China’s planned AI buildout in Xinjiang is moving ahead with an unresolved question at its core: how projects expecting restricted Nvidia hardware intend to get it. Bloomberg’s reporting points to a gap between the ambitions written into project documents and the limits imposed by U.S. export controls.
The issue is not only about one region or one supplier. It shows how demand for high-end AI infrastructure can keep rising even when access to the most desirable chips is formally restricted.
What Bloomberg found in the project plans
Bloomberg’s investigation found that 39 planned data centers in Xinjiang and Qinghai are expecting to use more than 115,000 Nvidia H100 or H200 chips. Those models have been subject to U.S. export restrictions since 2022.
The documents reviewed by Bloomberg do not state how the planned facilities intend to obtain the restricted hardware. That missing detail matters because the plans appear to assume access to chips that are not supposed to be freely available to China under U.S. rules.
One project highlighted in the reporting involves Chinese energy company Nyocor. In its first phase, Nyocor plans to install 250 H100 servers, equal to around 2,000 Nvidia chips. Those servers are intended to be leased to Infinigence AI, a company that wants computing power to become as broadly available as utilities such as electricity or water.
The scale of the plans makes the supply question difficult to treat as a minor procurement issue. If the data centers are built around expectations of Nvidia H100 or H200 access, the chips are not incidental. They are central to the computing capacity these projects are designed to deliver.
Why Xinjiang is central to the buildout
Roughly 70 percent of the planned computing capacity is expected to be concentrated in a single government-backed data center complex in Xinjiang, near the Gobi Desert city of Yiwu. The facility is designed to house thousands of high-performance GPUs.
Local authorities are offering incentives for the project. Bloomberg cites free electricity quotas, operating subsidies, and other benefits. The site is expected to rely mainly on wind, solar, and coal, while the cool desert climate is intended to help lower costs.
Those details explain why the location is attractive for large-scale AI infrastructure. High-performance computing requires substantial power and cooling. A site that combines energy access, government support, and a climate that can reduce operating costs can make a data center complex more appealing, especially when the planned hardware is measured in thousands of GPUs.
At the same time, the concentration of capacity in one large complex makes the hardware question more visible. A plan of this size depends on more than ordinary equipment sourcing. It requires confidence that a large number of restricted chips can be secured, installed, and operated.
The export-control dispute
According to Bloomberg, U.S. Commerce Department official Jeffrey Kessler considers the smuggling of restricted Nvidia chips into China an established reality. That view suggests U.S. authorities see systematic efforts to get around export controls as a serious concern.
Nvidia CEO Jensen Huang has taken a different public position, denying evidence of AI chip diversions. Nvidia also maintains that it does not provide technical support for illegal products and views large-scale smuggling as commercially and technically unviable.
Bloomberg did not find direct evidence of centralized procurement for the chips. Even so, the reporting points to U.S. investigations and accounts of increasingly sophisticated smuggling operations. The examples cited include routes through Malaysia and hidden shipments under live seafood.
The result is a sharp tension between three facts presented in the reporting:
- Planned projects expect to use restricted Nvidia H100 or H200 chips.
- The documents do not explain how those chips would be obtained.
- U.S. officials and Nvidia disagree over whether diversion is occurring at scale.
For the planned data centers, that uncertainty has not stopped construction from moving forward. Bloomberg’s account describes activity continuing as if the required hardware is expected to arrive.
Domestic chips have not closed the gap
The reliance on Nvidia chips also reflects the state of China’s domestic high-end AI chip options as described in the source reporting. Despite significant investment and state support, China has not yet developed internationally competitive high-end AI chips.
Huawei’s Ascend 910C is described as the leading domestic contender. Analysts estimate that it still trails Nvidia’s China-focused H20 chip by about a year. The upcoming 910D is expected to surpass the H100 in performance, although it will require more power.
That comparison helps explain why project planners would still be interested in restricted Nvidia hardware. If local alternatives are improving but have not yet matched the strongest available options, builders of large AI data centers may continue to design around Nvidia GPUs where they believe access is possible.
The power requirement also matters. A chip that can exceed another in performance may still create practical tradeoffs if it consumes more energy. For large data centers, performance is only one part of the operating equation. Power use, cooling, availability, and technical support all shape whether a chip can support a large deployment.
What the plans reveal
The Bloomberg report shows an AI infrastructure push that is being planned around constrained hardware. Xinjiang and Qinghai projects are not merely discussing general computing expansion. They are tied to specific Nvidia chips that fall under U.S. export restrictions.
That makes the plans a test of whether export controls can meaningfully limit access to advanced AI hardware when buyers, builders, and local authorities still have strong incentives to keep projects moving. The reporting does not prove a centralized procurement channel, and it does not show exactly how the chips would be acquired. It does show that major data center plans are written as though restricted hardware will be available.
For now, the central contradiction remains unresolved. U.S. officials say smuggling is an established reality, Nvidia rejects the idea that large-scale diversion is practical, and planned data centers in Xinjiang and Qinghai continue to point toward a future built on Nvidia H100 and H200 chips.