Economic Observer reporter Wang Yajie

Yu Shengqi's team recently moved their home. The screenwriters and reviewers stayed in their original cities, and a large number of video generation tasks were moved from Beijing to the hosting center in Suzhou. Yu Shengqi is the founder of an AI (artificial intelligence) anime production company, leading a team of several people to catch up with the peak season for anime releases at the end of the year, with machines running 24/7.

Yu Shengqi told the Economic Observer reporter that with the same one minute film, the cost of computing power has increased to two or three times that of the beginning of the year. "I hope to use this (team relocation) to shorten the production cycle and also reduce the constantly increasing cost of computing power

Similar to the actions of Yu Shengqi's team, in the past two months, there has been a trend of computing power shifting southward in the AI short drama industry, and many northern teams are moving their rendering clusters to the Yangtze River Delta. The Economic Observer reporter contacted multiple northern drama teams and learned that these teams are promoting this migration to varying degrees.

The "Digital Yangtze River Delta Development Report (2026)" jointly released by the development and reform departments of Shanghai, Jiangsu, Zhejiang, and Anhui shows that the intelligent computing power scale of the Yangtze River Delta accounts for nearly 30% of the country's total.

A computing industry insider analyzed to the Economic Observer that the domestic computing business has developed more maturely in the Yangtze River Delta, and the supply of computer rooms is also more abundant.

Yu Shengqi said that he has compared several cities in the Yangtze River Delta, and found that Suzhou has abundant supply of high-power cabinets and more practical hosting quotes. In addition, batch rendering does not require high latency, so overall it is more suitable for the business needs of his team.

The person in the computing industry added that compared to other cities, Suzhou's key lies not in having more surplus, but in its different form. The Suzhou hosting server room is based on long-term contracts with manufacturing, finance, and research clients. Idle computing power can be leased to project-based teams on a per card basis, and the hosting provider does not need to build a new server room for short dramas.

In the data center in Suzhou, this type of content team is becoming a new customer. Some will migrate and host their own servers, while others will rent computing power on demand and submit tasks remotely. The division of labor in content production has extended to cities.

The background of demand is the rapid growth of AI short drama supply. Wang Xiaoliang, Director of the Network Audiovisual Program Management Department of the State Administration of Radio and Television, introduced at the press conference of the State Council Information Office on September 17 that in the first eight months of this year, 430000 micro dramas were launched nationwide, which is 13 times that of last year. Among them, AI dramas accounted for more than 90%.

These newly added production demands have brought orders to computing power service providers.

In this wave of AI short drama craze that is sweeping the screen, profits are flowing along the industry chain, from the hands of drama creators to the computer rooms that supply computing power.

Cabinet welcomes new tenants

Before the migration, Yu Shengqi's team relied on self built office equipment and public cloud to complete video generation. Yu Shengqi told the Economic Observer reporter that when the 8-card equipment in the office is running at full capacity, there is a lot of noise and heat dissipation pressure. Due to the limitations of the venue conditions, team members are required to shut down after work at night. A 12 episode short drama requires eight to ten days for rendering alone. Renting public cloud computing power may result in resource queues during centralized rush work.

After transferring to the Suzhou data center, the equipment can operate all day long. Yu Shengqi introduced that high computing tasks such as large-scale video rendering can be continuously operated 24/7 in the Yangtze River Delta data center. At night, the server completes the rendering of the film, returns the film materials in the early morning, and during the day, the team focuses on reviewing and modifying the shots. The overall delivery cycle for short dramas of equal size has been compressed to three to four days.

The uninterrupted operation of servers is determined by the current industry production mode. The team needs to repeatedly generate ten versions of the same shot, and only keep the available version in the end. A large amount of computing power resources are consumed in the process of repeated trial and error and generating invalid images.

The head of a local computing power hosting service provider in Suzhou revealed to the Economic Observer that in the past six months, a large number of new short drama clients have come from cities such as Beijing, Xi'an, Chengdu, and Changsha. A medium-sized short drama enterprise with dozens of employees directly moved the entire set of 48 card GPU (graphics processing unit) servers into the computer room; There are also small teams of three to five people who do not purchase hardware and rent computing power based on card time (actual GPU usage time), while personnel stay in other places and submit rendering tasks remotely. During the concentrated online stage of the series, customers' GPU computing power will be instantly occupied, and the short-term computing power utilization rate of the data center will exceed 80%.

In the opinion of this person in charge, high-power cabinet resources in Beijing and other places continue to be tight, and the overall hosting cost in Suzhou has stronger competitiveness. Batch video generation tasks require low network latency, and cross city deployment can fully meet business needs.

The above-mentioned person in charge said that the original customers of the computer room were mostly manufacturing, financial, and research enterprises, signing long-term contracts for six months or one year, with equipment fixed on shelves and running 24/7. After one year, the load did not change much. The short drama team is very different. The equipment comes quickly, the tasks come fiercely, and once the project is over, they stop as soon as they say.

The staff of Suzhou Data Bureau stated that the computing power demand of AI short dramas has pulse characteristics, such as the instant saturation of computing power during the rush to launch stage of the drama, the rapid decline of computing power demand after project completion and release, and the large fluctuation gap between busy and idle time, which is not suitable for long-term locking of full cabinet computing resources. Compared to other cities in the Yangtze River Delta, other cities have also laid out short drama related computing power businesses, but they tend to focus more on local content incubation, large-scale model training, and other long-term intelligent computing businesses. This type of project, where northern content teams remotely host batch rendering across cities, is increasingly flowing to third-party hosting centers in Suzhou.There are also many hosting centers in Suzhou that use long-term contract customers in manufacturing, finance, and scientific research as their basic business chassis. The idle cabinet resources released from this are perfectly adapted to the flexible business of short drama, which occupies during busy hours and releases during idle hours.

