In recent years, local governments in mainland China have been introducing policies to subsidize and support intelligent computing center projects, leading to an investment scale of trillions. However, data from various sources indicate that the average utilization rate of various types of computing centers is less than 30%. Industry insiders point out that the biggest bubble in the computing power market is the high premium and high leverage computing power real estate model that is formed due to the mismatch between supply and demand. The biggest risk lies in the devaluation risk caused by technological iteration and the hindered commercialization realization of downstream applications. According to public reports, many enterprise-level computing projects have already fallen through.
According to a report by “C114 Communication Network,” Zhang Jin, the General Manager of Tencent Cloud Operator Solutions, stated at the “World Mobile Communications Conference 2026” in Shanghai that the average GPU utilization rate of domestic intelligent computing centers is less than 30%, mainly due to low efficiency and low loading coefficients. The efficiency coefficient measures the actual effectiveness of GPU computing power over time, currently only averaging 0.6. The loading coefficient measures the utilization density of GPUs over time, with an average of only 0.5.
According to Tencent Cloud’s formula “Computing power productivity = nominal computing power × efficiency coefficient × loading factor,” due to efficiency decay and loading losses, the nominal computing power results in only a 30% utilization rate, leaving 70% idle.
Other sources also indicate that the utilization rate of computing power in Chinese intelligent computing centers is around 30%. As reported by the “Securities Times” on July 10, the average cabinet utilization rate of a large number of intelligent computing centers in China lingers between 20% to 30%, with some enterprise-level centers even falling below 10%. On one side, rental prices for computing power remain high due to chip shortages, while on the other side, expensive GPUs remain idle in data centers – demonstrating that owning computing power does not necessarily equate to productivity, which has become a warning sign hanging over the industry.
Reuters previously reported that data from four different sources indicated that the utilization rate of data centers in China is around 20% to 30%. It is important to note that traditional data centers encompass physical spaces for IT equipment like servers, storage, and networks, whereas intelligent computing centers provide powerful computing, data, and algorithm services for AI applications such as training large language models, image recognition, and language processing.
Fang Chao, the Executive Director of the Beijing Yizhuang Institute of Artificial Intelligence, told Caixin reporters that from a planning perspective, early construction efforts mainly focused on traditional data centers and did not fully anticipate the explosive demand for AI computing power, resulting in a structural mismatch. From a market mechanism perspective, the decentralized nature of computing power resources, lack of standardized procedures, and insufficient regional scheduling mechanisms have made it difficult to efficiently utilize existing computing power resources.
While the average utilization rate of intelligent computing centers in China stands at only 30%, the demand remains high in the high-end fields of large-scale model training and inference, leading to a mismatch in supply and demand that creates a significant premium the computing power market.
According to a report by “Economic Observer” on June 2, Hu Rong, the Chairman of Wuhan Big Data Construction Investment and Operation Co., Ltd., stated that the biggest bubble in the computing power market lies in the “irrational premium under supply and demand mismatch” and “intensification of asset financialization.” Some enterprises even use GPUs as collateral for high-leverage financing, creating a computing power real estate model akin to “speculating on houses with loans,” causing asset valuations to significantly deviate from actual business returns.
Hu Rong also pointed out that the biggest risks in the computing power market are the “devaluation risk caused by technological iteration” and the “obstruction of downstream commercialization realization.” If AI applications fail to achieve profitable scale, the high cost of computing power will directly burden downstream customers. At that point, GPUs accumulated at high prices will face a dual blow of plummeting rental prices and drying up liquidity, rendering the blind expansion of heavy asset-based businesses susceptible to turning into bad debts, triggering systematic clearing of the industry chain.
According to the issuance recommendation document for the initial public offering and listing on the Science and Technology Innovation Board of Moore’s Chongqing Intelligent Technology Co., Ltd. by China Securities Co., Ltd. in June 2025, GPU computing power has continued to strengthen, significantly shortening its update cycle. Many computing centers may not have recouped their costs before new-generation chips hit the market, leading to their direct “elimination” in economic terms.
As the foundation of AI infrastructure, intelligent computing centers are considered to be the new infrastructure of the artificial intelligence era, akin to water and electricity. Local governments have introduced policies to subsidize and support intelligent computing center projects.
In a report by “Southern Metropolis Daily” in February 2025, Lan Yun, the director of Yuhong Artificial Intelligence Research Institute, stated that there are a total of 196 intelligent computing centers of various types nationwide, including newly constructed, expanded, and approved centers. Of these, 135 intelligent computing centers that disclosed their exact investment amounts had an average investment of 2.857 billion yuan per center.
