Computing power leasing: A lucrative track for leading players

In May 2026, in Wuxi, Jiangsu Province. Zhang Lei, the business director of a leading computing power leasing enterprise, has just signed a three - year long - term agreement worth 5 billion yuan. The customer is a leading Internet giant. The rental rate of the high - end computing power cluster in his hands has reached 100%, and the orders have been scheduled until 2028.

In Zhongwei, Ningxia, 1300 kilometers away, Ma Hongyuan, the owner of a small computing power enterprise, is staring at the server monitoring on his mobile phone. He only has a few hundred P (a unit of computing power, 1P is 1PFLOPS, approximately equal to one quadrillion floating - point operations per second) of domestic computing power. His customers are small startup teams and individual developers in fields such as artificial intelligence and software development. He has calculated that if the utilization rate of computing power is lower than 70%, he will surely incur losses.

Back in Wuxi, Lee, the AIDC (Artificial Intelligence Data Center) business director of Wuxi Lianyun Century Technology Co., Ltd. (871315, hereinafter referred to as "Lianyun Century"), a company listed on the New Third Board, describes the current industry as follows: "Currently, the domestic computing power leasing industry is in a period of explosive growth dividends. The market demand is continuously expanding, but the supply of high - quality computing power resources is relatively scarce. The industry competition has reached a white - hot stage. Facing the fierce competition among a large number of homogeneous service providers, we must match resources one step ahead."

Public information shows that Lianyun Century is a computing power leasing service provider. It obtains computing power resources in bulk from upstream suppliers and then provides them to enterprise customers in need in the form of leasing.

In the government building of a certain city in the Yangtze River Delta, Lao Zhou, who is in charge of attracting investment in the computing power industry, is reviewing the materials of another computing power enterprise applying for settlement. In the past two years, he has removed more than a dozen speculators who "took the subsidies and ran away", and has also witnessed several intelligent computing centers that were left unfinished after investing more than one billion yuan in construction funds. "The 'wind' is too strong now. We can't let just anyone rush in," he said.

The rise of computing power leasing started from the global large - model craze triggered by Chat GPT in 2023. In the past three years, domestic artificial intelligence (AI) applications have moved from the "hundred - model battle" to industrial implementation, and the demand for computing power has exploded accordingly. Data from the China Academy of Information and Communications Technology shows that the scale of the computing power leasing market has jumped from 32.8 billion yuan in 2025 to 68 billion yuan in the first quarter of 2026, and is expected to exceed 260 billion yuan for the whole year.

The computing power wave has arrived. But the wave doesn't reach everyone evenly. In the computing power track, participants are all competing, competing for resources, customers, and time. And the government, as Lao Zhou put it, is competing for "order".

Competing for Resources

The underlying logic of computing power leasing is, first of all, "whether you have the cards".

"Cards" refer to GPUs (Graphics Processing Units, here referring to computing power chips used for AI training). High - end GPUs, including H100, B100, A800, etc., are the "hard currency" for training large models. However, due to US export controls and chip production capacity limitations, these cards have been in short supply for a long time. Data from the China Academy of Information and Communications Technology shows that in the first quarter of 2026, the gap in high - end intelligent computing reached 35%, and the gap for just one type of H100 card was as high as 430,000 units.

Whoever has the cards has the right to speak.

The enterprise where Zhang Lei works is one of the players at the top of the resource chain. It has unique qualifications. Its card - acquisition cycle is 6 to 12 months shorter than that of its peers, and its cost is 30% to 50% lower than that of small and medium - sized players.

His strategy is clear: only focus on high - end training computing power, and avoid the low - end inference computing power and the domestic card market. When he entered the computing power industry in 2023, he set this iron - clad rule. "The low - end market has an oversupply and a fierce price war, and the gross profit margin is as thin as paper. We have no advantage in entering it. While high - end training computing power is in short supply in the long run due to (US) export controls and chip production capacity limitations."

This determination is first reflected in customer selection. While many small and medium - sized computing power providers are worried about acquiring customers, his service targets are only a few leading Internet giants and top large - model enterprises. He doesn't accept small and medium - sized customers at all.

"Leading customers have sufficient budgets, stable demands, and zero default risks. They can accept high prices and long - term contracts," Zhang Lei said. "What we sell to leading customers is not 'computing power' but 'certainty', including stable high - end computing power, reliable operation and maintenance, continuous technical support, and uninterrupted services."

This certainty has earned his enterprise a three - year long - term agreement worth billions of yuan from a large enterprise. In the first quarter of 2026, the enterprise he works for can dispatch 38,000 P of computing power, including 10,000 P of its own and 28,000 P through sub - leasing. The on - hand orders exceed 7 billion yuan, and the gross profit margin is stable between 53% and 60%.

