A 12 months or so in the past, Xiao Li was seeing floods of Nvidia chip offers on WeChat. An actual property contractor turned information middle undertaking supervisor, he had pivoted to AI infrastructure in 2023, drawn by the promise of China’s AI craze. 

At the moment, merchants in his circle bragged about securing shipments of high-performing Nvidia GPUs that have been topic to US export restrictions. Many have been smuggled by way of abroad channels to Shenzhen. On the top of the demand, a single Nvidia H100 chip, a form that’s important to coaching AI fashions, might promote for as much as 200,000 yuan ($28,000) on the black market. 

Now, his WeChat feed and trade group chats inform a unique story. Merchants are extra discreet of their dealings, and costs have come again right down to earth. In the meantime, two information middle initiatives Li is accustomed to are struggling to safe additional funding from buyers who anticipate poor returns, forcing undertaking results in dump surplus GPUs. “It looks like everyone seems to be promoting, however few are shopping for,” he says.

Simply months in the past, a growth in information middle building was at its top, fueled by each authorities and personal buyers. Nonetheless, many newly constructed amenities are actually sitting empty. Based on folks on the bottom who spoke to MIT Know-how Evaluation—together with contractors, an government at a GPU server firm, and undertaking managers—many of the firms operating these information facilities are struggling to remain afloat. The native Chinese language shops Jiazi Guangnian and 36Kr report that as much as 80% of China’s newly constructed computing assets stay unused.

Renting out GPUs to firms that want them for coaching AI fashions—the principle enterprise mannequin for the brand new wave of information facilities—was as soon as seen as a certain wager. However with the rise of DeepSeek and a sudden change within the economics round AI, the trade is faltering.

“The rising ache China’s AI trade goes by way of is essentially a results of inexperienced gamers—firms and native governments—leaping on the hype practice, constructing amenities that aren’t optimum for right now’s want,” says Jimmy Goodrich, senior advisor for expertise to the RAND Company. 

The upshot is that initiatives are failing, vitality is being wasted, and information facilities have turn into “distressed belongings” whose buyers are eager to unload them at below-market charges. The scenario might finally immediate authorities intervention, he says: “The Chinese language authorities is more likely to step in, take over, and hand them off to extra succesful operators.”

A chaotic constructing growth

When ChatGPT exploded onto the scene in late 2022, the response in China was swift. The central authorities designated AI infrastructure as a nationwide precedence, urging native governments to speed up the event of so-called good computing facilities—a time period coined to explain AI-focused information facilities.

In 2023 and 2024, over 500 new information middle initiatives have been introduced in every single place from Interior Mongolia to Guangdong, in response to KZ Consulting, a market analysis agency. Based on the China Communications Business Affiliation Information Middle Committee, a state-affiliated trade affiliation, at the very least 150 of the newly constructed information facilities have been completed and operating by the tip of 2024. State-owned enterprises, publicly traded corporations, and state-affiliated funds lined as much as spend money on them, hoping to place themselves as AI front-runners. Native governments closely promoted them within the hope they’d stimulate the economic system and set up their area as a key AI hub. 

Nonetheless, as these expensive building initiatives proceed, the Chinese language frenzy over massive language fashions is dropping momentum. In 2024 alone, over 144 firms registered with the Our on-line world Administration of China—the nation’s central web regulator—to develop their very own LLMs. But in response to the Financial Observer, a Chinese language publication, solely about 10% of these firms have been nonetheless actively investing in large-scale mannequin coaching by the tip of the 12 months.

China’s political system is very centralized, with native authorities officers sometimes shifting up the ranks by way of regional appointments. Consequently, many native leaders prioritize short-term financial initiatives that display fast outcomes—usually to achieve favor with higher-ups—reasonably than long-term improvement. Massive, high-profile infrastructure initiatives have lengthy been a software for native officers to spice up their political careers.

The post-pandemic financial downturn solely intensified this dynamic. With China’s actual property sector—as soon as the spine of native economies—slumping for the primary time in many years, officers scrambled to search out various development drivers. Within the meantime, the nation’s as soon as high-flying web trade was additionally coming into a interval of stagnation. On this vacuum, AI infrastructure grew to become the brand new stimulus of selection.

