Economic Observer reporter Wang Yajie

Zhou Yong, a "computing power landlord" with orders worth tens of billions of yuan, took the initiative to approach his major clients since the end of the second quarter of this year and completed a rare deal in the industry.

He reduced the rental service fee for the long-term operation of the GPU (graphics processing unit) dedicated cabinets, and replaced it with a revenue-sharing scheme that is based on the increasing usage volume measured in Token (the unit of measurement for artificial intelligence workload).

This is not an isolated case.

A reporter from Economic Observer recently interviewed several enterprises along the upstream and downstream of the computing power industry chain and learned that some computing power service providers have already followed suit with similar practices.

For example, some computing power service providers have reduced their fixed service fees by about 10% in the contract, while the revenue sharing of Token has increased by more than twice.

Zhou Yong told the Economic Observer reporter, "This is not simply about lowering the base rent, but rather about exchanging the compression of fixed guaranteed earnings for the right to share in the incremental earnings from downstream reasoning services. Essentially, it is a restructuring of the long-term order internal earnings structure."

The signals from the consumer side are also transmitted synchronously. During this year's "Double Eleven" promotional period, the workload of artificial intelligence became evident in the products on the shelves. Public information shows that top live streamers sold nearly one hundred thousand monthly subscriptions to the large model Token in just half an hour, with approximately two trillion Token being purchased by ordinary consumers.

Zhou Yong stated that although there have been previous flexible pricing models in the industry such as pay-as-you-go, time-based leasing, and short-term packages, these models only charged based on resource consumption and did not participate in the profit sharing of downstream business growth. The core difference between the new business model that has emerged now and the previous ones is that this time, on top of a guaranteed base, it binds to Token output and shares the benefits of increased reasoning, with risks being shared together.

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

Mining power landlords set their sights on dividends

Starting from the first quarter of 2026, a computing power pricing scheme that includes a guaranteed minimum plus excess commissions based on Token has begun to be piloted among leading domestic computing power rental service providers, major large model developers, and Agent application manufacturers. Currently, this scheme is still in the early stages of exploration and is focused on strategic customers with a high proportion of inference services and expected growth in usage volumes. It has not yet been extended to small and medium-sized customers or training-related orders. The company where Zhou Yong works is one of the earlier adopters of this pilot program.

After raising the amount of the computing power lease contract twice within the year, the computing power service provider where Zhou Yong works terminated the original contract with a leading customer and signed a new one.

Zhou Yong provided reporters with a set of comparative calculations. 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 reduces the basic minimum service fee by about 10% to 108 million yuan. When the customer's monthly Token usage exceeds the agreed threshold, the excess amount is shared additionally according to the agreed ratio. Based on a neutral scenario, if the customer's usage steadily increases, the service provider's annual comprehensive income is expected to increase by 20% to 50% compared to the original fixed rental plan. However, if the downstream model applications do not start up as expected and the usage remains below the baseline for a long time, the service provider's annual income could potentially decline by up to 10%, while the customer can also avoid high initial rigid expenditures.

In the past, computing power leasing was mainly priced based on hardware, with charges calculated per cabinet, per GPU duration, or per amount of computing power used. Lenders typically charged a service fee based on the usage of resources. Several companies interviewed mentioned that they had rarely seen contracts where landlords shared the profits from inference services with tenants and jointly borne risks with them.

The person in charge of computing power procurement at a leading cloud provider told reporters that the usage of Agent (artificial intelligence programs capable of executing tasks autonomously) and large industry models is increasing rapidly, while the growth in general casual chat scenarios is relatively slower. The structural changes on the demand side have led customers to wish to reduce upfront fixed costs, while computing power service providers hope to share in the benefits brought by the growth in inference services.

Wang Daowei, the person in charge of Token business at a computing power service provider listed in South China, told reporters that the company has included profit-sharing terms in the new contract. For one type of project, they adopt a MaaS (model as a service) joint profit-sharing model, where the service provider provides the computing power infrastructure and cluster operations, while the model manufacturer provides the model capabilities. Both parties share the profits according to Token; for another type of project, they directly sell Token. Overseas orders are based on an annual minimum guarantee plus pay-per-use post-payment.

Wang Daowei said, 'More than one competitor is conducting similar pilot projects, it's just that the degree of disclosure varies from company to company.'

Zhou Yong explained that the core of this adjustment is to exchange a portion of fixed income for deeper customer loyalty. While the certainty of customer business growth is relatively strong, there is also a desire not to bear excessive fixed costs in the early stages. Moderately reducing fixed service fees is essentially a strategic move in business negotiations to gain deeper customer engagement and secure a foundation for long-term cooperation.

The newly signed contract is usually divided into two layers. One layer consists of a basic fixed service fee, which ensures the certainty of income; the other layer is divided according to the Token method, which is used to share the incremental profits brought about by the growth of the reasoning business. Zhou Yong believes that the relevant split is based on a fixed amount per computing unit and does not directly fluctuate with the customer's own operating income.

In other words, what the lessor is betting on is not how much money the model manufacturer ultimately makes, but how many Tokens the machine actually runs.

Before finalizing the terms, the company where Zhou Yong works conducted calculations for scenarios such as usage exceeding expectations and business falling short of expectations. He said that currently, traditional fixed-duration leasing remains their main business, while the revenue-sharing model is mainly being piloted among key strategic customers. Both models are being advanced in parallel and have not yet replaced existing contracts.

