A debate has erupted in China following claims that a significant portion of AI computing resources is being used to produce short-form dramas. The controversy was sparked by a survey result showing that short-form dramas and video generation account for 55% of China's total AI token usage.

While the analysis suggests that AI demand is concentrated more on video content creation than software development, the controversy is spreading as ByteDance raises doubts about the statistics' reliability. Amid the rapid growth of the AI industry, this debate centers on where and how much computing resources are actually being deployed.

The 55% Statistic That Turned China Upside Down

According to Chinese economic media outlet Caixin on the 21st, Chinese-online, a listed company on the Shenzhen Stock Exchange, stated in a recently released investor briefing material that short-form dramas and video generation account for 55% of AI token usage in China. E-commerce and marketing, such as advertising, followed with 24%, while software development stopped at 15%.

Tokens are the basic unit by which AI models process information and serve as an indicator to measure AI service usage and costs. The China Academy of Information and Communications Technology also cited the same figure in a research report on token economy development released last month.

As this figure spread to mean that half of China's entire computing resources are being used for short-form dramas, ByteDance stepped up to refute it. ByteDance's AI service, Doubao, recently pointed out via Weibo that the statistics in question do not match reality. ByteDance explained that the proportion of its total computing resources used for the training and inference of video generation models is only about 10%.

However, the measurement criteria differ between the two sides. Chinese-online based its criteria on the quantity of tokens processed by the AI model, whereas ByteDance presented the allocation ratio of total computing resources.

The process of pre-training AI models consumes massive computing resources, but hardly generates any output tokens provided to users. Therefore, a large volume of token usage alone does not mean that it occupies an equivalent share of total computing resources.

Doubts are also being raised about the representativeness of the 55% figure. A Chinese-online official explained to Caixin that the statistics were calculated targeting the first quarter of this year, combining the token usage generated during the company's own AI content production process and external research reports.

Due to the characteristics of a company expanding its AI short-form drama business, the tokens poured into content production are more than twice as those for software development. Objective and authoritative third-party statistics objectively aggregating token usage by purpose across all Chinese AI services are still lacking.

More AI Used for Dramas Than Coding

There is little disagreement within the industry over the fact that massive amounts of tokens are used to produce short-form dramas. AI video production requires the repetitive process of generating videos until the desired scene is achieved and selecting the final results.

According to TokenHub, an AI model service platform by Tencent, generating a 5-second, 1080P silent video with Kuaishou’s Kling Video V3 model consumes about 80,000 tokens per second. Creating a single photo using a related image generation model requires about 20,000 tokens.

China's short-form drama market is also growing explosively. According to the China Netcasting Services Association, out of about 128,000 short-form dramas released in the first quarter, works produced with AI exceeded 95% at 122,000 titles. In the second quarter, another 153,000 AI short-form dramas were added. A total of 275,000 titles were produced in the first half of the year alone, averaging 1,519 titles per day.

ByteDance’s release of the video generation model Seedance 2.0 last February accelerated changes in the production market.

The demand for video generation is also reflected in the performance of AI companies. According to Volcano Engine, ByteDance's cloud computing business division, the daily average token usage of the Doubao model reached 180 trillion in June, increasing more than tenfold over a year. Market research firm IDC analyzed that Volcano Engine took first place with a 49.5% market share in China's enterprise model-as-a-service calls last year, while Alibaba Cloud ranked second with 28%.

ByteDance's influence is also growing in the content distribution market. According to market research firm QuestMobile, monthly active users of ByteDance's short-form drama platform Hongguo reached 368 million in June. This scale accounts for 89% of China's total short-form drama user base of 414 million.

Profitability remains a homework assignment despite the growth of the video generation model market. Chinese AI company MiniMax exceeded an annualized recurring revenue of 800 million dollars last August, but recorded a loss of 358 million dollars (approx. 475.8 billion won) in the first half. This is due to the massive costs spent on training and inferencing high-performance video models.

An industry insider noted, "While video production has established itself as a major demand destination in China's AI industry, it is time to evaluate by distinguishing token usage, actual computing costs, and corporate profitability."

[ⓒ Global Economic Times. 무단전재-재배포 금지]