Economic Observer reporter Wang Yajie

Zhou Yong, the "computing power landlord" with orders worth billions of yuan, has actively approached his major clients since the end of the second quarter of this year and made a rare transaction in the industry.

He lowered the rental service fee for fixed cabinets with long-term GPU (graphics processing unit) computing power and replaced it with a revenue sharing plan that follows the growth of call volume for artificial intelligence tokens (a unit of measurement for artificial intelligence workload).

This is not an isolated case.

Economic Observer reporters recently interviewed multiple companies in the upstream and downstream of the computing power industry chain and learned that computing service providers have followed up with similar practices.

For example, some computing service providers have reduced their fixed service fees by about 10% in their contracts, while increasing their token revenue sharing by more than twice.

Zhou Yong told the Economic Observer reporter, "This is not simply lowering the basic rent, but in exchange for the qualification to share the incremental income of downstream inference business by compressing the fixed guaranteed income. Essentially, it is a restructuring of the internal income structure of long-term orders

The signal from the consumer end is also transmitted synchronously. During this year's "Double Eleven" promotion, the "workload" of artificial intelligence has become a commodity on the shelves. Public information shows that the top live streaming hosts sold nearly 100000 large model token monthly cards in half an hour, and about 2 trillion tokens were purchased by ordinary consumers.

Zhou Yong stated that although there have been flexible pricing methods in the industry such as volume based billing, card time leasing, and short-term packaging, these models only charge based on resource consumption and do not participate in downstream business incremental profit sharing. The core difference between the new business model that has emerged and the previous model is that this time it is based on a minimum guarantee, binding token output, sharing reasoning growth dividends, and risk sharing.

In his view, the pricing logic of computing power business is being rewritten.

The computing power landlord is targeting dividends

Starting from the first quarter of 2026, the guaranteed+Token excess sharing computing power pricing scheme will be piloted and implemented among top domestic computing power leasing service providers, as well as top large model and agent application vendors. Currently, it is still a small-scale exploration, focusing on strategic customers with a high proportion of inference business and expected growth in the use of measurement tools. It has not yet been rolled out to small and medium-sized customers and training orders. Zhou Yong's company is one of the early service providers to test the waters.

After twice increasing the amount of the computing power leasing contract within the year, Zhou Yong's computing power service provider terminated the original contract and re signed it with a top customer.

Zhou Yong provided a set of calculations and comparisons to the reporter. For example, under the original pure fixed long-term rental model, the customer's annual rigid computing power expenditure was about 120 million yuan. The new contract reduced the basic guaranteed service fee by about 10% to 108 million yuan. When the customer's monthly token usage exceeds the agreed threshold, the excess will be divided into additional shares according to the agreed proportion. According to the neutral scenario calculation, if the number of customer inference calls steadily increases, the service provider's annual comprehensive income is expected to increase by 20% to 50% compared to the original fixed rent plan. But if the cold start of downstream model applications falls short of expectations and the call volume remains below the baseline for a long time, the annual revenue of service providers may decline by up to 10%, and customers can also avoid high rigid expenses in the early stage.

In the past, computing power leasing mainly focused on hardware pricing, charging based on server cabinets, GPU hours, or computing power usage. The lessor usually charged service fees based on resource usage. Previously, the interviewed companies had seen few contracts where landlords shared the benefits of inference services based on the amount of demand, and shared risks with customers' businesses.

The head of computing power procurement at a leading cloud vendor told reporters that the use of enterprise agents (AI programs capable of autonomously executing tasks) and industry models is growing rapidly, while the growth rate of general chat scenarios is relatively flat. The structural changes on the demand side make customers hope to reduce fixed costs in the early stage, while computing power service providers hope to share the benefits brought by the growth of inference business.

Wang Daowei, the head of Token business for a listed computing power service provider in South China, told reporters that the company has included the sharing clause in the new contract. A type of project adopts MaaS (Model as a Service) joint venture profit sharing, where the service provider provides the computing power base and cluster operation and maintenance, the model vendor provides the model capability, and both parties generate profit sharing according to the Token output; Another type of project sells tokens directly, and overseas orders adopt an annual minimum guarantee plus pay as you go policy.

Wang Daowei said, "More than one peer is conducting similar pilot projects, but the degree of disclosure varies among them

Zhou Yong explained that the core of this adjustment is to exchange a portion of fixed income for deeper customer binding. The certainty of customer business growth is relatively strong, but they also do not want to bear excessive rigid costs in the early stage. Moderately reducing fixed service fees is essentially a business game to gain deeper customer loyalty and lock in long-term cooperation opportunities.

The re signed contract is usually divided into two layers. The first layer is the basic fixed service fee, which preserves the certainty of income; The other layer is Token sharing, which is used to share the incremental benefits brought by the growth of inference business. Zhou Yong believes that the relevant revenue sharing is based on a fixed amount agreed upon by the computing power unit, and does not directly fluctuate with the customer's own operating income.

In other words, the lessor is not betting on how much money the model manufacturer ultimately earns, but on how many tokens the machine actually runs over.

