At 11:18 AM on October 19, Liu Xing, the head of public relations at Doubao, shared a post on Weibo.

"The computing power data released by a certain company seems rather questionable. Take ByteDance as an example—only about 10% of our total computing power is allocated to the training and inference of video generation models," he stated.

In an investor relations activity record from early June, ChineseAll revealed that 55% of domestic AI Token consumption was used for short dramas and video generation, 24% for e-commerce advertising, and merely 15% for research and development. This information resurfaced in mid-October, sparking a Weibo trend with the topic "Over half of domestic computing power is used for AI short dramas."

When tracing the origin of the 55% figure, various sources point to the Research Report on the Development of the Token Economy published by CAICT this year. The report analyzes the usage structure of Token calls across the entire industry, covering every real generation on platforms such as Douyin, Kuaishou, Jimeng, and Kling. When ChineseAll referenced this data in its investor relations materials, it prefaced it with a disclaimer, noting that it was based on industry research data.

ChineseAll's circumstances at the time also shed light on why it cited this set of data. On October 8, the company had just terminated its H-share listing and A-share private placement. Its 2025 revenue stood at 1.657 billion yuan, with a net attributable loss of 671 million yuan, and sales expenses accounting for 57.54% of revenue. In the first half of the year, revenue from short dramas and IP derivatives reached 411 million yuan, accounting for 71.16% of total revenue. For ChineseAll, this data served as crucial support.

Only ByteDance truly knows the exact proportion of its total computing power dedicated to video generation. In 2026, ByteDance increased its capital expenditures from 160 billion yuan to 200 billion yuan, with approximately 85 billion yuan earmarked for AI chips. In September, it also secured a $29.6 billion syndicated loan. The computing power supported by these chips is not solely for Doubao; it also encompasses Douyin's recommendation and advertising systems, search, Feishu, and the computing power sold by Volcano Engine to external clients, all included in these capital expenditures.

Therefore, it's not surprising that the proportion of computing power ByteDance actually uses for AI short dramas (video generation) is relatively low.

CAICT counts Tokens, not GPU hours. Video Tokens are calculated based on pixels multiplied by the frame count. A 5-second 720p clip amounts to about 100,000 Tokens on mainstream platforms, equivalent to the input of a 100,000-character script.

A 90-minute AI short drama requires 110 million Tokens just for the final product, and with re-generations, the actual consumption can soar to hundreds of millions.

This conversion ratio is detailed in each company's price list and is part of commercial pricing. In other words, a video can range from 100,000 Tokens to 3 million Tokens, depending on how the pricing table is structured.

Furthermore, the method used to derive the proportion from this set of conversions varies when corresponding to computing power distribution. Some media outlets incorrectly used "computing power distribution" instead of "call volume structure" in their headlines, which is inherently misleading.

ByteDance's role in this matter is far from insignificant. Hongguo Short Drama boasted 168 million daily active users in July this year, surpassing the combined daily active users of Tencent Video, iQIYI, Youku, and Mango TV. Douyin launched 220,000 new AI short dramas in the first half of the year. Jimeng sells Tokens, and Volcano Engine sells computing power. Amid this frenzy, ByteDance is akin to a landlord collecting rent. Its claim that videos account for only 10% is also the only statement from a company with a motive to downplay the proportion, and this figure cannot be verified by any third party.

An unaudited industry estimate has thus been "cherry-picked" three times by three different parties, each selecting the portion that suited their narrative. Everyone ended up with the half-sentence they desired.

Speaking of short dramas, they have indeed enjoyed immense popularity this year. In the first eight months, 430,000 micro-short dramas were released domestically, 13 times the total for last year. Based on the break-even line of 50 million views, less than 1.3% meet the threshold, with an average of only 1 out of every 77 dramas recouping costs. In the first quarter, the number of active production companies dropped from 1,216 to 698, a decrease of 42%.

The computing power cost for a single episode of an AI animated drama has surged from 17 yuan at the beginning of the year to 80-100 yuan. Since April, mainstream cloud providers have adjusted their computing power card prices, with increases of up to 34%. Daily advertising spending across the industry ranges from 150 million to 170 million yuan, with acquisition costs accounting for 70-90% of revenue.

More and more short dramas are being produced, yet fewer and fewer are being watched. The computing power unit price is rising, while the probability of recouping investments is declining. The intersection of these two trends represents the current state of this industry chain.

On October 19, the concept of AI animated dramas surged during trading. Zhangyue Technology hit its daily limit, while ChineseAll's stock price initially fell 2% in the morning but later recovered. The market was digesting Liu Xing's debunking while continuing to price the short drama concept. A figure without a unified standard can still cause billions of yuan in market value to fluctuate.