The most dangerous signal is to speculate on GPUs like real estate.
Why does the local state-owned asset platform in Wuhu emphasize light-asset operation in a capital-intensive sector? When regions engage in a subsidy competition to attract computing power enterprises, how can we ensure that the money is spent effectively? Against the backdrop of restrictions on high-end chips, when will domestic computing power truly be able to step in?
In 2026, the Chinese computing power market is experiencing a sharp divide: on one hand, leading enterprises have orders scheduled until 2028 and their net profits have skyrocketed by over 800%; on the other hand, a large number of small and medium-sized players in the west are struggling on the verge of losses with a utilization rate of less than 70%, and some intelligent computing centers have even become unfinished projects.
In the national strategic layout of the "Eastern Data, Western Computing" initiative, Wuhu is one of the two clusters in the Yangtze River Delta hub and the only national-level data center cluster in central China. It has a large-scale intelligent computing capacity and has gathered multiple super-large projects such as Huawei, Douyin, and the three major telecommunications operators, undertaking over 70% of Anhui Province's intelligent computing tasks.
As one of the core operators of Wuhu's computing power industry, Hu Rong, the chairman of Wuhu Big Data Construction Investment and Operation Co., Ltd., is driving the operation of a provincial-level computing power coordination and scheduling platform. They are not satisfied with being simple "cabinet landlords" but are trying to build a cross-regional computing power trading platform to enable efficient and market-oriented circulation of computing power, just like water and electricity.
Why does the local state-owned asset platform in Wuhu emphasize light-asset operation in a capital-intensive sector? When regions engage in a subsidy competition to attract computing power enterprises, how can we ensure that the money is spent effectively? Against the backdrop of restrictions on high-end chips, when will domestic computing power truly be able to step in?
She warned that the biggest bubbles in the current computing power market lie in "irrational premium due to supply-demand mismatch" and "financialization of heavy assets." Some enterprises even use GPUs (Graphics Processing Units) as collateral for high-leverage financing, forming a computing power real estate model similar to "buying real estate with a loan."
Economic Observer: What is the real temperature of the current computing power leasing market? Which sectors are truly booming, and which are just showing false prosperity?
Hu Rong: The truly booming sector in the current computing power market is the high-end intelligent computing field, especially the computing power required for large model training and inference. Enterprises have a large demand for it, but due to various factors, there is often a serious shortage of supply, leading to a sharp increase in the rental price of some graphics cards. Leading enterprises with strong financial resources have pre-arranged relevant computing power resources in advance. They have no shortage of lessees in the rental market and naturally have sufficient profit margins. However, reasonable planning should be made for the investment and construction layout of computing power resources to make the best use of things and avoid falling into the dilemma of supply-demand mismatch.
In the early days of the computing power leasing market, there were extensive and high profits. Whoever got the graphics cards first could rent them out at sky-high prices. This was a phased bonus under the extreme supply-demand mismatch. As more capital flows in and more data centers are put into operation, the window for earning excessive profits simply through resource price differences is indeed narrowing.
However, there are still great opportunities to "enter the market" in 2026, but the nature of the opportunities has changed. The computing power industry should move from "renting resources" to "selling services." The opportunities mainly lie in three aspects: First, build differentiated ecological capabilities, such as providing customized computing power solutions for specific industries; Second, seize the structural opportunity of the explosion of inference demand and provide in-depth operation capabilities such as refined scheduling, cost optimization, and on-demand services around inference scenarios; Third, keep up with the development trend of the domestic computing power industry, such as establishing the ability to operate Tokens (the smallest unit when an artificial intelligence model processes text). The period of high profits may have passed, but the period of value creation has just begun.
Economic Observer: Where are the biggest bubbles in the current computing power market? What is the most dangerous signal?
Hu Rong: The biggest bubbles lie in "irrational premium due to supply-demand mismatch" and "financialization of heavy assets." For example, the price of the B300 server soared to 7 million yuan at one point, and the rental price increased by 40%. The core driving force is not simply technological demand but a combination of multiple factors such as overseas export controls, blocked compliance channels, and speculation in the gray market. Some enterprises even use GPUs as collateral for high-leverage financing, forming a computing power real estate model similar to "buying real estate with a loan," resulting in the asset valuation being seriously divorced from the actual commercial return.
