China’s multi-billion-dollar AI infrastructure gold rush is now unraveling, in accordance with a brand new investigative report from Caiwei Chen, a reporter masking know-how matters associated to China at MIT Expertise Assessment.
“Simply months in the past, a growth in knowledge middle building was at its top, fueled by each authorities and personal traders. Nonetheless, many newly constructed amenities are actually sitting empty,” Chen writes, noting the publication has sources on the bottom together with contractors, an govt at a GPU server firm, and challenge managers.
The sources say corporations who constructed datacenters throughout the growth are actually treading water, and in accordance with Chinese language information shops, as much as 80% of China’s newly constructed computing assets stay unused.
A Frenzy of Funding Meets a New Actuality
Why all these idle datacenters? Chen’s report reveals a cooling mixture of speculative investments primarily based on AI hype converging with the explosive launch of DeepSeek.
The so-called “DeepSeek Impact” has rippled via the AI market and international economic system, because the mannequin’s effectivity at coaching with diminished infrastructure has undercut demand for large-scale datacenter deployments. It’s probably reshaping, although not but redefining, the present trajectory of AI infrastructure technique. China could also be feeling the earliest and most seen impacts.
The first enterprise mannequin for China’s new datacenters has been renting out GPU clusters for mannequin coaching. However as DeepSeek’s open-source mannequin R1 matched ChatGPT-level efficiency at a fraction of the fee, smaller gamers have deserted their plans to pretrain massive fashions. The demand for coaching infrastructure has collapsed, simply as a whole lot of recent amenities had been prepared to come back on-line.
In parallel, the technical calls for of AI are shifting. At the moment’s compute bottleneck isn’t in coaching however in inference. Particularly, working fashions in actual time, particularly for step-by-step reasoning duties. These workloads require {hardware} optimized for low-latency efficiency and geographic proximity to main tech hubs. But a lot of China’s new knowledge facilities had been in-built rural or distant areas, optimized for the flawed type of AI work, and too removed from expertise and finish customers, Chen reveals.
After ChatGPT’s debut in late 2022, China moved quick to place AI infrastructure as a nationwide precedence. The central authorities inspired native officers to develop AI datacenters, dubbed “good computing facilities,” and by 2024, over 500 had been introduced. At the least 150 had been accomplished, backed by state-owned companies and public corporations desperate to trip the AI wave.
However the growth was short-lived. Of the greater than 140 corporations registered in 2024 to construct massive fashions, solely about 10% remained lively by yr’s finish. Most of the companies and traders concerned had little AI experience and had been chasing authorities subsidies, low cost electrical energy, or entry to state loans, in accordance with this report.
“The rising ache China’s AI business goes via is essentially a results of inexperienced gamers—companies and native governments—leaping on the hype practice, constructing amenities that aren’t optimum for immediately’s want,” says Jimmy Goodrich, senior advisor for know-how to the RAND Company, as quoted in Chen’s report. Goodrich says it’s probably that the Chinese language authorities will step in to imagine administration and reassign these property to extra succesful operators, now that traders are in search of to promote at below-market charges.
The AI Arms Race Isn’t Slowing Down
Whilst a lot of China’s new knowledge facilities stay idle, the nation’s central authorities reveals no indicators of retreating from its AI ambitions. In early 2025, it convened a high-level AI symposium, doubling down on infrastructure as a nationwide precedence and emphasizing the necessity for technological self-reliance. Main companies are following swimsuit: Alibaba has pledged greater than $50 billion towards cloud and AI {hardware} over the subsequent three years, whereas ByteDance plans to spend $20 billion on GPUs and datacenters.
(Supply: Dabarti CGI/Shutterstock)
The USA is pushing simply as laborious, Chen notes. The Stargate initiative, backed by OpenAI, Oracle, and SoftBank, goals to speculate as much as $500 billion into superior knowledge facilities globally. For Beijing, the thought of scaling again now could be unlikely. As RAND’s Jimmy Goodrich places it, underused infrastructure could also be considered not as failure, however as “a essential evil” in constructing long-term functionality.
Demand for Nvidia chips just like the H20 (a chip custom-designed for the Chinese language market) and the H100 stays sturdy, pushed partially by Chinese language corporations racing to deploy their very own variations of DeepSeek’s open-source fashions. However infrastructure alone is now not the bottleneck.
As Fang Cunbao, a longtime knowledge middle govt, instructed Chen: “The market is simply too chaotic. The early adopters profited, however now it’s simply individuals chasing coverage loopholes.” He’s since left the business, satisfied that future progress relies upon not on constructing extra, however on having clear, deployable plans for utilizing AI effectively. “What stands between now and a future the place AI is definitely in all places,” he says, “is just not infrastructure anymore, however strong plans to deploy the know-how.”
You possibly can learn the complete reporting by Caiwei Chen at MIT Expertise Assessment right here.