Before determining the terms, Zhou Yong's company conducted calculations on scenarios such as exceeding expectations in call volume and falling short of expectations in business. He said that currently, traditional fixed duration leasing is still its main business, and the revenue sharing model is mainly piloted among top strategic customers. The two modes are advancing in parallel and have not yet replaced existing contracts.

New Model

For computing service providers, the sharing is not a single template, and customers are combining different resources according to business stages.

The head of computing and inference business at a leading large modeling company told reporters that the company is adopting a hybrid computing architecture. Our own cluster undertakes core and stable long-term business, while external computing power is used for elastic expansion. Fixed long-term leasing provides cost certainty for mature businesses, while on-demand resources serve new agent scenarios with high traffic fluctuations.

The head of computing power procurement for the aforementioned cloud vendors also stated that traditional fixed duration leasing is suitable for businesses with uncertain traffic and the need to control costs in advance. The "guaranteed fixed amount plus token sharing" is more suitable for the cold start stage, which can reduce rigid expenses in the early stage. But if the volume of transfers grows rapidly, the total expenditure in the hybrid mode may exceed that of traditional long-term leasing, and the business data and token measurement caliber of both parties will be deeply bound, increasing the migration cost of changing suppliers.

The computing power landlord is most afraid of equipment idling, and the customer is most afraid of bearing high fixed costs before the business is fully operational. The revenue sharing model allows both parties to share some of the uncertainty.

Zhou Yong stated that revenue sharing will be affected by the delivery and acceptance of computing power units and customer business activity, and "risks are indeed shared in both directions". This means that the service provider is not only giving up a portion of fixed cash flow, but also includes the originally more certain revenue expectations.

The mode change is also transmitted to chip manufacturers.

A business representative from a leading domestic artificial intelligence chip manufacturer told reporters that the focus of the revenue sharing model assessment is no longer just on the peak performance of the chip on paper, but also includes the actual output of tokens per unit of hardware and the cost of a single token. This provides a new evaluation standard 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 artificial intelligence chips is roughly one and a half to two years, and the accounting depreciation cycle may reach three or four years, but the commercially available life cycle is often shorter. If the actual output of hardware falls short of expectations, the project revenue of the lessee will be directly affected.

Wang Daowei believes that profit sharing will become an important supplementary model for the industry, but it is difficult to completely replace fixed price long-term leasing in the short term. We can't talk about a comprehensive turning point yet. Currently, the overall project is still in the small-scale pilot stage. He expects that projects with a high proportion of top strategic customers and inference businesses will adopt a mixed solution of guaranteed bottom line and profit sharing; Ordinary customers and training businesses will still be dominated by traditional long-term leasing.

Risk

The first to tighten is the funding side.

The head of the computing power financing business of a joint-stock bank's corporate business department told reporters that the risks of computing power leasing projects are significantly higher than those of ordinary manufacturing industries. Banks will increase the risk premium and require a combination of equipment collateral, joint guarantee from actual controllers, main guarantee from listed companies, and accounts receivable pledge to enhance creditworthiness (additional risk protection measures).

The first source of repayment is the actual cash flow from downstream businesses, "said the bank insider." The number of GPUs itself cannot represent debt repayment ability, and customer qualifications, formal contract quality, and actual listing rate are the focus of due diligence (investigation and verification before loan disbursement).

For anonymous customers, banks will verify terms such as minimum purchase quantity, default compensation, payment rhythm, etc., distinguish between intentional framework agreements and binding formal contracts, cross check payment receipts, and if necessary, go to the data center to verify equipment operation status. Accounts receivable generated from computing power services will also be subject to pledge and collection supervision.

If the counterparty remains anonymous throughout the process and lacks third-party evidence, credit access will be significantly tightened. Public information shows that some cross-border enterprises have terminated or received regulatory inquiries in advance after disclosing large computing power contracts, which has also made banks pay more attention to the safety cushion of the enterprise's main business, the ability of the industrial team, and the performance of the contract.

The above-mentioned bank personnel stated that the second-hand AI equipment circulation market is not yet mature, the chip iteration is fast, and there is significant uncertainty in the disposal and realization of equipment. It cannot be expected that the auction hardware will cover all exposures (the risk amount that the bank has not yet recovered). Therefore, pre loan access and in loan monitoring are more crucial than post loan disposal.

There are also fluctuations on the demand side.

The person in charge of computing power and inference business at the top model company mentioned above said that the growth rate of Agent related calls is significantly higher than that of ordinary dialogue scenarios, but there are significant differences among different customers and industries. The current growth is characterized by a steady increase in some scenarios combined with periodic pulses, with mature business calls gradually increasing, while pilot projects may experience a decline after termination.

This is also the biggest variable in the revenue sharing model, where downstream calls go up and service providers can obtain excess returns. If the business cools down and revenue shrinks, rigid expenses such as depreciation, electricity, and operation and maintenance will not stop.

Token sharing has become a new option for computing power leasing, but whether machines can continue to operate and customer business can continue to grow still determines whether this business can ultimately be established.

(At the request of the interviewee, Zhou Yong uses a pseudonym)

(Author Wang Yajie)