The most dangerous signals are "the risk of asset depreciation caused by the delay of technological iteration" and "the obstruction of downstream commercialization and monetization." Once the delivery of the new generation of chips is delayed due to problems such as the yield rate of advanced packaging, or the artificial intelligence (AI) applications cannot achieve large-scale profitability, the high computing power cost will directly crush downstream customers. At that time, the old GPUs hoarded at high prices will face a double blow of plummeting rental prices and exhaustion of liquidity, and the heavy assets blindly expanded in the early stage will easily become non-performing debts, triggering a systematic clearance of the industrial chain.
Economic Observer: What are the main bottlenecks you encounter when conducting cross-regional computing power scheduling?
Hu Rong: First, there is a hard constraint of network latency: the latency problem caused by long-distance transmission is still an insurmountable barrier for AI inference tasks with high real-time requirements; Second, the technical standards and interfaces are not unified: data centers in different clusters and from different manufacturers have their own ways in computing power measurement, billing methods, API (Application Programming Interface) interfaces, and scheduling protocols. When enterprises want to use the computing power of multiple regions at the same time, they often need to do a lot of adaptation and development, and the transaction cost remains high; Third, it is difficult to guarantee cross-regional services: after the cross-provincial use of computing power, there is currently a lack of unified rules on how to settle the fees, how to guarantee the service quality, and how to define the liability for faults.
Economic Observer: The computing power subsidy standards vary greatly among different cities, and some people specifically "seek subsidies." What do you think of this phenomenon?
Hu Rong: The Anhui Provincial Computing Power Coordination and Scheduling Platform assists the Provincial Department of Science and Technology in the accurate and efficient distribution of subsidy funds. Through in-depth penetration of business scenarios, the platform accurately matches the supply and demand sides of computing power, builds a risk control line from the source, and effectively eliminates the behavior of seeking and defrauding subsidies. In the past two years, more than 26 million yuan of provincial intelligent computing subsidies have been distributed in Anhui.
As for whether the subsidy competition is a healthy competition or a waste of resources, my judgment is: it depends on the result. A healthy subsidy should "use a small effort to achieve a great effect," effectively reducing the cost of enterprises using computing power, cultivating the industrial ecosystem, and driving the growth of real computing power consumption. However, if the subsidy is for a "show project" or "investment promotion figures," it is a typical waste of resources.
Economic Observer: How much has Wuhu Big Data Company invested in computing power infrastructure? What is the return period? Is it profitable? What is the relationship between you and private computing power service providers like Lianyun Century? What are the respective advantages and disadvantages of both sides?
Hu Rong: Our investment is mainly in platform construction. The company has independently invested in building the Anhui Provincial Computing Power Coordination and Scheduling Platform, and the construction of data center infrastructure is mainly invested by enterprises such as Huawei and the three major telecommunications operators. Our role is "platform + service" rather than "heavy asset holder," and this model itself reduces the scale pressure of direct investment.
In terms of the return period, the heavy-asset model of pure computing power leasing does face a long cycle and high capital pressure. However, we are taking a diversified path: achieving profitability through comprehensive revenues such as platform scheduling fees, data operation services, and the ecological benefits brought by the "computing power coupon" policy. What we pursue is not the financial return of a single project but to attract the agglomeration of the artificial intelligence industry by lowering the threshold of computing power and drive the value improvement of the entire digital economy industrial chain in Wuhu.
We are first of all ecological partners with private computing power service providers like Lianyun Century. After these data centers are built, the computing power resources will be connected to our scheduling platform, and we will help them "sell" computing power to achieve supply-demand matching. We are the "platform provider" and "service provider," and they are the "supplier" and "tenant."
The advantages of local state-owned asset platforms lie in government credit endorsement, policy resource integration ability, and data security and compliance guarantee. The advantages of private enterprises lie in their agile and flexible market-oriented mechanism, strong motivation for technological innovation, and fine cost control. The disadvantages of both sides are also corresponding: state-owned asset platforms are not as flexible and have less incentive mechanisms as private enterprises, while small and medium-sized private enterprises have natural disadvantages in obtaining scarce chips and large-scale financing. This model of "state-owned enterprises setting the stage and private enterprises performing" has initially proven feasible in Wuhu.
Economic Observer: Some people think that building computing power centers by local state-owned assets is "heavy assets, low returns, and high risks." How can we avoid the risk of "idle upon completion?"