“AI felt like a shot of adrenaline,” says Li. “Some huge cash that used to stream into actual property is now going into AI information facilities.”

By 2023, main firms—a lot of them with little prior expertise in AI—started partnering with native governments to capitalize on the development. Some noticed AI infrastructure as a method to justify enterprise enlargement or increase inventory costs, says Fang Cunbao, a knowledge middle undertaking supervisor based mostly in Beijing. Amongst them have been firms like Lotus, an MSG producer, and Jinlun Know-how, a textile agency—hardly the names one would affiliate with cutting-edge AI expertise.

This gold-rush strategy meant that the push to construct AI information facilities was largely pushed from the highest down, usually with little regard for precise demand or technical feasibility, say Fang, Li, and a number of on-the-ground sources, who requested to talk anonymously for concern of political repercussions. Many initiatives have been led by executives and buyers with restricted experience in AI infrastructure, they are saying. Within the rush to maintain up, many have been constructed rapidly and fell in need of trade requirements. 

“Placing all these massive clusters of chips collectively is a really troublesome train, and there are only a few firms or people who know do it at scale,” says Goodrich. “That is all actually state-of-the-art laptop engineering. I’d be stunned if most of those smaller gamers know do it. Loads of the freshly constructed information facilities are shortly strung collectively and don’t provide the steadiness that an organization like DeepSeek would need.”

To make issues worse, undertaking leaders usually relied on middlemen and brokers—a few of whom exaggerated demand forecasts or manipulated procurement processes to pocket authorities subsidies, sources say. 

By the tip of 2024, the thrill that when surrounded China’s information middle growth was  curdling into disappointment. The reason being easy: GPU rental is not a very  profitable enterprise.

The DeepSeek reckoning

The enterprise mannequin of information facilities is in principle simple: They earn cash by renting out GPU clusters to firms that want computing capability for AI coaching. In actuality, nevertheless, securing purchasers is proving troublesome. Just a few high tech firms in China are actually drawing closely on computing energy to coach their AI fashions. Many smaller gamers have been giving up on pretraining their fashions or in any other case shifting their technique for the reason that rise of DeepSeek, which broke the web with R1, its open-source reasoning mannequin that matches the efficiency of ChatGPT o1 however was constructed at a fraction of its price. 

“DeepSeek is a second of reckoning for the Chinese language AI trade. The burning query shifted from ‘Who could make one of the best massive language mannequin?’ to ‘Who can use them higher?’” says Hancheng Cao, an assistant professor of knowledge methods at Emory College. 

The rise of reasoning fashions like DeepSeek’s R1 and OpenAI’s ChatGPT o1 and o3 has additionally modified what companies need from a knowledge middle. With this expertise, many of the computing wants come from conducting step-by-step logical deductions in response to customers’ queries, not from the method of coaching and creating the mannequin within the first place. This reasoning course of usually yields higher outcomes however takes considerably extra time. Consequently, {hardware} with low latency (the time it takes for information to cross from one level on a community to a different) is paramount. Information facilities have to be positioned close to main tech hubs to reduce transmission delays and guarantee entry to extremely expert operations and upkeep employees. 

This transformation means many information facilities in-built central, western, and rural China—the place electrical energy and land are cheaper—are dropping their attract to AI firms. In Zhengzhou, a metropolis in Li’s house province of Henan, a newly constructed information middle is even distributing free computing vouchers to native tech corporations however nonetheless struggles to draw purchasers. 

Moreover, quite a lot of the brand new information facilities which have sprung up in recent times have been optimized for pretraining workloads—massive, sustained computations run on large information units—reasonably than for inference, the method of operating skilled reasoning fashions to answer person inputs in actual time. Inference-friendly {hardware} differs from what’s historically used for large-scale AI coaching. 

GPUs like Nvidia H100 and A100 are designed for enormous information processing, prioritizing velocity and reminiscence capability. However as AI strikes towards real-time reasoning, the trade seeks chips which can be extra environment friendly, responsive, and cost-effective. Even a minor miscalculation in infrastructure wants can render a knowledge middle suboptimal for the duties purchasers require.