Public computing platforms are the first to feel the trend of concentrated customer settlement.

The person in charge of the operation of a public service platform for parents' triangle computing power told Economic Observer reporters that since the beginning of the second half of the year, the platform in his area has significantly increased the number of new AI short drama and comic drama customer registrations. From the perspective of computing power consumption structure, large model training tasks account for more than 40% of the platform's total computing power, and nearly 60% of the computing power flows into inference business. AI short drama and anime batch rendering are important components of inference business.

New demands and tenants are also changing the charging methods of service providers.

The person in charge of the hosting company mentioned above said that the company retains long-term rental of server cabinets and sells some idle computing power according to card usage. At present, the two types of businesses account for 65% and 35% of its revenue, respectively. The latter also includes model fine-tuning, digital humans, and other businesses, not all of which come from short dramas. To accommodate flexible orders, the company has upgraded its task scheduling, resource isolation, and metering systems.

A staff member from the Suzhou Data Bureau told the Economic Observer that AI video and short drama content teams currently do not have the highest proportion of local computing power consumption, but their growth rate is relatively fast among all customer types.

An Unavoidable Bill

The cost structure changes accordingly.

The pressure felt by Yu Shengqi comes from the two ends of production pricing and computing power expenses.

According to his recollection, in the early days, the production price of AI short dramas could reach four to five thousand yuan per minute. Nowadays, orders worth several hundred yuan per minute are not uncommon in the market. At the same time, his team's computing power expenditure per minute has increased from about 100 yuan at the beginning of the year to 200 to 300 yuan. For projects that require full process AI generation and repeated re drawing of lenses, the computational cost already accounts for 35% to 50% of the production cost, compared to just over 10% in the early stages.

Yu Shengqi said that this account is not only affected by resource prices, but also depends on how many times it needs to be generated to obtain usable images.

He began to reduce high difficulty and large-scale scenes in the script stage, trying to minimize the number of repetitions. But in his opinion, as long as the mass is maintained, there are boundaries to throttling.

The business leader of a listed computing power leasing company told Economic Observer reporters that since the beginning of this year, the spot rental prices of high-end inference GPUs they have been exposed to have increased, and the increase in demand for video generation is one of the reasons. But GPU card rental and the comprehensive computing power cost per minute are two different measures. The latter is also affected by the model calling method, bandwidth cost, and the number of repetitions, and cannot attribute all the increase in downstream costs to hardware price increases.

The person in charge said that the prices of long-term framework agreement clients are relatively stable, and small and medium-sized teams that temporarily purchase resources are more susceptible to fluctuations in the spot market.

Yu Shengqi said that even if some projects have good playback data, after deducting labor, streaming, and computing power expenses, there is little profit left. He is now more inclined to accept platform customized orders, which require higher image quality and character consistency, and offer relatively better prices. He also utilized the ecosystem of a comic drama creation platform to undertake IP (intellectual property) adaptation projects, using existing scripts and character settings to reduce trial and error, and retaining a small amount of overseas comic drama business.

Some enterprises still hope to receive policy subsidies support.

According to Suzhou's "Several Measures to Further Accelerate the Construction of 'Artificial Intelligence+' Cities (2026 Edition)" (Su Fu Gui Zi [2025] No. 18), technology research and development enterprises that meet the conditions will be subsidized at 30% of their actual payment for intelligent computing power rental fees, with an annual subsidy limit of 2 million yuan per enterprise.

The person in charge of operating the computing power public service platform mentioned above stated that some project-based studios do not have local legal entities, while other teams purchase computing power in a scattered manner and lack verifiable transaction and usage records on the platform, which poses obstacles in their subsidy application process. Relevant departments and platforms are trying to help small and micro teams understand the application requirements through policy promotion and process optimization.

Don't gamble on the trend

After the machine was settled down, Yu Shengqi's attitude towards selecting orders became more cautious, and the team left a small piece of overseas anime business. Overseas payment sharing is more friendly than some domestic commercial orders.

The person in charge of the hosting company mentioned above said that after the large model training client initiates the task, it usually continues to occupy resources, and the project is calculated on a weekly or monthly basis with a relatively long contract. Short play customers will focus their computing power with the delivery progress, and release resources quickly after going online. He has seen teams betting on a certain genre, but the investment falls short of expectations, the project is cancelled, and the computing power demand immediately drops to zero.

Therefore, he positioned the short drama clients as supplementary needs for idle computing power: maintaining stable income through long-term contract business, and then using idle resources to undertake short-term orders. He did not specifically build a new computer room.

The person in charge of the listed company mentioned above also stated that AI short drama orders currently account for a limited proportion of the company's revenue and are not sufficient to determine the overall profitability level. The company plans to expand customers through the CALI platform, IDC (Internet Data Center) partners and content ecological platform, focusing on docking with cultural and creative enterprises with stable project reserves and overseas short drama companies, so as to reduce the management cost and payment collection risk of sporadic orders.

He reminded that fluctuations on the content side will propagate along the industry chain. If a large amount of hardware is purchased for short-term needs, expenses such as equipment depreciation and cabinet rental will still exist after the order is withdrawn. For this new business, he summarized the company's strategy as follows: "Take advantage of the trend and do not gamble on the trend

(Author Wang Yajie)