Hu Rong: The key to the solution lies in how to operate. Our response strategy has three levels. First, adhere to "platformization" rather than "assetization." We do not blindly build a large number of data centers on our own but build a computing power scheduling platform to manage the existing computing power resources in society and carry out light-asset operation work.
Second, establish a dynamic monitoring and elastic adjustment mechanism. The Wuhu cluster implements full-process tracking and management of data center projects, and uses means such as price guidance and demand matching on the platform to flexibly adjust and quickly direct idle computing power to users with needs, avoiding "idle upon completion."
Third, develop diversified sources of income. We are not just engaged in computing power leasing but also form an income matrix through data operation and other sectors. If the demand for computing power suddenly cools down, we can also transform the computing power resources into the infrastructure for data operation and apply the computing power to fields and scenarios such as government services and smart cities. A rich variety of scenarios is an important manifestation of strong demand. With usage scenarios, the computing power resources can be fully revitalized.
Economic Observer: High-end GPUs are in short supply. What is the gap between domestic computing power and NVIDIA?
Hu Rong: There is indeed a gap between domestic computing power and NVIDIA, and the core lies in the software ecosystem. NVIDIA's CUDA (GPU General Computing Development Platform) ecosystem is mature, and the toolchain is complete. Developers can get started quickly and the migration cost is low. These are exactly the areas that domestic computing power needs to improve. However, this gap is narrowing. At present, mainstream domestic chips already have good competitiveness in inference scenarios, and there is more room for cost control. The gap is mainly concentrated in the single-card performance and cluster stability in ultra-large-scale training scenarios, but through system optimization and scheduling improvement, this gap is gradually narrowing.
Economic Observer: When will domestic computing power truly be able to "step in?" What can local state-owned asset platforms do?
Hu Rong: I think the initial results will be seen in three years, and it will be basically achievable in five years. Within three years, domestic computing power can "step in" in specific vertical industries and specific application scenarios, especially in inference scenarios. However, to fully replace NVIDIA's position in high-performance training scenarios, especially in the most "core" scenarios such as pre-training of trillion-level large models, at least five years of ecological construction cycle are needed.
Local state-owned asset platforms can do a lot of things. First, participate in the testing of domestic computing power capabilities, accumulate operation data, and provide optimization suggestions upwards; Second, promote the dual adaptation of "computing power + algorithm" to help AI enterprises complete the model migration from NVIDIA to domestic chips and lower the technical threshold; Third, cultivate the developer ecosystem, gather application scenarios through the computing power platform, and let more developers use domestic computing power in real business to form a positive cycle.
Economic Observer: What capabilities of enterprises does Wuhu value most in investment promotion? What signals are you most vigilant about?
Hu Rong: We value three things most: first, having real computing power demand, not just telling stories; second, having a technical team and continuous R & D ability, not just making quick money; third, having a clear business model and profit expectation, not relying on subsidies to survive. The three signals we are most vigilant about are: first, talking about government subsidies before having anything; second, only setting up a "shell company" in Wuhu while the core team and technology are all outside; third, having a main business completely unrelated to the digital economy and just trying to ride on the "computing power" concept.
The computing power industry is a long-distance race, not a 100-meter sprint. What we want to do is to cultivate the local computing power industrial ecosystem and make Wuhu a well-deserved "City of Intelligent Computing."
Economic Observer: What suggestions do you want to give to entrepreneurs or investors entering the computing power leasing sector?
Hu Rong: First, rely on computing power resources to create unique industry value. Many entrepreneurs ask "how to get NVIDIA cards" as soon as they start, simply understanding computing power leasing as "buying cards and renting them out." But the real competitiveness lies in how you make computing power create value. Is it to build a computing power scheduling platform, provide computing power solutions for vertical industries, or engage in Token sharing operation? Different positions correspond to completely different ability requirements.
Second, do a good job in operation and service. The computing power leasing industry is capital-intensive, with large upfront investments and a long return period. Without strong capital reserves and a clear contingency plan, do not easily enter the "heavy asset holding" model. You can consider doing "light asset" service operation and realize commercial value by scheduling existing computing power.
Third, attach great importance to compliance and security. Any problem in any link may make all the previous investments go to waste.
This article is from the WeChat official account "Economic Observer", author: Wang Yajie, published by 36Kr with authorization.