The "rent-collecting business" of AI computing power is going through a sea change.
Zhou Yong, a "computing power landlord" with tens of billions of yuan in orders, has taken the initiative to reach out to his key clients since the end of the second quarter of this year to strike a rarely seen deal in the industry.
He lowered the long-implemented fixed cabinet rental service fee for GPU (Graphics Processing Unit) computing power, replacing it with a revenue-sharing scheme that grows along with the call volume of AI Tokens (the unit of measurement for AI workload).
This is not an isolated case.
Reporters from Economic Observer recently interviewed multiple enterprises upstream and downstream in the computing power industry chain and learned that many computing power service providers have followed up with similar practices.
For example, some computing power service providers have cut the fixed service fee by about 10% in the contract, while raising the Token revenue share by more than 200%.
Zhou Yong told the Economic Observer reporter: "This is not simply lowering the base rent, but taking the reduction of fixed guaranteed returns as consideration in exchange for the qualification to share the incremental returns of downstream inference business. In essence, it is the reconstruction of the internal revenue structure of long-term orders."
Signals from the consumer side have also been transmitted simultaneously. During this year's "Double 11" shopping promotion, the "workload" of artificial intelligence has become a commodity on the shelves. Public information shows that top live streamers sold nearly 100,000 large model Token monthly cards in half an hour in their live rooms, and about 2 trillion Tokens were purchased by ordinary consumers.
Zhou Yong said that although flexible pricing methods such as pay-as-you-go, hourly card rental, and short-term packages have existed in the industry before, these models only charge according to resource consumption and do not participate in the profit sharing of downstream business increments. The core difference between the emerging new business model and the previous models is that on the basis of a guaranteed minimum, it binds Token output, shares the dividends of inference growth, and shares risks together.
In his view, the pricing logic of the computing power business is being rewritten.
Computing Power Landlords Set Sights on Profit Sharing
Since the first quarter of 2026, the computing power pricing scheme of "guaranteed minimum + excess Token share" has been piloted and implemented between leading domestic computing power rental service providers and leading large model and Agent application manufacturers. At present, it is still a small-scale exploration, focusing on strategic customers with a high proportion of inference business and expected growth in call volume, and has not yet been extended to small and medium-sized customers and training-type orders. The enterprise where Zhou Yong works is one of the service providers that tested the water earlier.
After raising the amount of computing power lease contracts twice within the year, the computing power service provider where Zhou Yong works terminated the original contract with a leading customer and re-signed a new one.
Zhou Yong provided the reporter with a set of calculation comparisons. For example, under the original pure fixed long-term rental model, the customer's annual rigid computing power expenditure is about 120 million yuan. The new contract reduces the basic guaranteed service fee by about 10% to 108 million yuan. When the customer's monthly Token call volume exceeds the agreed threshold, the excess part will be divided additionally according to the agreed proportion. According to the neutral scenario calculation, if the customer's inference call volume rises steadily, the service provider's annual comprehensive income is expected to rise by 20% to 50% compared with the original fixed rent scheme. However, if the downstream model application cold start fails to meet expectations and the call volume remains below the baseline for a long time, the service provider's annual revenue may drop by up to 10%, and the customer can also avoid high rigid expenditures in the early stage.
In the past, computing power rental was mainly priced around hardware, charging by cabinet, GPU duration or computing power usage, and the lessor usually charged service fees according to resource usage. Many interviewed enterprises have rarely seen contracts where landlords share the revenue of inference business according to call volume and share risks with the customer's business before.
The person in charge of computing power procurement at a leading cloud vendor told the reporter that the call volume of enterprise Agent (AI programs that can perform tasks autonomously) and industry large models is growing rapidly, while the growth rate of general casual chat scenarios is relatively gentle. The structural change on the demand side makes customers want to reduce the upfront fixed costs, while computing power service providers hope to share the benefits brought by the growth of inference business.
Wang Daowei, head of Token business at a listed computing power service provider in South China, told the reporter that the company has included share clauses in new contracts. One type of project adopts MaaS (Model as a Service) joint operation profit sharing, where the service provider provides the computing power base and cluster operation and maintenance, and the model manufacturer provides model capabilities, and the two parties share profits according to Token output; the other type of project directly sells Tokens, and overseas orders adopt annual guaranteed minimum plus post-pay-as-you-go.
Wang Daowei said: "More than one peer is doing similar pilots, but the degree of disclosure varies from one to another."
Zhou Yong explained that the core of this adjustment is to exchange part of fixed income for deeper customer binding. The growth of the customer's business is relatively certain, but they do not want to bear too heavy rigid costs in the early stage. Moderately reducing the fixed service fee is essentially to exchange for deeper customer binding in business games and lock the chassis of long-term cooperation.
The newly signed contract is usually divided into two layers. One is the basic fixed service fee, which retains the certainty of income; the other is Token sharing, which is used to share the incremental income brought by the growth of inference business. Zhou Yong believes that the relevant share is agreed at a fixed amount per computing power unit, and does not directly fluctuate with the customer's own operating revenue.
In other words, what the lessor bets on is not how much money the model manufacturer finally makes, but how many Tokens the machine actually runs.
