China has spent too much money in AI and disaster is not far away

While the United States invests $ 500 billion in a data center for AI, China has launched an overly ambitious project. Between 2023 and 2024, public and private colossal investments were made to anticipate the growth of demand for the rental of graphic processors (GPU).

IA data centers unused in China

But the demand has finally dropped and many operators find themselves in a precarious financial situation. Enthusiasm was powered by the media excitement. The Chinese government has done everything to position the country as the world leader in AI and encouraged local officials to accelerate the construction of data centers.

According to state sources, more than 500 projects were announced nationally, and at least 150 were completed at the end of 2024. However, according to information from MIT Technology Review, up to 80 % of this new calculation capacity would remain inactive.

According to MIT Technology Review, geographic location is a major obstacle. The installations in the central and western regions of China are less expensive in electricity but face difficulties in meeting the requirements of latency. In cities like Zhengzhou, operators would distribute free computing vouchers to attract users.

In regions, developers sell their GPUs, for lack of long -term customers. Xiao Li, data center project manager interviewed by MIT Technology Review, weChat groups formerly very active for the transaction of Nvidia fleas are now inactive. “Everyone seems to sell, but few buy”he observed.

In the event of a global flood due to accidental capacity, this could cause significant difficulties for data centers developers and saturate an already weakened sector with even lower prices.
The emergence of Deepseek is among the main causes of down demand. The company upset the global technological economy at its launch in January 2025. Its open source model, R1, offers comparable performance to Chatgpt-O1 but at a lower cost.

Market interest has therefore been redirected in the training of models towards inference, that is to say in real time of AI models, which requires a different infrastructure. Many data centers built during the effervescence period had been designed for large -scale training, not to meet the requirements of low reasoning in real time.

Despite this catastrophic situation, China would continue its commitment. The authorities organized a symposium on the AI ​​in early 2025, and companies like Alibaba and Bytedance, owner of Tiktokhave announced significant investments.

But for investors in these data centers, expectations have collapsed. The infrastructure is there, but demand has not manifested itself.