In these circumstances, the GPU rental value has dropped to an all-time low. A current report from the Chinese language media outlet Zhineng Yongxian mentioned that an Nvidia H100 server configured with eight GPUs now rents for 75,000 yuan per 30 days, down from highs of round 180,000. Some information facilities would reasonably go away their amenities sitting empty than run the chance of dropping much more cash as a result of they’re so expensive to run, says Fan: “The income from having a tiny a part of the information middle operating merely wouldn’t cowl the electrical energy and upkeep price.”

“It’s paradoxical—China faces the very best acquisition prices for Nvidia chips, but GPU leasing costs are terribly low,” Li says. There’s an oversupply of computational energy, particularly in central and west China, however on the similar time, there’s a scarcity of cutting-edge chips. 

Nonetheless, not all brokers have been seeking to earn cash from information facilities within the first place. As a substitute, many have been interested by gaming authorities advantages all alongside. Some operators exploit the sector for backed inexperienced electrical energy, acquiring permits to generate and promote energy, in response to Fang and a few Chinese language media stories. As a substitute of utilizing the vitality for AI workloads, they resell it again to the grid at a premium. In different instances, firms purchase land for information middle improvement to qualify for state-backed loans and credit, leaving amenities unused whereas nonetheless benefiting from state funding, in response to the native media outlet Jiazi Guangnian.

“In direction of the tip of 2024, no clear-headed contractor and dealer out there would nonetheless go into the enterprise anticipating direct profitability,” says Fang. “Everybody I met is leveraging the information middle deal for one thing else the federal government might provide.”

A mandatory evil

Regardless of the underutilization of information facilities, China’s central authorities remains to be throwing its weight behind a push for AI infrastructure. In early 2025, it convened an AI trade symposium, emphasizing the significance of self-reliance on this expertise. 

Main Chinese language tech firms are taking word, making investments aligning with this nationwide precedence. Alibaba Group introduced plans to speculate over $50 billion in cloud computing and AI {hardware} infrastructure over the subsequent three years, whereas ByteDance plans to speculate round $20 billion in GPUs and information facilities.

Within the meantime, firms within the US are doing likewise. Main tech corporations together with OpenAI, Softbank, and Oracle have teamed as much as decide to the Stargate initiative, which plans to speculate as much as $500 billion over the subsequent 4 years to construct superior information facilities and computing infrastructure. ​Given the AI competitors between the 2 international locations, specialists say that China is unlikely to cut back its efforts. “If generative AI goes to be the killer expertise, infrastructure goes to be the determinant of success,”  says Goodrich, the tech coverage advisor to RAND.

“The Chinese language central authorities will possible see [underused data centers] as a mandatory evil to develop an vital functionality, a rising ache of types. You might have the failed initiatives and distressed belongings, and the state will consolidate and clear it up. They see the tip, not the means,” Goodrich says.

Demand stays robust for Nvidia chips, and particularly the H20 chip, which was custom-designed for the Chinese language market. One trade supply, who requested to not be recognized beneath his firm coverage, confirmed that the H20, a lighter, quicker mannequin optimized for AI inference, is at the moment the preferred Nvidia chip, adopted by the H100, which continues to stream steadily into China although gross sales are formally restricted by US sanctions. Among the new demand is pushed by firms deploying their very own variations of DeepSeek’s open-source fashions.

For now, many information facilities in China sit in limbo—constructed for a future that has but to reach. Whether or not they may discover a second life stays unsure. For Fang Cunbao, DeepSeek’s success has turn into a second of reckoning, casting doubt on the belief that an limitless enlargement of AI infrastructure ensures progress.

That’s only a fantasy, he now realizes. At the beginning of this 12 months, Fang determined to stop the information middle trade altogether. “The market is just too chaotic. The early adopters profited, however now it’s simply folks chasing coverage loopholes,” he says. He’s determined to enter AI schooling subsequent. 

“What stands between now and a future the place AI is definitely in every single place,” he says, “will not be infrastructure anymore, however strong plans to deploy the expertise.”