China's AI infrastructure boom was built on a simple expectation: more artificial intelligence would require more computing power. But many of the new facilities built to serve that demand are now struggling to find customers.
The result is a market where expensive GPU clusters can sit idle, investors are pulling back, and project managers are questioning whether the rush to build actually matched what AI companies now need.
A Buildout Fueled By The AI Rush
After ChatGPT emerged in late 2022, China moved quickly to make AI infrastructure a national priority. Local governments were encouraged to accelerate the development of smart computing centers, a term used for AI-focused data centers.
That push triggered a wave of construction. In 2023 and 2024, over 500 new data center projects were announced across regions including Inner Mongolia and Guangdong, according to KZ Consulting. At least 150 of the newly built data centers were finished and running by the end of 2024, according to the China Communications Industry Association Data Center Committee.
State-owned enterprises, publicly traded companies and state-affiliated funds invested in the sector. Local governments promoted projects as a way to stimulate their economies and position their regions as AI hubs.
But the buildout was not always driven by proven demand. Sources described a market crowded with companies and investors that had limited experience in AI infrastructure. Some projects were assembled quickly, and some fell short of the reliability and technical quality that major AI customers would require.
“The growing pain China’s AI industry is going through is largely a result of inexperienced players—corporations and local governments—jumping on the hype train, building facilities that aren’t optimal for today’s need,” says Jimmy Goodrich, senior advisor for technology to the RAND Corporation.
The GPU Rental Bet Is Weakening
The business plan behind many of these facilities was direct: rent out GPU clusters to companies training AI models. During the peak of demand, access to high-end Nvidia chips became a major part of the market.
Xiao Li, a real estate contractor who moved into AI infrastructure in 2023, saw Nvidia chip deals flood WeChat. Traders discussed high-performing Nvidia GPUs that were subject to US export restrictions, with many smuggled through overseas channels to Shenzhen. At the height of demand, a single Nvidia H100 chip could sell for up to 200,000 yuan ($28,000) on the black market.
That environment has changed. Li now sees traders acting more quietly, while prices have fallen. Two data center projects he knows are struggling to raise more funding, and project leads have been trying to sell surplus GPUs.
“It seems like everyone is selling, but few are buying,” he says.
Reports from local Chinese outlets Jiazi Guangnian and 36Kr say up to 80% of China's newly built computing resources remain unused. People involved in the sector, including contractors, a GPU server company executive and project managers, also described companies running data centers as under pressure.
GPU leasing prices show the stress. A report from Zhineng Yongxian said an Nvidia H100 server configured with eight GPUs now rents for 75,000 yuan per month, down from highs of around 180,000. Some operators would rather leave capacity unused than run part of a facility at a loss because electricity and maintenance costs remain high.
DeepSeek Changed What Customers Need
DeepSeek has become a turning point for the market. Its open-source reasoning model R1 matched the performance of ChatGPT o1 while being built at a fraction of its cost, according to the source article.
That shifted the question for many companies. Instead of focusing mainly on who can build the best large language model, the market is paying more attention to who can use models effectively.
“DeepSeek is a moment of reckoning for the Chinese AI industry. The burning question shifted from ‘Who can make the best large language model?’ to ‘Who can use them better?’” says Hancheng Cao, an assistant professor of information systems at Emory University.
This matters because reasoning models change the type of computing demand. Much of the need comes from inference: running trained models in response to user inputs in real time. That is different from pretraining, where large sustained computations run across massive data sets.
For inference, low latency is critical. Data must move quickly, which makes proximity to major tech hubs more important. Facilities also need access to skilled operations and maintenance staff.
That weakens the appeal of data centers built in central, western and rural China, where land and electricity may be cheaper but distance from major AI customers can become a disadvantage. In Zhengzhou, a newly built data center has distributed free computing vouchers to local tech firms but still struggles to attract clients.
A Mismatch In Hardware, Location And Incentives
Many new data centers were designed around pretraining workloads, not inference-heavy use. That creates a technical mismatch. GPUs such as Nvidia H100 and A100 are built for massive data processing, with an emphasis on speed and memory capacity. As AI shifts toward real-time reasoning, customers increasingly want hardware that is efficient, responsive and cost-effective.
Even small errors in planning can make a data center less attractive for the workloads companies actually need. In a market with falling rental prices, that can quickly damage the financial case for a project.
The sector also attracted middlemen and brokers. Sources said some exaggerated demand forecasts or manipulated procurement to capture government subsidies. Fang Cunbao, a data center project manager based in Beijing, said some operators pursued benefits linked to subsidized green electricity, permits, land, loans and credits rather than direct data center profitability.
“Towards the end of 2024, no clear-headed contractor and broker in the market would still go into the business expecting direct profitability,” says Fang. “Everyone I met is leveraging the data center deal for something else the government could offer.”
Why The Buildout May Continue Anyway
Despite the underused capacity, China's central government is still backing AI infrastructure. In early 2025, it convened an AI industry symposium and emphasized self-reliance in the technology.
Major companies are responding to that priority. Alibaba Group announced plans to invest over $50 billion in cloud computing and AI hardware infrastructure over the next three years. ByteDance plans to invest around $20 billion in GPUs and data centers.
Goodrich expects China may not simply walk away from failed projects. He said the government is likely to step in, take over and hand distressed assets to more capable operators.
Demand for Nvidia chips remains strong, especially the H20 chip, which was custom-designed for the Chinese market. One industry source said the H20, a lighter and faster model optimized for AI inference, is currently the most popular Nvidia chip, followed by the H100.
The central question is no longer whether China has built AI infrastructure. It has. The harder question is whether those facilities are in the right places, equipped for the right workloads and connected to real deployment plans.
“What stands between now and a future where AI is actually everywhere,” Fang says, “is not infrastructure anymore, but solid plans to deploy the technology.”