Ghost Infrastructure
How China raced to build data centers—and why so many are now left underutilized or idle.
Western coverage tends to frame China as an “AI boogeyman,” highlighting its breakthroughs (e.g., DeepSeek) and geopolitical ambitions (e.g., the “AI Plus” initiative) as a justification to plow ahead with AI advancement at all costs. But seldom is the material basis that underpins those advances closely examined.
That’s what I’ll do in this post. A closer look at the recent history of China’s data center buildout reveals a far more complex story. Behind the dazzling progress of its frontier AI models is an infrastructure story rife with inefficiency—hundreds of facilities built in a frenzy, now sitting idle or chronically underused. Much like the ghost cities that have long plagued the country’s real-estate sector, China’s zealous data center buildout have resulted in the proliferation of what I call ghost infrastructure.
Data center boom years
China’s data center boom began well before the ChatGPT moment. In 2020, Beijing announced a “new infrastructure” initiative that focused on seven key infrastructures to drive the next phase of the country’s modernization—among which were data centers. The announcement triggered a wave of investment and hundreds of firms rushed into the sector.
The ensuing build out was driven more by political enthusiasm than by market logic. Local governments competed fiercely to attract firms to build, offering subsidies, discounted land, and other kinds of financing. Data centers were erected across the country almost overnight. This enthusiasm was further intensified by another policy directive issued in early 2022 called East Data, West Computing, which encouraged the construction of more data centers in western provinces where land and energy are abundant.
The release of ChatGPT—and the realization that AI progress would hinge on ever-larger models demanding exponentially more compute—lit an even greater fire under the ongoing building spree. Dozens of startups rushed to develop their own LLMs, each needed vast computing resources to train them. Local governments and firms once again raced to expand capacity, retrofit older facilities, and position themselves as the backbone of China’s AI economy.
However, the facilities required to train today’s frontier AI models are far more complex than the commodity data centers of the past. They demand an enormous and steady power supply, advanced cooling systems, and round-the-clock operational stability—and, above all, access to vast clusters of GPUs capable of training large-scale models. In the West, such AI-oriented facilities are often called “neo-clouds”; in China, they’re known as zhi suan (智算), or “smart compute” centers. Many of the firms that had entered the market in the initial 2020-rush could make do essentially as real-estate ventures that focused on easy financing, sourced commodity equipment, and built cheaply at scale. But this new generation of AI-capable data centers demanded a level of technical sophistication that most players didn’t have.
This of course didn’t stop companies from building. In the years that followed, hundreds of new projects were announced as state-owned enterprises, publicly listed firms, and government-backed funds all sought a piece of the AI infrastructure buildout. Many started construction long before securing any reliable sources of demand. An MIT Tech Review article hilariously pointed out that “among them were companies like Lotus, an MSG manufacturer, and Jinlun Technology, a textile firm,” hardly the type of firm with relevant technical competencies.
Without market discipline, however, this reckless rush started to unravel. By 2024, hundreds of new AI-focused centers had been built, yet many fell far short of the technical and operational standards required to train modern AI models. Meanwhile, China’s AI model ecosystem began to consolidate. Dozens of smaller LLM startups folded as the market coalesced around a handful of dominant players—the major internet giants among them—leaving much of the newly built facilities without customers.
These more sophisticated users also had far higher technical standards, demanding facilities capable of 24/7 uninterrupted power and flawless operational reliability. This meant that inexperienced players—many of whom had treated data centers as real-estate construction projects rather than highly technical systems—were in no position meet them. And when DeepSeek’s R1 model was released in early 2025, showing that cutting models could be trained on a fraction of the computing capacity, demand contracted even further. Soon, it became clear that the hundreds of facilities built in the preceding years were heavily underutilized, with many operating well below capacity, some running at just 10%.
China’s industrial model
There are two underlying dynamics worth pointing out in the story above. First is the gold rush. Once Beijing designated data centers as a national priority, the policy itself became a signal of future demand—prompting both state and private actors to pile in, often well before any actual demand had materialized. This leads to the second related dynamic: the rapid influx of new players—many with little relevent competency beyond access to financing, sometimes from local governments—makes it difficult for more serious players to pursue longer-term, quality-oriented investments. These speculative entrants drive prices down, forcing capable firms to compete on cost rather than capability, and leaving them unable to reinvest in R&D, technology upgrades, or other improvements to the production stack. The result is overcapacity at the low end, underinvestment at the high end, and an industry trapped in a race to the bottom.
