A Bloomberg News analysis of investment approvals, tender documents and company filings shows that Chinese firms aim to install more than 115,000 Nvidia Corp. AI chips in some three dozen data centers across the country’s western deserts. Operators in Xinjiang intend to house the lion’s share of those processors in a single compound — which, if they can pull it off, could be used to train foundational large-language models like those of Chinese AI startup DeepSeek.

The complex as envisioned would still be dwarfed by the scale of AI infrastructure in the US, but it would significantly boost China’s computing prowess as President Xi Jinping pushes for technological breakthroughs. Such a project also would raise serious concerns for officials in Washington, who restricted leading-edge Nvidia chip sales to China in 2022 over worries that advanced AI could give Beijing a military edge.

Yet the Chinese documents contain no explanation of how companies plan to acquire the chips, which cannot be legally purchased without licenses from the US government, permits that haven’t been given. The companies listed in the filings, state officials and central government representatives in Beijing declined to comment when asked to explain.

To gauge whether Chinese entities could realistically procure that quantity of restricted processors, Bloomberg News spoke with more than a dozen people who’ve been involved in or privy to US government investigations into the matter, as well as several people with direct knowledge of the black market in China.

Listen: Big Take: China’s Got Big Plans for AI — In the Desert (Podcast)

None of those familiar with the US probes said they previously knew of the data center buildout in Xinjiang. All said that while they believe there are indeed banned chips in China, they’re not aware of an illicit trade network sophisticated enough to procure more than 100,000 such processors and direct that hardware to a centralized location.

But the US government doesn’t appear to have reached a consensus on the number of restricted Nvidia chips currently in the Asian country. Most of the people interviewed for this story said they were unaware of an agreed-upon estimate, while some offered rough numbers that differed by tens of thousands of processors.

Two senior Biden administration officials said they believe there are around 25,000 banned Nvidia chips in China — a number that, one of them added, would not be terribly concerning. That volume of semiconductors, assuming they are integrated into servers and designated for the same facility, could power at most one mid-sized data center.

The US Commerce Department — whose Bureau of Industry and Security, known as BIS, is tasked with implementing and enforcing chip trade restrictions — did not answer detailed questions for this story, including how many banned Nvidia chips the Trump administration believes are in China, nor whether Trump officials were previously aware of the projects in Xinjiang.

“Posting a web page asking about restricted products is not the same as successfully licensing, building, and operating a datacenter,” Nvidia said in an emailed response to questions about the Chinese companies’ claims. “Datacenters are massive and complex systems, making smuggling extremely difficult, and we do not provide any support or repairs for restricted products.”

The California-based company also said that “trying to cobble together a datacenter from smuggled, previous-generation products makes no business or engineering sense,” especially since chips and servers made by Huawei Technologies Co. are widely available in China. Jensen Huang, Nvidia’s chief executive officer, made his position clear at a May conference in Taipei: “There’s no evidence of any AI chip diversion,” he said.

Yet the head of BIS pointedly contradicted that assertion just weeks later, telling US lawmakers that there is clearly a problem with AI chip smuggling. “It’s happening,” said Commerce Under Secretary Jeffrey Kessler. “It’s a fact.” Although Kessler didn’t mention Nvidia by name, the company is by far the dominant provider of such semiconductors. Kessler also said that US efforts to restrict Huawei’s chipmaking capabilities will keep China’s output at just 200,000 AI processors this year — a number far short of domestic demand.

To be sure, Bloomberg News has not found evidence that China has amassed, or can amass, 115,000 banned Nvidia chips — nor evidence that smaller volumes of restricted semiconductors that US officials believe are in the country have been directed to centralized locations.

And yet in Yiwu, the construction goes on.

Looming out of the desert, a tower the height of the Golden Gate Bridge radiates an intense light that pierces the surrounding dust clouds. Arrays of reflectors focus the sun’s energy onto a receiver that allows the daytime heat of the arid plains to be stored, ensuring continuous power generation.

It’s one main reason for the choice of Yiwu, just to the south over a mountain pass. On the barren hill behind one new building stands a wall with a slogan picked out in red Chinese letters two meters high: “Data-electricity fusion shows great promise.”

Xinjiang, and especially the Hami region which includes Yiwu County, is rich in wind and solar energy, as well as abundant in coal, offering a ready source of affordable power. Local governments there are at the forefront of a state strategy to take advantage of those energy resources — along with cheap land and cool weather at altitude, helping counter the heat generated by racks of servers — to meet the AI computing-power demand of more economically developed regions such as Shanghai and Shenzhen.