The enterprise has also self - developed immersion liquid - cooling technology, reducing the PUE (Power Usage Effectiveness) to between 1.09 and 1.1, far lower than the industry average of 1.5. The density of a single cabinet reaches 100 kW (kilowatts), which is twice that of traditional air - cooling, and the power consumption is reduced by 30%. Relying on the precision machining capabilities of traditional manufacturing, it produces its own liquid - cooling hardware, with a gross profit margin of over 60%, forming an integrated closed - loop of "computing power + liquid cooling".

But not everyone can grab resources.

Ma Hongyuan's computing power team has only 8 people, and he holds a few hundred P of domestic computing power, including domestic AI chips such as Ascend 910 and Hygon DCU. He doesn't dare to touch H100 and A800.

Ma Hongyuan said that, firstly, he can't get the quota, and secondly, they are too expensive. One card costs hundreds of thousands of yuan, and he can't let them sit idle.

His survival logic relies on the cost advantage in the western region. The electricity price in the eastern region is 0.8 to 1.2 yuan per kilowatt - hour, while in the Zhongwei area, it is 0.3 to 0.4 yuan per kilowatt - hour, more than half the difference. For 100 P of computing power, the monthly electricity cost in the western region is less than 20,000 yuan, while in the eastern region, it is 30,000 to 50,000 yuan. Coupled with the policy subsidies and site rent exemptions of the "Eastern Data, Western Computing" project, these "hidden profits" form the foundation for his survival.

But he has calculated: if he rents out all 50 P of domestic computing power, at 3,000 yuan per P per month and a utilization rate of 70%, the monthly income is 105,000 yuan. However, the purchase cost is 100,000 yuan, and with cabinet electricity costs, operation and maintenance bandwidth, the monthly cost exceeds 110,000 yuan. "If the utilization rate is lower than 70%, it's a sure loss. With a utilization rate of over 80%, the profit is thin. Only when the utilization rate is over 90% can there be some profit."

Caught between the leading computing power equipment suppliers and the small and medium - sized computing power leasing enterprises in the west is Lianyun Century, headquartered in Wuxi, Jiangsu. In 2025, the company's computing - power - related business revenue was approximately 13.88 million yuan, accounting for 12.05% of the total revenue. Currently, the enterprise mainly integrates upstream computing power resources, supplemented by its own computing power. It is gradually increasing its investment in self - built and self - owned computing power, adopting a service model of "computing power leasing + supporting operation and maintenance + customized solutions".

Lianyun Century has self - built an intelligent computing center in Wuhu, Anhui, planning a cluster of tens of thousands of cards, which will be delivered and put into operation by the end of 2026.

"The outside world defines computing power leasing enterprises as 'computing power sub - landlords'. This positioning only applies to our business model before 2020," Lee said. "Since 2021, we have no longer just engaged in simple light - asset sub - leasing business. We have continuously made heavy investments in the field of AI computing power. Currently, Lianyun Century is self - building a data center in Wuhu, a node of the 'Eastern Data, Western Computing' project. The overall plan covers 120 mu. The first phase of the project is expected to deliver 40 MW by the end of 2026, and the second phase is expected to be delivered in 2027. After the project is officially put into operation, it will completely break away from the single model of the traditional'sub - landlord'."

Competing for Customers

Resources are a matter of the supply side, while customers are a matter of the demand side. With the arrival of the wave, customers are also differentiating.

The top - tier customers are leading Internet giants and top AI model companies. Their demand is "certainty", including stable high - end computing power, reliable operation and maintenance, continuous technical support, and uninterrupted services. They have sufficient budgets and are willing to pay a high premium for this certainty.

Relying on this strategy of "selling certainty", the enterprise where Zhang Lei works has won a three - year long - term agreement worth billions of yuan from a large enterprise, with a 100% rental rate and orders scheduled until 2028.

The second - tier customers are small and medium - sized enterprises in the Yangtze River Delta, including those in the digital transformation of manufacturing, the implementation of AI applications, fintech, and healthcare. These customers have a demand for computing power, but they can't afford to buy, maintain, or operate an entire cluster. What they want is "flexibility, affordability, and someone to manage".

This is the main battlefield of Lianyun Century. Lee's customers cover financial payment enterprises, quantitative private equity companies, leading medical manufacturers, Internet platforms, and game companies. One of the financial payment enterprises he serves has over 30 million daily transactions; a quantitative private equity company manages assets of approximately 4 billion yuan and makes investments based on AI strategies.

Lee's strategy is to take "localized and customized services" to the extreme. "The services provided by large enterprises are standardized and remote, with a slow response. We have a local team, and we can visit customers promptly when they have problems," he said.