This of course is not unique to data centers. In fact, to make sense of China’s current data center crisis, we need to understand it within the country’s broader industrial model, which goes something like this: A new national priority is announced—or a promising technology emerges—and firms rush in to stake their claim. Local governments, eager to meet central targets and stimulate growth, each back their own “champion,” showering them with subsidies, tax breaks, and cheap financing. Production ramps up rapidly, competition intensifies, and prices spiral downward, creating a deflationary race to the bottom. Yet rather than allowing weaker firms to fail, local officials continue to prop them up, unwilling to see their homegrown projects collapse. The result is a proliferation of “zombie” firms that survive not on their commercial strength, but on policy support.
China’s electric vehicle (EV) manufacturers offers the clearest illustration of this dynamic, and is perhaps the most talked about case in recent years as they threaten incumbent automakers around the world. China now produces some of the world’s most competitively priced EVs and the top firms have started to export them across the world. Yet competition is so cutthroat that few automakers are profitable. Many persist only through a lifeline of local subsidies, as officials refuse to let their regional champions die. The industry is thus trapped in chronic overcapacity, with too many firms vying for an already saturated domestic market. (I’ve written more about this dynamic in the EV sector here.)
This pattern of hyper-competition has become so severe that it has drawn concern from the very top. Earlier this year, President Xi Jinping publicly warned automakers to exercise “self-discipline” to curb involution—a term borrowed rom Chinese internet culture that is now used to describe the self-defeating cycle of competition and overproduction. Indeed, this combination of industrial overcapacity and price wars has long plagued China’s economy, from steel and cement to solar panels and EVs.
And as the story of the data center boom shows, the same logic has also taken hold in the AI economy. The rush to build data centers has replicated the very pattern that once defined China’s manufacturing ascent: rapid expansion without market discipline, leading to impressive capacity on paper but widespread inefficiency in practice. Yet unlike manufacturing that can be easily exported, its much harder for domestic data center firms to sell excess capacity abroad.
Fighting involution
The AI data center boom hasn’t gone bust for everyone. Over the past few weeks, I’ve started to take a closer look at the data center firms that have managed to survive, and have noticed some important patterns about what distinguishes them from the many that didn’t.
First, however, it’s worth understanding briefly where AI compute demand in China has shifted to today. Model development has consolidated around a handful of major internet companies—Alibaba, Tencent, ByteDance, and Baidu—and a few well-capitalized AI startups such as DeepSeek and Moonshot. These firms, flush with capital and wary of security and reliability risks, increasingly build and operate their own data centers. As a result, there is relatively little demand for independent, third-party providers of AI-grade facilities. Most other companies that need compute lease capacity from these third-party players; these third-party players also supply overflow capacity for the top firms when they exceed their in-house resources.
In this environment, success hinges on a very different logic than the kind that characterized the earlier data center boom years. Data center firms have to focus not on quantity but quality—serving large, stable (mostly big tech) clients with real compute demand and focusing relentlessly on operational efficiency. Companies like Chindata Group’s success rests on cultivating deep relationships with major customers and maintaining an exacting operational model that guarantees uptime, stability, and energy efficiency.
This shift has implications for the industrial structure of the data center industry. Many of the failed entrants treated data centers as real-estate projects that can be cheaply replicated anywhere. The more successful firms, by contrast, know that the industry is fundamentally an engineering business—one that demands capabilities in efficient energy use, cooling, networking, and operations. To achieve this, firms like BCI Group have pursued tightly integrated supply chains, investing in in-house energy management, bare metal servers, and other IT equipment. Even the internet giants are increasingly vertically integrating down the stack and investing in their own capabilities in data center construction.
Successful firms have also benefited from long-term policy stability. The founder of BCI Group described choosing to concentrate operations in Datong, a city with abundant and diversified energy sources, a reliable regulatory environment, and proximity to Beijing. But most importantly, rather than chasing short-term subsidies, he emphasized the importance of consistent, long-horizon industrial policy:
Achieving a vertically integrated model requires certain conditions at the local level: long-term policy stability and resistance to excessive commercialization; not just access to electricity, but also a complete and sustainable set of resources—energy, surface water, land, and labor. (translated with AI)
Data centers and the compute layer on which the AI race is being waged represent just one layer of a much larger technology stack. And already, the biggest players are looking to take control over this layer to enable tighter integration across their value chain. But the rise and fall of China’s data center boom, and the few success cases illustrated above, offers a lesson that can be carried forward to other domains. It is that success doesn’t always come from speed and scale and capturing short-term profits, especially when demand is untested. Rather, it depends on a much longer-term vision where firms embed themselves in stable ecosystems, cultivate deep customer relationships with actual demand, and invest in the technical and organizational depth that can turn commodity infrastructure into value-added capabilities.