Before determining the terms, the enterprise where Zhou Yong works has calculated scenarios such as call volume exceeding expectations and business falling short of expectations. He said that at present, the traditional fixed-duration lease is still its main business, and the sharing model is mainly piloted among leading strategic customers. The two models are advancing in parallel and have not replaced the existing contracts.
New Model
For computing power service providers, sharing is not a single template, and customers are combining different resources according to their business stages.
The person in charge of computing power and inference business at a leading large model company told the reporter that the company is adopting a hybrid computing power architecture. Its own cluster undertakes core, stable long-term business, and external computing power is used for elastic expansion. Fixed long-term rental provides cost certainty for mature businesses, while pay-as-you-go resources serve newly launched Agent scenarios with large traffic fluctuations.
The aforementioned person in charge of computing power procurement of the cloud vendor also said that the traditional fixed-duration lease is suitable for businesses with uncertain traffic and that need to control costs in advance. The "guaranteed fixed plus Token sharing" model is more suitable for the cold start stage, which can reduce the upfront rigid expenditure. However, if the call volume grows rapidly, the total expenditure under the hybrid model may exceed that of traditional long-term rental, and the business data of both parties and the Token measurement caliber will also be deeply bound, increasing the migration cost of replacing suppliers.
Computing power landlords are most afraid of idling equipment, while customers are most afraid of bearing excessively high fixed costs before their business is proven feasible. The sharing model allows both parties to share part of the uncertainty.
Zhou Yong said that the shared revenue will be affected by the delivery and acceptance of computing power units and the activity level of customer business, "the risk is indeed shared in both directions". This means that service providers give up not only part of the fixed cash flow, but also the originally relatively certain income expectation.
The model change has also been transmitted to chip manufacturers.
A business person from a leading domestic AI chip manufacturer told the reporter that the focus of assessment under the sharing model is no longer just the nominal peak performance of the chip, but also how many Tokens are actually output per unit of hardware and the cost of a single Token. This provides a new evaluation scale for domestic chips, and also requires chip manufacturers to further improve software adaptation, cluster interconnection and inference optimization.
The person added that the iteration cycle of AI chips is roughly 1.5 to 2 years, and the accounting depreciation cycle may reach 3 to 4 years, but the commercially available life cycle is often shorter. If the actual output of hardware fails to meet expectations, the project revenue of lessors will be directly affected.
Wang Daowei judges that sharing will become an important supplementary model for the industry, but it is difficult to completely replace the fixed-price long-term rental in the short term. "It is not yet at a comprehensive inflection point, and the whole is still in the small-scale pilot stage at present." He predicts that leading strategic customers and projects with a high proportion of inference business will more adopt the hybrid scheme of guaranteed minimum plus sharing; ordinary customers and training-type businesses will still be dominated by traditional long-term rental.
Risks
The first to tighten is the capital side.
The person in charge of computing power financing business in the corporate business department of a joint-stock bank told the reporter that the risk of computing power leasing projects is significantly higher than that of ordinary manufacturing industries. Banks will raise the risk premium and require combined credit enhancement (additional risk protection measures) such as equipment mortgage, joint and several guarantee by the actual controller, guarantee by the listed company entity, and accounts receivable pledge.
"The first source of repayment is the cash flow of the downstream real business." The aforementioned bank person said that the number of GPUs itself cannot represent the solvency, and the key points of due diligence (investigation and verification before loan disbursement) are customer qualification, the quality of formal contracts and the actual online rate of equipment.
For anonymous customers, banks will verify clauses such as minimum procurement volume, default compensation, and payment rhythm, distinguish between intentional framework agreements and binding formal contracts, cross-check the payment return flow, and go to the computer room to verify the equipment operation status when necessary. Accounts receivable formed by computing power services will also be included in the pledge and payment supervision.
If the counterparty is anonymous throughout the process and lacks third-party verification, credit access will be significantly tightened. Public information shows that some cross-border enterprises have experienced early termination or regulatory inquiries after disclosing large computing power contracts, which also makes banks pay more attention to the main business safety cushion of enterprises, the capability of the industrial team and contract performance.
The aforementioned bank person said that the circulation market for second-hand AI equipment is not yet mature, and chips are iterating rapidly, so there is great uncertainty in the disposal and realization of equipment, and it is impossible to expect the auction of hardware to cover all exposures (the amount of risk not yet recovered by the bank). Therefore, pre-loan access and in-loan monitoring are more critical than post-event disposal.
There is also fluctuation on the demand side.
The aforementioned person in charge of computing power and inference business of the leading large model company said that the growth rate of Agent-related calls is significantly higher than that of ordinary dialogue scenarios, but there are large differences among different customers and industries. The current growth is characterized by "steady rise in some scenarios superimposed with periodic pulses": the call volume of mature businesses increases gradually, while the pilot projects may fall back after termination.
This is also the biggest variable of the sharing model: when downstream calls rise, service providers can obtain excess returns. If the business cools down, the shared revenue shrinks, but rigid expenditures such as depreciation, electricity and operation and maintenance will not stop.
Token sharing has thus become a new choice for computing power leasing, but whether the machine can run continuously and whether the customer's business can continue to grow still determines whether this business can be established in the end.
This article is from WeChat official account "Economic Observer", author: Wang Yajie, authorized for release by 36Kr.