TL; Dr: In view of the explosive beginning of the chip in the late 2022, the AI industry of China experienced enthusiasm and investment. However, this initial enthusiasm has given way to a frightening reality as the country has stated to the undertiliated data centers and shifting market dynamics with oversuply.
Jio Lee, a former real estate contractor, was published for the AI Infrastructure in 2023, has seen this change for the first time through demand for ups and downs for NVidia GPU. A year ago, traders in their networks claimed to receive high performance NVidia GPU despite US export restrictions. Many of these chips illegally got funnel in Shenzhen through international channels. On the peak of the market, an Nvidia H100 – AI is important to train models – can bring more as 200,000 yuan ($ 28,000) on the black market.
Today, Li noticed that traders have become more prudent and GPU prices have become stable. Additionally, two data centers projects they are familiar with are struggling to attract further investment as the backers estimate weak returns. This financial stress has forced the project leaders to unload additional GPU. MIT Technology Review said, “Everyone is selling, but there are no buyers.”
In short, leasing the GPU for businesses for AI model training – a main strategy for the latest generation data centers – once considered a guaranteed success. However, the emergence of Deepsek and shifting economic factors in the AI region has placed the country’s data center industry on unstable ground.
Data centers across China from Inner Mongolia to Guangdong were given fuel by rapid construction, government instructions and a combination of private investment. More than 500 new projects were announced in 2023 and 2024, with at least 150 by the end of 2024. However, the boom of this building has created a state of contradiction: an abundance of computational power, especially in Central and Western China, coupled with lack of chips, which meets current requirements for current requirements and regulatory realities.
The rise of Deepsek, a company that developed an open-sour Reasoning model, matches the performance of CHATGPT, but has further interrupted the market, at a fraction of the cost. Hancheng Cao, an assistant professor at Emori University, said that this success has focused on practical applications from model development. “Who can make the burning question ‘the biggest language model?” ‘Who can use them better?’ ,
This change has highlighted the boundaries of many haste data centers. Several features adapted to large-scale AI training are ill for the low-oppression requirements of the estimated tasks required for real-time logic models. As a result, data centers in remote areas with cheap electricity and land are losing their appeal to AI companies.
Oversupply of computational power has caused a dramatic fall in GPU fare prices. An Nvidia H100 server with eight GPU is now rented for 75,000 yuan per month (about $ 10,345) below the last high level of around 180,000 yuan ($ 25,141). Some data center operators chose to neutralize their features instead of leaving their features.
Jimmy Gudich, Senior Technology Advisor to Rand Corporation, credited this prediction for inexperienced players who jumped on AI Bandwagan. “Rising pain is undergoing China’s AI industry, it is a large extent of inexperienced players – corporations and local governments – jumping on promotional train, construction facilities that are not optimal to today’s needs,” they explain.
China’s political system has played an important role in the data center boom with emphasizing short -term economic projects for career advancement. Local authorities were demanding to promote their political career and stimulate the economy in front of a post-pandemic recession, turning into AI infrastructure as a new development driver.
This top-down approach often disregarded real demand or technical feasibility. Many projects were led by officials and investors with limited expertise in the AI infrastructure, resulting in hasty facilities that decreased from industry standards.
The rise of logic models such as R1 and OpenAI’s chatgate of Deepsek has transferred computing requirements from large scale training to real -time estimates. This change requires hardware with low delay, which is often located near the major tech hub, to reduce the transmission delay and to ensure access to skilled employees.
As a result, several data centers manufactured in Middle, Western and Rural China are struggling to attract customers. Some, like a newly created feature in Zengjhou, even distribute free computing vouchers to local technical firms, but still struggle to find users.
Despite the challenges, the Central Government of China prioritizes the development of AI infrastructure. In early 2025, it called an AI industry seminar, emphasizing the importance of self -sufficiency in this technique.
Prominent technical companies such as Alibaba and Bidens have announced significant investment in cloud computing and AI hardware infrastructure.
Gudich suggests that the Chinese government sees the current situation as an essential growing pain. “The Chinese Central Government will probably see as an essential evil (under data centers) to develop an important ability … they see the end, not the instrument,” they say.
As the industry develops, the demand for Nvidia chips remains strong, especially the H20 model designed for the Chinese market. However, for many people in the region, such as the data center project manager Fong Qunbao, the current status of the market has inspired the revaluation.
At the beginning of the year, Fong completely abandoned the data center industry. “The market is very chaotic. Early adopters made profits, but now it is only chasing policy flaws to the people,” they explain. He is now focusing his focus on AI education.