On a midweek day in March, workers loaded windmill blades onto the back of trucks traveling the road between the prefectural capital of Hami City and Yiwu, over bleak terrain past occasional camels grazing, and through a new tunnel leading out to a plain with views of snow-capped mountains. The main road into town leads past the first data center, still under construction, with a man welding from his perch on metal scaffolding.

Hami is best known for its sweet melons, and Yiwu claims to be the site of the last battle on the mainland of the Chinese civil war in 1949. There’s a monument downtown dedicated to a horse that played a role in the final engagement between Communist forces and nationalists loyal to Chiang Kai-shek.

The authorities in Xinjiang are particularly suspicious of foreigners due to Western allegations of human-rights abuses against ethnic Uyghurs. Interview requests sent to eight data center operators in Yiwu were ignored, rejected or agreed to and then cancelled at short notice. The Xinjiang government and Ministry of Industry and Information Technology (MIIT), the central government ministry overseeing data center development, didn’t reply to Bloomberg requests for comment.

The most important part of a giant data center is relatively small. Nvidia dominates the market for so-called AI accelerators, highly coveted components that have propelled the chipmaker’s valuation to nearly $4 trillion. The processors are connected together in giant arrays numbering tens of thousands and used to sift through mountains of data to create new computer code that can in many ways approximate human intelligence.

The US barred China from importing Nvidia’s best chips in October 2022, a month before OpenAI’s ChatGPT debut roiled the tech industry and sparked a global race that now includes DeepSeek among its top players. Washington several times has ratcheted up those curbs, restricting sales to China of a variety of advanced semiconductors and the machines used to make them — with additional sanctions levied on specific Chinese tech companies. That sweeping effort, which dates back to Trump’s first term, has become a primary source of tension with Beijing — one that Chinese officials repeatedly raised in recent trade talks with the US after the Trump administration imposed punitive tariffs.

“All the greatest chips in the world are American, right? So of course they want them,” Commerce Secretary Howard Lutnick told CNBC last month, speaking about China’s position during negotiations in London. “And of course we said ‘absolutely not.’”

The Xinjiang effort suggests that China’s AI ambitions — which hinge in large part on locally produced chips from the likes of Huawei — still include some hope of accessing restricted Nvidia hardware too. Project approval documents show that in the fourth quarter of 2024, local governments in Xinjiang and in neighboring Qinghai province green-lit a total of 39 data centers that intend to use more than 115,000 Nvidia processors.

All of the companies stated in their investment plans that they aim to obtain H100 or H200 chips, two Nvidia GPUs, or graphics processing units, that were the industrial standard for training large language models such as OpenAI’s GPT4o and Google’s Gemini through last year. Nvidia this year debuted a new, more advanced model — dubbed the Grace Blackwell — that is banned along with the H100 and H200 from export to China without a US government license.

Seven Xinjiang projects that aim to use those processors had started construction or won open tenders for AI computing service as of June 2025, according to tender documents obtained by Bloomberg. One operator says it’s already using advanced hardware facilities to support cloud access to DeepSeek’s R1 model, according to local news reports. Still, the provincial projects’ description of their intended computing capabilities may be somewhat aspirational: Local party officials try to signal to Beijing that they are working toward national priorities, but Chinese companies frequently launch initiatives that are never completed.

Nyocor declined to comment. China Bester and China Energy Investment didn’t reply to requests for comment. Infinigence AI couldn’t be reached for a response.

Around 70% of computing power planned by the identified projects is in a single compound set up by the local government in Xinjiang. That makes the region — the epicenter of Western charges of Chinese rights abuses including forced labor and religious persecution — pivotal to China’s efforts to seize the lead from the US in a sphere seen as key to future global technological, and geopolitical, dominance.

Even if successful, the Xinjiang complex would only involve the number of Nvidia chips that one major hyperscaler — a term for massive data center operators like Microsoft Corp. and Amazon Web Services — deploys in a single week, according to data Nvidia provided on a recent earnings call. Still, Chinese companies like DeepSeek are beginning to show they can do more with less.

“The gap between leading US and Chinese AI labs is closing,” said Kevin Xu, a tech investor and founder of US-based Interconnected Capital, who put it at around three months. Players like DeepSeek, which says it trained its R1 model using less-advanced Nvidia chips, are “very serious and sincere” about pursuing artificial general intelligence, Xu said. The fact that leading Chinese models are open source means they spread faster globally, he added, while noting that diffusion is hard to track: “Beijing sees this trend as a source of technological soft power worth embracing.”

DeepSeek and other Chinese AI startups have already expressed interest in collaborating with the data center projects in Xinjiang, according to an employee of one of the largest investors in the Yiwu sites. That employee, whose name has been withheld to protect their identity, said in a message exchange that their company will invest more than 5 billion yuan ($700 million) in data center projects there in 2025 and 2026.

