separate business discussion about demand for AI services.

Why Zhao appeared on an international stage

Creators labelled many of the videos as a world tour, combining Zhao's face with Michael Jackson's stage appearance and adding garlic, corn and other visual references to northeastern China. Some edits also placed Nvidia founder Jensen Huang and Tesla chief executive Elon Musk in the scenes.

Zhao's appeal rests on decades of television familiarity. His Spring Festival Gala comedy sketches and series including Ma Dashuai and Country Love made him recognisable across generations. Moving that rural comedy persona into the setting of an international pop performance created the contrast driving the joke.

More accessible AI tools, including widely available models, helped ordinary users produce and circulate variations. Face replacement changes the identity seen in an existing or generated image, while image generation can create the surrounding costumes, props and settings. Creators can combine these steps into a short music video.

The repeated appearance of the same familiar face across feeds gave users an easy way to describe the craze: enormous amounts of computing power must be going into Zhao videos. The phrase spread as an exaggeration about what people were watching, then acquired an apparent numerical foundation through the unrelated 55% claim.

What Chinese Online's figures actually measured

Chinese Online's investor relations record attributed 55% of domestic AI token consumption to short dramas and video generation, 24% to ecommerce advertising or marketing, and 15% to research and development. Descriptions of the last category varied, with some identifying it more specifically as software development.

The company also said token consumption for AI short dramas and video generation exceeded twice that for software development. Using the stated figures of 55% and 15%, the first category is about 3.7 times the second. Together, the three listed shares add up to 94%, leaving 6% outside those categories.

A token is a small unit a model processes. In a language model, it can be a word, part of a word or punctuation. Prompts, scripts, descriptions and subtitles all generate token usage, making token counts useful for measuring activity and billing for many AI services.

Video systems can also process representations of images and other media. The computing work associated with a token therefore depends on the model and task. To understand a combined token total, readers need to know which services were counted and how their usage was measured.

Why computing power is a different measure

Computing capacity describes the resources available to perform calculations, including the work done by graphics processing units, or GPUs. Those processors can train models, generate images and video, and run many other demanding applications.

Token consumption measures units processed by models. Computing usage measures the resources spent processing tasks. A useful comparison is the difference between counting jobs sent to a factory and measuring the machine time used to complete them: some jobs require much more work than others.

Video generation often demands more computation than a text response because the system must create many images while keeping movement, faces and objects consistent across time. The length, resolution and complexity of the output all affect the workload.

The 55% figure therefore concerns the distribution of a reported usage measure, while the 10% figure concerns ByteDance's allocation of computing resources. A direct calculation treating the two as competing estimates of the same national total would combine different measures and different populations.

What Doubao's response covers

Liu Xing, the public relations chief for ByteDance's AI assistant Doubao, challenged the circulating computing claim in his October 19 response. He described the company's allocation in these words, translated from Chinese:

Taking ByteDance as an example, only about 10% of our company's total computing power is used for training and inference of video generation models.

Training is the process through which a model learns patterns from data. Inference is the process of using the trained model to produce an answer, image or video. Liu's percentage includes both activities for video generation models.

His statement covers the broader video model category at ByteDance, rather than a measured allocation for the Zhao clips. That category can serve many kinds of creators and productions, from short dramas to other generated video.

Several details remained unspecified when the response circulated: Chinese Online had not published the underlying research method or further clarified its figures; Liu had not disclosed ByteDance's total capacity, the measurement period or the allocation of the remaining approximately 90%; and no authoritative independent national comparison was available. The company had given no date for a further explanation.

How the claim travelled

The sequence places a business statistic, a company response and a comedy trend within the same period. The dates help explain why the claims became connected in public discussion.

• A Chinese Online investor relations activity record cited industry research assigning 55% of domestic AI token consumption to AI short dramas and video generation. Its publication date was not specified in the accounts of the dispute.
• Around October 2026, Zhao Benshan world tour edits circulated on Bilibili, Douyin and WeChat Channels.
• On October 19, 2026, Liu Xing said approximately 10% of ByteDance's total computing capacity went to training and inference for video generation models.
• By November 8, 2026, coverage of the Zhao video craze described how the token statistic had become attached to the joke about half of China's computing power.

The change in wording mattered. A percentage presented as token consumption became a percentage of computing power, while the subject narrowed from short dramas and video generation generally to clips featuring a single entertainer.

Why AI dramas can require repeated computing work

Chinese Online also described a practical reason for heavy demand from video producers: repeated generation. Teams may request several versions of a scene before selecting footage that meets the director's requirements for storytelling, camera movement and appearance.

A failed or rejected version still consumes resources. Producing a convincing face in one frame is one task; keeping that face, its expressions and its surroundings consistent throughout a shot adds further work. Revisions multiply the number of attempts behind the finished clip.

According to the company's analysis, AI dramas aiming for a realistic human appearance typically consume two to three times as many tokens as AI animated dramas. It linked the difference to stricter requirements for frame counts and 1080p resolution.

A reported production example put computing costs for a drama of 12 episodes at close to 200,000 yuan. That illustrates the potential expense of a production involving substantial generation and revision. Budgets depend on episode length, visual quality, model prices and the number of discarded attempts, so producers face very different costs from project to project.

A large audience sits behind the production rush

China's short drama audience was reported at 851 million users by May 2026, approaching seven in ten mobile internet users. That scale gives platforms and producers a large potential audience for both conventional productions and AI generated stories.

Reported figures put the number of microdramas released during the first eight months of 2026 at 430,000, about 13 times the total for the whole of the previous year. Douyin alone was reported to have added 220,000 AI short dramas during the first half of 2026. These counts describe releases, with productions potentially varying widely in length and format.

ByteDance's Hongguo short drama service was reported to have 168 million daily active users, compared with approximately 155 million combined for Tencent Video, iQiyi, Youku and Mango TV. Using those figures, Hongguo's audience was about 13 million larger, or roughly 8% above the combined total.

Hongguo users were also reported to spend about 125 minutes a day on the service. Audience time on that scale helps explain the commercial interest in producing more stories and testing cheaper production methods.

For creators, accessible video tools make it easier to turn a familiar personality or a visual joke into something shareable. For production businesses, repeated generation creates demand for model services and computing infrastructure. Those are concrete pressures behind the debate, even as the national allocation of resources remains a separate measurement problem.

Key Points

• AI videos placed Zhao Benshan in Michael Jackson styled performances, prompting jokes about their computing demands. Zhao did not make the clips.
• Chinese Online cited industry research assigning 55% of domestic AI token consumption to short dramas and video generation.
• Doubao public relations chief Liu Xing said on October 19 that approximately 10% of ByteDance's total computing capacity supported video model training and inference.
• The percentages cover different measures and populations: reported national token consumption and one company's computing allocation.
• Repeated attempts, visual consistency and resolution requirements can make generated video costly, especially for dramas seeking realistic human appearances.
• China's large short drama audience helps explain demand, with 851 million users reported by May 2026.