The data center industry in China has seen better days. According to domestic media outlets, a notable portion of the country’s recently built data centers remain unused. The National Development and Reform Commission even set a goal of getting at least 60% of the country’s data centers online by the end of the year.

While the industry briefly thrived alongside the AI boom that followed ChatGPT’s 2022 launch, dramatic changes in the overall AI sector have left companies looking to sell off assets that have become a drag on their finances. That raises the question: How did a once flourishing industry founder so quickly? And what does the future hold for these now-silent facilities?

The rise in data center projects began in early March 2023, after the release of GPT-4, but really took off a month later as companies raced to stockpile as many AI chips as possible — and build the data centers needed to house them. At the height of this boom, from late 2023 to early 2024, over 500 new data center projects were announced nationwide.

Contributing to the boom was the government initiative known as “Eastern Data Western Computing.” Announced by the National Development and Reform Commission (NDRC) in 2021, the project aims to strengthen China’s tech industry and support its western provinces by turning them into hubs for data storage and processing.

The plan calls for creating 10 cloud computing hubs across eight regions. With the support of local governments, private companies such as Huawei and state-owned enterprises like China Mobile rushed in, with the NDRC estimating the overall cost of the plan at as much as 400 billion yuan ($56 billion). This massive outlay also drew in speculators, who sought to cash in by investing in and building their own data centers.

Then came DeepSeek.

When the Hangzhou-based hedge fund High-Flyer unveiled its DeepSeek large language model in late 2023, it was instantly hailed as a milestone for the Chinese AI industry — proof that the country could compete and even surpass the West despite restrictions on its access to U.S. tech. That success came with a price, however.

DeepSeek R1, an open-source model, was able to match ChatGPT o1 performance at a fraction of its cost in computing. That meant it needed less infrastructure to function, including data centers. Previously, the rule of thumb in the industry was the more data centers, the better the model. Now the focus is on smaller data centers positioned closer to tech hubs, ensuring faster transmission time.

DeepSeek’s success also caused other tech companies to pull back on their own model development. In 2024, 144 companies registered with the Cyberspace Administration of China to develop LLMs; currently, only 10% of them are still actively investing in large-scale model training, according to the Economic Observer. Instead, the majority of Chinese AI firms are focusing their efforts on optimizing currently available models — typically DeepSeek — for their own uses.

Ultimately, what this means is there is currently not enough demand for the computing power provided by China’s new data centers.

Nevertheless, despite the gloom surrounding the private investors and local governments that funded the construction of these dormant data centers, there remains potential for their future down the line. They could remain useless, a drag on the economy until they can be sold off at a heavy loss. Or they could be the seeds of something more, like how the dark fiber of the 2000s bloomed into the modern internet.

During the internet boom of the late 1990s, the United States invested heavily in building up a fiber network system to better facilitate faster internet speeds and stronger connections. Companies spent hundreds of millions of dollars on this project, only to have their plans derailed by the collapse of the dot-com bubble. For a long time, massive amounts of these already built “dark fiber” cables were left unused — as much as 85 to 95% by some estimates.

But those same cables later became the foundation for the next internet revolution, as the availability of cheap internet infrastructure helped companies like Google, Facebook, and Netflix grow rapidly.

Dark fiber is not the only collapsed technological bubble that later laid the groundwork for economic expansion. In the 1840s, Britain went through a similar speculative bubble with railways, leading to a massive rail network that initially seemed useless but eventually helped drive the Industrial Revolution.

The “dark fiber” potential of China’s data centers is real. Although optimized for pre-DeepSeek AI, they can be retrofitted to offer new services. One of the goals of the Eastern Data Western Computing initiative was to promote the digital transformation of less-developed western regions. Now, with these data centers in place, nascent startups in these regions can access high volumes of computing power at a low price. There are already signs of this with a newly built data center in Zhengzhou, capital of the central Henan province, reportedly distributing free computing vouchers to local tech firms. As tech companies and investors look beyond AI, idle data centers could fuel new opportunities from cloud computing to green technology, manufacturing to AAA gaming. The potential is there, but only if the data centers are kept around.

There is still a possibility China’s data centers become another example of overcapacity — useless pieces of infrastructure for an AI industry that has advanced beyond them. However, past experiences of speculative tech booms show that this is not always the case. While speculators are most likely in trouble in the short term, the tech industry itself could be on the verge of another leap forward. That is, if local governments and investors are willing to let these expensive pieces of infrastructure stay idle for a little while.

Editor: Cai Yiwen; portrait artist: Zhou Zhen.