New Model

For computing power service providers, the division is not based on a single template; customers are combining different resources according to the stages of their business.

The person in charge of computing power and inference business at a leading large-model company told reporters that the company is adopting a hybrid computing power architecture. The company's own clusters handle core, stable long-term businesses, while external computing power is used for elastic scaling. Fixed long-term rentals provide cost certainty for mature businesses, while pay-as-you-go resources serve new projects with larger traffic fluctuations.

The person in charge of computing power procurement from the aforementioned cloud providers also stated that traditional fixed-duration leasing is suitable for businesses with uncertain traffic that need to control costs in advance. The "guaranteed fixed fee plus Token commission" model is more suitable for the cold start phase, as it can reduce initial rigid expenses. However, if the usage volume grows rapidly, the total expenses under the hybrid model may exceed those of traditional long-term leasing. Additionally, the business data of both parties and the Token measurement criteria will be closely linked, increasing the migration costs associated with changing suppliers.

Miners fear that their equipment will run idle, while customers are most concerned about incurring high fixed costs before their businesses are fully operational. The revenue-sharing model allows both parties to share some of the uncertainties.

Zhou Yong stated that the distribution of income is affected by the delivery and acceptance of computing power units as well as the activity level of customers, meaning that 'the risks are indeed shared in both directions.' This implies that service providers not only give up a portion of their fixed cash flows but also their relatively certain income expectations.

The change in patterns also affects chip manufacturers.

A business executive from a leading domestic artificial intelligence chip manufacturer told reporters that the assessment criteria under the performance-based incentive model no longer focus solely on the theoretical peak performance of the chips, but also include how much Token is actually produced per unit of hardware, as well as the cost per individual Token. This provides a new benchmark for evaluating domestic chips and also requires chip manufacturers to further improve software adaptation, cluster interconnection, and inference optimization.

The individual added that the iteration cycle for artificial intelligence chips is roughly one and a half to two years, while the accounting depreciation cycle can reach three to four years, but the commercial usable life cycle is often shorter. If the actual hardware output falls short of expectations, the project profits of the lessors will be directly affected.

Wang Daowei believes that the split-payment model will become an important supplementary approach in the industry, but it is difficult to completely replace the fixed-price long-term rental model in the short term. “We can’t yet talk about a complete turning point; at present, we are still in the stage of small-scale pilot projects.” He expects that for projects with top strategic clients and a high proportion of reasoning-based services, a hybrid model combining a guaranteed minimum payment with a commission will be more commonly adopted; for ordinary clients and training-related services, traditional long-term rentals will still be the main option.

Risks

The first to tighten measures was on the funding side.

The person in charge of the computing power financing business in the corporate business department of a joint-stock bank told reporters that the risks associated with computing power leasing projects are significantly higher than those of ordinary manufacturing industries. The bank will increase the risk premium and require additional risk mitigation measures such as equipment collateral, joint guarantees from actual controllers, guarantees from listed companies, and pledged accounts receivable.

"The primary source of repayment is the real cash flow from downstream business operations," said the banking official mentioned above. The quantity of GPU itself does not represent solvency; customer qualifications, the quality of formal contracts, and the actual deployment rate are the key focuses of due diligence (the investigation and verification before loan disbursement)."

For anonymous customers, banks will verify terms such as minimum purchase quantities, breach of contract compensation, and payment schedules. They will distinguish between intent-based framework agreements and binding formal contracts, and cross-check payment receipts. If necessary, they will visit the data center to verify the operational status of the equipment. Accounts receivable generated from computing power services will also be subject to pledge and payment collection supervision.

If the counterparty remains anonymous throughout the process and there is a lack of third-party verification, credit approval will be significantly tightened. Public information indicates that some cross-industry companies have encountered early terminations of large-scale computing power contracts or regulatory inquiries after disclosing such contracts, which has also made banks pay more attention to the safety reserves of these companies' main businesses, the capabilities of their industry teams, and their compliance with contract obligations.

The aforementioned banking professionals stated that the secondary market for artificial intelligence equipment is not yet mature, and chip iterations are rapid. There is considerable uncertainty in the disposal and monetization of such equipment, so it is not feasible to expect auctions of hardware to cover all of the exposed risks (the amounts that banks have not yet recovered). Therefore, pre-loan approval and ongoing monitoring during the loan period are more critical than post-loan disposal.

There are also fluctuations on the demand side.

The person in charge of computing power and inference business at the aforementioned large-scale model companies stated that the growth rate of Agent related calls is significantly higher than in ordinary dialogue scenarios, but there are significant differences among different customers and industries. The current growth pattern is characterized by a steady increase in some scenarios, coupled with periodic spikes. The number of calls for mature businesses is gradually increasing, while pilot projects may see a decline after they are terminated.

This is also the largest variable in the revenue-sharing model; only when downstream calls increase can service providers obtain excess profits. However, if business activity declines and revenue from revenue sharing shrinks, fixed expenses such as depreciation, electricity costs, and operational maintenance will not cease.

Token has become a new option for computing power leasing, but whether the machines can continue to operate and whether customer business can continue to grow will ultimately determine whether this business can be successful.

At the request of the interviewee, Zhou Yong is using a pseudonym.

(Author: Wang Yajie)