Riding the AI ​​boom, China built data centers — now hundreds of them are useless

During the “neural network rush,” China poured billions into AI infrastructure, declaring the projects a national priority. But many of the newly built facilities now sit empty, while others struggle to stay afloat. DeepSeek’s emergence has made the lucrative business of renting out GPUs for AI training unprofitable. The economics of the industry have changed dramatically in a short period of time.

Image source: unsplash.com

When ChatGPT kicked off the AI ​​boom in late 2022, the Chinese government made AI infrastructure a national priority, calling for accelerating the development of so-called “intelligent computing centers” — a term coined to describe AI-focused data centers.

According to analytics firm KZ Consulting, more than 500 new data centers were laid down in China in 2023 and 2024, with at least 150 of them operational by early 2025. State-owned enterprises, public companies, and government-affiliated funds competed to invest in them, hoping to take the lead in the AI ​​race. Local governments hoped that the new data centers would boost the economy and make the region a key AI hub.

However, many of the newly built facilities are now empty, and most of the companies operating these data centers are struggling to stay afloat. By some estimates, up to 80% of newly commissioned AI infrastructure facilities are unused. Renting out GPUs for AI training was seen as the primary business model for the new wave of data centers and was considered a sure bet. But with the advent of reasoning models like DeepSeek R1, which lowered the bar for training AI models by orders of magnitude, and the sudden change in the economics of AI, the industry is reeling.

The rise of reasoning models such as DeepSeek R1 and OpenAI ChatGPT o1 and o3 has changed the requirements for data centers. With these models, most of the computing power is spent on step-by-step inference in response to user queries, rather than training the model. The reasoning process produces high-quality results, but it takes more time. As a result, low latency is becoming a determining factor, which means data centers need to be located near major tech hubs. Many data centers built in remote areas of China, where electricity and land are cheaper, have become unattractive to AI companies.

Additionally, many of the new data centers that have emerged in recent years have been optimized for training workloads—large, time-consuming computations performed on huge data sets—rather than for responding to user input in real time.

Demand has shifted to low-latency hardware for inference, making remote data centers less attractive. DeepSeek has shifted its approach from “who can build the best big language model?” to “who can use that model best?”

In 2024 alone, more than 144 companies registered with the Cyberspace Administration of China, the country’s central internet regulator, to develop their own large language models. However, only about 10% of these companies were still actively investing in these projects by the end of the year, according to the Chinese publication Economic Observer.

«The growing pains facing China’s AI industry are largely the result of inexperienced players — corporations and local governments — jumping on the hype train by building facilities that are not optimal for today’s needs,” said Jimmy Goodrich, senior technology adviser at the RAND Corporation.

As a result, projects fail, energy is wasted, and data centers become “distressed assets” that are put up for sale at below-market prices. The situation could eventually prompt government intervention, which Goodrich says would “hand them over to more capable operators.”

It is also necessary to note the abuses that accompany the construction of new data centers. Not all market participants sought to make money on data centers; many use them to obtain subsidized green energy, which they then resell to the grid at a premium. Sometimes companies acquire land for the development of data centers in order to receive government loans and credits, without even planning to use the facility for its intended purpose.

Analysts identify the following objective reasons for market oversaturation:

• Excess initial financing amid a declining real estate market and stagnant internet industry;
• Decline in the GPU rental market, significant price reductions;
• Disappointment of investors after completion of construction, difficulties in attracting further financing;
• The inadequacy of many built data centers to meet current needs;
• Local authorities encourage short-term projects in order to enhance their political status;
• Limited management experience, haste and failure to comply with industry standards;
• Inflated forecasts and data manipulation by intermediaries and brokers to obtain government subsidies.

Under these circumstances, the rental price of GPUs has fallen to an all-time low. According to Chinese publication Zhineng Yongxian, a server with eight Nvidia H100 GPUs is now rented for 75,000 yuan ($10,200) per month, compared to 180,000 yuan ($24,600) previously. Some data centers are choosing to shut down their facilities because revenues do not cover electricity and maintenance costs.

Ironically, China faces the highest costs of acquiring Nvidia chips, but GPU rental prices are unusually low. The country has a glut of computing power, particularly in central and western China, but also a shortage of advanced chips. Some of the new demand is driven by companies rolling out their own versions of DeepSeek models. Nvidia’s H20, optimized for AI inference, is in the highest demand, followed by the H100, which continues to steadily arrive in China despite U.S. sanctions.

Despite the downtime of many of the data centers it has built, the Chinese government continues to develop AI infrastructure. Major Chinese tech companies continue to invest in this area, which has been declared a national priority. Alibaba Group plans to invest more than $50 billion in cloud computing and AI infrastructure over the next three years, and ByteDance is planning to spend about $20 billion on graphics processors and data centers.

It’s worth noting that US tech companies are doing the same, taking part in the $500 billion Stargate program to build advanced data centers and computing infrastructure over the next four years. Given the competition in AI, China is unlikely to scale back its efforts.

«“Infrastructure is going to be the defining factor of success,” Goodrich says. “The Chinese government is likely to view [idle data centers] as a necessary evil to develop an important opportunity, a kind of growing pain. You have failed projects and distressed assets, and the government will consolidate and clean them up. They see the end, not the means.”