Among his customer groups, AI startup teams account for approximately 40%, small and medium - sized enterprises account for approximately 35%, and government agencies, enterprises, and research institutions account for approximately 25%. AI startup teams have strong and volatile demands, short cycles, and tight budgets. They pursue high - end cards and mainly focus on model training; small and medium - sized enterprises have stable demands, long cycles, and pursue cost - effectiveness. They mainly engage in AI inference and data processing; government and enterprise customers have sufficient budgets, pursue stability, security, and compliance, and mainly engage in scientific research and government AI.

In the Yangtze River Delta, another private computing power service provider has also smelled the structural opportunity. The company is based in Hangzhou, providing local 7×24 - hour operation and maintenance, on - site docking, and small - batch customization. The price of its computing power services is more than 30% lower than that of large enterprises. Zhou Bo, an employee of the company, told Economic Observer: "In the past, when talking to customers, they would ask 'What is computing power'. Now, when talking to customers, they immediately ask 'How many H100s do you have? Can you lock in the supply for three months?' Only those who have experienced it can understand this gap."

Zhou Bo's customers are also mainly small and medium - sized enterprises in the Yangtze River Delta. He feels the squeeze from two directions: firstly, large upstream enterprises are moving down to compete for small and medium - sized customers. "They have capital, resources, brands, and technology. Our medium - sized companies simply can't compete with them." Secondly, the supply is becoming increasingly unstable. "For example, we once locked in a batch of A800s, but 30% of the quota was cut by the upstream supplier before delivery, and the price increased by 15%. On the one hand, we have to fulfill the contract with customers, and on the other hand, we have to bear the increased cost. We have to absorb the price difference ourselves."

The third - tier customers are "leftovers", including small local enterprises in the west, small startup teams in the east, and individual developers. They have limited budgets and require short - term rentals and flexible payment methods. Large enterprises don't care about them, and medium - sized companies also find them troublesome.

This is Ma Hongyuan's market. His customers are tough negotiators, saying that others only charge 2,800 yuan per P. He can only explain, "We include operation and maintenance and technical support, with a 7×24 - hour response. Others offer bare rentals, and no one will take care of problems when they occur." After an hour of negotiation, they finally reach a deal at 2,900 yuan per P for a one - month short - term contract. "We can earn some hard - earned money. It's better than leaving the computing power idle," Ma Hongyuan said.

He is also trying to provide value - added services, such as 7×24 - hour monitoring, fault repair, model deployment, and parameter optimization for customers, charging by the time or by the month. "We earn hard - earned money from sub - leasing and profit from services. Without services, small players can't survive," he said.

There are different strategies for the three tiers of customers. Ma Hongyuan said: "The leading players eat the meat, and we drink the soup. There are still some'sand' in the soup, and we have to filter it slowly. But after filtering it clean, we can still quench our thirst."

Competing for Time

After grabbing resources and customers, the remaining thing is to compete for time. Because the industry is changing too fast, if you're slow, you'll miss out.

Zhou Bo uses one word: "Compete". Compete for time, for implementation, and for customers.

Lee describes it like this: "Currently, the entire computing power leasing industry is in a stage of explosive growth and a highly active market." He has a high - intensity workload. He spends most of the day outside, connecting with customers, communicating about needs, and following up on project plans. He can only sort out internal processes and review the day's work at night. This is "the norm for Internet industry practitioners".

His team can visit customers promptly when they have problems. This response speed is something that the standardized services of large enterprises can't achieve.

Zhou Bo added: "What large enterprises find troublesome is our livelihood. But we have to hold this livelihood more and more steadily."

The way the enterprise where Zhang Lei works "competes for time" is different. It doesn't compete for the moment but locks in the next 3 to 5 years with long - term contracts.

"Small and medium - sized players pursue quick entry and exit and short - term arbitrage. We only do ultra - long - term contracts of 3 to 5 years," Zhang Lei said. "The core value of long - term contracts is, firstly, to lock in high gross profit margins. In the first quarter of 2026, the rent of high - end computing power increased by 20% to 30%. Long - term contracts directly lock in the high gross profit margins after the price increase. Secondly, the cash flow is extremely certain. Customers pay a 30% to 50% deposit in advance and make monthly payments. The advance payments can cover approximately 60% of the equipment cost."

But he also has concerns. The iteration cycle of high - end GPUs is only two or three years. Once new cards are released, the prices of old cards will plummet. At the end of 2025, NVIDIA released the new B100 card. He signed all the H100s for three - year long - term contracts in advance and got the B100 quota first, so he smoothly passed through the technology iteration period.

"In the short term, it's a game. In the long term, it's a monopoly. Resources are concentrated in the hands of the leading players. Eventually, a few leading enterprises will monopolize the high - end market.