With 196 intelligent computing centers calculated to have an average investment of 2.857 billion yuan per center, the total amount of investment reaches 559.9 billion yuan. When considering planned constructions and centers that have been built but did not disclose information, the total investment in the industry approaches trillions of yuan.
According to “Southern Finance and Economics,” citing data from the National Development and Reform Commission of the Communist Party of China, as of the end of March this year, China had built an intelligent computing capacity of 1.882 million P (PFlops, representing tens of trillions of floating-point operations per second), which is equivalent to 2.5 times that of the same period last year and is expected to continue to grow rapidly. Experts estimate that the investment volume of the computing power infrastructure system could exceed trillions of yuan.
The construction frenzy of computing centers is heavily reliant on financial subsidies from local governments. According to public records, in November 2025, Hubei Province distributed a total of 50 million yuan in “computing power vouchers,” providing a 10% subsidy on the actual costs of enterprise computing power.
In February 2025, the Hunan provincial government introduced a policy that offers a 10% subsidy on the equipment and software construction costs for intelligent computing infrastructure projects with a total computing power of over 50 PFlops (tens of trillions of floating-point operations per second), with a maximum of 20 million yuan.
In June 2025, the city of Wuhu in Anhui Province specifically allocated 100 million yuan in fiscal funds for issuing “computing power vouchers,” providing a 25% fund support for smart computing service contract fees, with each recipient eligible for a maximum annual payment of 20 million yuan.
In 2023, Kunming, Yunnan, introduced ten measures to accelerate the development of the artificial intelligence industry, proposing to support the construction of intelligent computing centers and large-scale computing model service platforms, offering a 15% subsidy on the actual investments in platform software and hardware equipment, with a maximum subsidy of 50 million yuan.
He Yun (pseudonym), who works in the AI data center industry in China, told The Epoch Times that there is an oversupply of computing centers, with many built to cater to large model enterprises. However, due to government encouragement to establish computing centers, there has been an excess of such centers constructed.
Looking back at past examples, anything that the Chinese Communist Party has pushed forward with policies and developed with “national efforts” has never been spared from a fate of ending in a mess. China has taken a similar path in solar energy, new energy vehicles, and lithium batteries – rushing to expand production capacity, resulting in structural surplus, followed by rounds of integration and clearing, ultimately leaving only a few enterprises surviving, while many others are eliminated by the market, leading to more projects failing. Local governments’ investments have gone to waste, burdening taxpayers with heavy debts, inevitably spreading this burden to every taxpayer.
Against the backdrop of a large amount of idle computing power, how far are some intelligent computing center projects from failing? At the beginning of 2026, National Committee member Shi Ke cautioned in a proposal to the National Committee of the Chinese People’s Political Consultative Conference to guard against intelligent computing centers becoming “digital problematic properties.”
As reported by the “Securities Times” on June 23, established children’s clothing manufacturer Annai in 2024 briefly became a trending stock after acquiring a computing power company, but after a four-month acquisition process, the deal fell through and they received fines.
Lotus Holding, primarily engaged in seasoning products, disclosed in November 2024 its purchase of 330 GPU servers but only delivered a few units.
Leisure and entertainment brand operator Oya Co. Ltd. initially planned to acquire 128 GPU servers in August 2024 but stopped purchases after delivering only 8 units.
Even after crossing the computing power threshold, it is not a smooth path. In January 2024, Qunxing Toys, without talent or intent contracts, expanded into computing power leasing but generated only 53 million yuan in revenue from intelligent computing business in 2025, with a gross loss of 15 million yuan.
Huafu Fashion’s Akshay Wakka Cluster project, which started two to three years ago, has progressed slowly, prompting investors to question in April 2026 whether it is headed for a miscarriage.
In September 2023, Sichuan Hui-Su Intelligent Computing claimed to have purchased the first batch of 200 A100 servers, with a total computing power of 6.2 EFLOPS (hundreds of billions of floating-point operations per second), aiming to make a profit of 110 million yuan within three years. However, the actual results were: a net profit of 2.51 million yuan in 2023, edging to 15.065 million yuan in 2024, and a net profit of only 3.03 million yuan in 2025.
“Xinhua Finance and Economics” compiled the performance of 99 A-share listed companies in the computing power leasing concept: in the third quarter of 2025, 59 companies saw a year-on-year decline in gross profit margin, accounting for 56.6%. More than half of the listed companies in the computing power concept are becoming less profitable as they expand their businesses.