China’s data center industry is expected to surpass 300 billion yuan in scale this year, according to the Securities Times. Chinese entities are collectively expected to invest nearly that amount on an annual basis by 2028, according to the China Communications Industry Association — a more than threefold increase from a half-decade prior.

Xinjiang has already brought its first “intelligent computing center” online, and constructed 24,000 petaflops of computing power for demand from the logistics hub of Chongqing, Chairman of the People’s Government of Xinjiang Erkin Tuniyaz said in an annual government work report in January, without specifying the type of chips installed. The cited computing power is equivalent to roughly 12,000 server-integrated Nvidia H100s.

Prospective investors in such projects are attracted with the promise of free electricity worth up to 20% of total power costs. Data center operators also can access government support ranging from one-off payments for construction to operation incentives for up to five years, depending on company size, according to local government documents reviewed by Bloomberg. Experts in “green computing” areas are also eligible for favorable terms on accommodation, children’s education and research funding.

From a standing start, “Xinjiang’s intelligent computing has achieved a historic breakthrough,” Tuniyaz said in January.

Policymakers in Washington for years have been aware that limiting China’s access to US technology is not as simple as writing a regulation. Not two months after the chip restrictions took effect, Chinese officials caught a woman hiding forbidden hardware in a baby bump. The American AI company Anthropic recently said smugglers have packed GPUs next to live lobsters. Nvidia has dismissed both examples as “tall tales” that ignore the complexity of building data centers, which require operational support to run properly — support that Nvidia does not provide for restricted products in China.

Still, conversations with people privy to illicit semiconductor transactions, as well as media reports from a range of outlets, indicate that smuggling networks have gotten more sophisticated over time. Those stories — which have helped inform US investigations, people familiar with the matter said — have cited examples ranging from dozens of illicit processors to more than a thousand.

Potential smuggling in Malaysia has become a big concern for the Trump administration, which plans to restrict Nvidia sales there to halt possible diversion to China, and also has asked Malaysian authorities to crack down on the issue — a request the government has said it’ll heed. Officials in Singapore, meanwhile, are prosecuting three men for alleged fraud in exports to Malaysia of AI servers that likely contained advanced Nvidia processors — bound for an unknown final destination.

In response to queries about Washington’s export control plans, Malaysia’s Ministry of Investment, Trade & Industry said the country will “act firmly against any company or individual should there be strong evidence” of misuse or diversion of advanced tech.
The ministry added that Malaysia welcomes a dialogue with the US and other nations to “clarify any misunderstandings and to strengthen mutual trust.”

Trump officials are separately investigating whether DeepSeek may have accessed restricted chips through intermediaries in Singapore, and a bipartisan congressional committee focused on China recently requested Nvidia’s customer data for 11 Asian countries, related to concerns that DeepSeek may have circumvented US export controls. (None of the documents viewed or interviews conducted through the course of this investigation indicated any link between the Xinjiang projects and supply chains in Singapore or Malaysia. Nvidia is not accused of any wrongdoing in Singapore’s probe or in the US investigation into DeepSeek.)

Read More: Lutnick Urges Tougher Enforcement of Export Curbs on China

Nvidia consistently has said it abides by all US rules, but Huang has made no secret that he doesn’t like Washington’s strategy. Years of curbs — including on crucial semiconductor manufacturing equipment — have “failed” to contain Huawei’s rise, he said at the May conference in Taipei. Nvidia now sees Huawei as a formidable competitor, and the company worries its Chinese rival will continue to improve and gain market share — unless the US government allows Nvidia to compete on Huawei’s home turf.

Washington isn’t buying it. The Trump administration has already further limited the types of chips Nvidia can sell in China, at a $5.5 billion hit to the company. White House AI Advisor Sriram Krishnan, asked about Huang’s urge to lift those curbs, said that “there is still bipartisan and broad concern about what can happen to these GPUs once they’re physically inside” the Asian country.

Meanwhile, Chinese companies continue to build their data centers, a sign they expect to receive AI chips from somewhere.

Two such construction projects were approved by the Qinghai government in December 2024, with a total investment of 13.5 billion yuan, documents from Qinghai’s investment review website show.

The companies applying for construction permits for both projects were founded that same month. China’s company registry services show both entities can be traced by shareholding data to the same group of controlling companies: one real estate firm in Qinghai named Qinghai Borong Group and one AI tech company in Sichuan called Chengdu Qingshu Technology. They didn’t respond to requests for comment.

Neither is on Nvidia’s official resellers list.