Newsletter
On the edge of the Gobi Desert in China’s Xinjiang region, construction crews are racing to build something unusual. Not factories or highways, but rows of data centers: large, futuristic buildings that will house racks of servers designed to power artificial intelligence.
At the center of this story is a small county called Yiwu, tucked away in the Hami prefecture of eastern Xinjiang. Local authorities there have approved one of China’s most ambitious AI infrastructure projects yet: a data center complex that, according to official filings reviewed by Bloomberg, aims to run on over 115,000 Nvidia AI chips.
That’s a surprising number, not because of the scale, but because of the restrictions. Since 2022, the United States has banned the export of Nvidia’s most advanced AI chips like the H100 and H200 GPUs to China, citing national security concerns. These chips are used to train large language models, the same kind that power tools like ChatGPT.
How, then, do Chinese firms plan to acquire more than 100,000 of them?
That question has no clear answer. Bloomberg’s analysis of Chinese investment filings, procurement records, and tender documents found repeated references to these banned chips. Yet there is no mention of how they will be obtained, legally or otherwise. US officials interviewed said they were unaware of any black-market network capable of sourcing chips at that volume. Nvidia itself has stated it offers no technical support for restricted chips in China and called large-scale smuggling infeasible.
Still, the projects are moving forward. Construction is active across at least 39 planned AI data centers in Xinjiang and neighbouring Qinghai province. One of the biggest is backed by Nyocor, a solar and wind energy firm from Tianjin. The company has begun installing servers for a data center in Yiwu, intended to be used by Infinigence AI, a fast-growing AI infrastructure firm that says it wants to make computing power as accessible as electricity.
Yiwu’s location isn’t accidental. The region offers an unusual combination of cheap land, abundant renewable energy from wind and solar, and a high-altitude climate ideal for keeping servers cool. Local governments are offering power subsidies and five-year incentive packages to attract investment in what they call “green AI computing.” A slogan painted on the wall behind one of the data center buildings reads: “Data-electricity fusion shows great promise.”
The facilities are part of China’s broader push to close the AI infrastructure gap with the United States. While the US still dominates with hyperscale data centers run by Amazon, Microsoft, and Google, China is investing heavily to catch up. According to industry estimates, China’s data center industry will exceed 300 billion yuan in value this year, and investments are expected to nearly triple by 2028.
But none of that solves the chip problem. Washington’s export controls, introduced under the Trump administration and tightened further since, specifically aim to slow down China’s access to high-performance AI hardware. US officials have flagged Malaysia and Singapore as potential diversion routes and are investigating whether companies like DeepSeek—one of China’s leading AI startups—may have indirectly obtained restricted chips through overseas intermediaries.
Even so, the numbers remain unclear. US government estimates suggest around 25,000 banned Nvidia chips might already be in China. That’s a fraction of what Chinese documents claim is needed. Bloomberg found no evidence that the 115,000 figure has been achieved—or that China currently has the supply chain to get there.
Yet the belief that these chips will eventually arrive seems to be enough for now. Companies are building the infrastructure. Workers are laying cables and installing cooling systems. And in some cases, data centers are already online, serving demand from major cities like Shanghai and Chongqing.
What makes this story notable is not just the number of chips or the secrecy around their sourcing. It’s the clarity with which it shows that the global race for AI dominance is no longer just about talent or algorithms. It’s about supply chains, energy access, and geopolitical control over the smallest but most powerful pieces of silicon.
For now, no one has proven how the chips will get in. But the desert construction continues.
FAQs
Why is China building data centers in Xinjiang?
China is building AI data centers in Xinjiang due to its cheap land, abundant renewable energy, cool climate, and government incentives. The region offers ideal conditions to host power-intensive server farms that support artificial intelligence computing for companies across the country.
What are Nvidia H100 and H200 chips used for?
Nvidia H100 and H200 chips are high-performance GPUs used in artificial intelligence, especially for training large language models like ChatGPT. These chips provide the computing power needed for deep learning and are considered critical infrastructure in the global AI race.
Why has the US banned Nvidia’s AI chips from being exported to China?
The US banned Nvidia’s top-end AI chips in 2022 over concerns they could be used for military, surveillance, or geopolitical advantage. The goal is to limit China’s access to high-performance computing that could accelerate its defence and AI capabilities.
How many banned Nvidia chips does China reportedly want?
Chinese documents reviewed by Bloomberg show that companies plan to install over 115,000 Nvidia AI chips in data centers across Xinjiang and Qinghai. However, there’s no official confirmation on whether China has actually acquired these chips.
Can China legally buy Nvidia AI chips?
No, China cannot legally purchase Nvidia’s H100, H200, or newer Grace Blackwell chips without a US government license. These chips are subject to export restrictions aimed at preventing advanced AI tech from reaching Chinese companies.
Is there proof that China has obtained banned Nvidia chips?
There is no confirmed evidence that China has amassed 115,000 banned Nvidia chips. US officials estimate about 25,000 may be present in China, but large-scale smuggling or legal sourcing has not been verified.
What is DeepSeek AI and its role in this story?
DeepSeek AI is a Chinese AI startup developing large language models. It is one of the companies reportedly interested in using the new data centers in Xinjiang. Some reports suggest it may already be using older-generation chips for training.
Why is Xinjiang’s role in AI development controversial?
Xinjiang is under global scrutiny for alleged human rights violations, particularly against Uyghurs. Building advanced AI infrastructure there raises concerns about surveillance, control, and the use of restricted technologies in a politically sensitive region.
What are the US concerns about chip smuggling to China?
The US fears that banned AI chips could be reaching China through grey markets or third countries like Malaysia or Singapore. Some investigations have uncovered potential rerouting and illicit shipments, though no large-scale networks are confirmed.
Is Nvidia accused of wrongdoing in this situation?
No, Nvidia is not accused of any wrongdoing. The company has stated it complies with all US export rules and does not support or service restricted products in China. It has also denied the feasibility of large-scale chip smuggling.
