TL;DR: Within the wake of ChatGPT’s explosive debut in late 2022, China’s AI trade skilled a surge of pleasure and funding. Nevertheless, this preliminary fervor has given solution to a sobering actuality because the nation grapples with an oversupply of underutilized knowledge facilities and shifting market dynamics.
Xiao Li, a former actual property contractor who pivoted to AI infrastructure in 2023, has witnessed this transformation firsthand via the fluctuating demand for Nvidia GPUs. A yr in the past, merchants in his community boasted about buying high-performance Nvidia GPUs regardless of U.S. export restrictions. Many of those chips have been illegally funneled into Shenzhen via worldwide channels. On the market’s peak, an Nvidia H100 – essential for coaching AI fashions – may fetch as a lot as 200,000 yuan ($28,000) on the black market.
Right now, Li seen that merchants have turn out to be extra discreet and GPU costs have stabilized. Moreover, two knowledge middle initiatives he’s acquainted with are struggling to draw additional funding as backers anticipate weak returns. This monetary pressure has pressured mission leaders to dump extra GPUs. “Everybody appears to be promoting, however there aren’t many consumers,” he informed MIT Expertise Evaluate.
Briefly, leasing GPUs to companies for AI mannequin coaching – a core technique for the newest technology of knowledge facilities – was as soon as thought-about a assured success. Nevertheless, the emergence of DeepSeek and shifting financial elements within the AI sector have put the nation’s knowledge middle trade on unstable floor.
The fast development of knowledge facilities throughout China, from Internal Mongolia to Guangdong, was fueled by a mixture of presidency directives and personal funding. Over 500 new initiatives have been introduced in 2023 and 2024, with at the very least 150 accomplished by the tip of 2024. Nevertheless, this constructing growth has led to a paradoxical scenario: an abundance of computational energy, notably in central and western China, coupled with a scarcity of chips that meet the present wants for inference and regulatory realities.
The rise of DeepSeek, an organization that developed an open-source reasoning mannequin matching the efficiency of ChatGPT however at a fraction of the price, has additional disrupted the market. Hancheng Cao, an assistant professor at Emory College, famous that this breakthrough has shifted the main focus from mannequin growth to sensible purposes. “The burning query shifted from ‘Who could make the most effective giant language mannequin?’ to ‘Who can use them higher?'”
This shift has uncovered the restrictions of many rapidly constructed knowledge facilities. Many amenities optimized for large-scale AI coaching are ill-suited for the low-latency necessities of inference duties wanted for real-time reasoning fashions. Because of this, knowledge facilities in distant areas with cheaper electrical energy and land are dropping their attraction to AI corporations.
The oversupply of computational energy has led to a dramatic drop in GPU rental costs. An Nvidia H100 server with eight GPUs now rents for 75,000 yuan monthly (round $10,345), down from earlier highs of round 180,000 yuan ($25,141). Some knowledge middle operators selected to go away their amenities idle reasonably than function at a loss.
Jimmy Goodrich, senior expertise advisor to the RAND Company, attributes this predicament to inexperienced gamers leaping on the AI bandwagon. “The rising ache China’s AI trade goes via is essentially a results of inexperienced gamers – companies and native governments – leaping on the hype prepare, constructing amenities that are not optimum for at present’s wants,” he explains.
China’s political system, with its emphasis on short-term financial initiatives for profession development, has performed a big position within the knowledge middle growth. Native officers, searching for to spice up their political careers and stimulate the economic system within the face of a post-pandemic downturn, turned to AI infrastructure as a brand new progress driver.
This top-down method usually disregarded precise demand or technical feasibility. Many initiatives have been led by executives and traders with restricted experience in AI infrastructure, leading to rapidly constructed amenities that fell in need of trade requirements.
The rise of reasoning fashions like DeepSeek’s R1 and OpenAI’s ChatGPT has shifted computing wants from large-scale coaching to real-time inference. This modification requires {hardware} with low latency, usually situated close to main tech hubs, to attenuate transmission delays and guarantee entry to expert employees.
Because of this, many knowledge facilities inbuilt central, western, and rural China are struggling to draw shoppers. Some, like a newly constructed facility in Zhengzhou, even distribute free computing vouchers to native tech companies however nonetheless battle to search out customers.
Regardless of the challenges, China’s central authorities prioritizes AI infrastructure growth. In early 2025, it convened an AI trade symposium emphasizing the significance of self-reliance on this expertise.
Main tech corporations like Alibaba and ByteDance have introduced important investments in cloud computing and AI {hardware} infrastructure.
Goodrich means that the Chinese language authorities views the present scenario as a needed rising ache. “The Chinese language central authorities will possible see [underused data centers] as a needed evil to develop an necessary functionality… They see the tip, not the means,” he says.
Because the trade evolves, demand stays sturdy for Nvidia chips, notably the H20 mannequin designed for the Chinese language market. Nevertheless, for a lot of within the discipline, like knowledge middle mission supervisor Fang Cunbao, the present state of the market has prompted a reevaluation.
In the beginning of the yr, Fang left the info middle trade completely. “The market is simply too chaotic. The early adopters profited, however now it is simply folks chasing coverage loopholes,” he explains. He is now shifting his focus to AI training.
