Chinese AI projects still specify restricted Nvidia chips in designs

Persistence in project designs

Chinese infrastructure projects worth over 23 billion yuan ($3.2 billion) continue to specify Nvidia processors in their official plans. As reported by Unite.AI, these figures come from a review of more than 150 procurement notices, investment plans, and construction contracts. The data covers a wide range of activities, including the building of data centers and the development of power systems.

The scope of these projects is geographically diverse, stretching from coastal technology hubs to remote western provinces. These initiatives are not limited to major technology companies; they also involve regional computing centers, universities, and ventures backed by local governments. This indicates that the desire for specific high-end hardware is widespread across different sectors of the Chinese economy.

While these documents show a clear intent to use Nvidia chips, they do not confirm that the hardware has actually been delivered. There is a significant gap between a project's design phase and its final hardware installation. Because of US export controls, many of these projects may face delays or be forced into costly redesigns to accommodate domestic alternatives.

Specific hardware targets

The tender documents highlight a specific interest in the Nvidia H100 and H200 processors for large-scale AI development. For example, a 6.2 billion yuan computing center in Xianyou county, Fujian, is currently seeking commercial partners while specifying H100 chips. Similarly, a 3 billion yuan project in Shaanxi won a contract that specifically ties the facility to H100 and H200 processors.

Other regional developments show a mix of hardware preferences. A 2.26 billion yuan initiative in Jilin pairs H100 units with processors built in China. Meanwhile, a development in Inner Mongolia valued at 1.5 billion yuan proposes the use of 625 H200 servers. These projects illustrate how different regions are balancing restricted technology with domestic options.

In Gansu, a local company was awarded a 1.16 billion yuan construction contract for a computing center. This facility plans to house approximately 3,072 H100 and H200 processors. These specific numbers provide a clear picture of the scale at which Chinese entities are planning to deploy restricted computing power.

Operational status of clusters

While many projects are still in the planning phase, some restricted hardware is already operational in China. Local authorities in Wuxi announced in March 2025 that an H800 cluster had begun functioning in Jiangsu. This suggests that some clusters have successfully bypassed or secured the necessary licenses to begin functioning.

Damingcheng also disclosed that a venture in Fujian built an H800 cluster. This project originally involved a computing-services agreement worth 759 million yuan, though that specific contract ended early. These instances show that despite the stated goal of the US government to limit access, some high-end clusters are already active.

The existence of these operational clusters complicates the narrative of a total blockade. It raises questions about how the restricted hardware reached these users. The current data does not establish the specific pathways used to move these chips into the country or the exact identities of the end-users.

The limits of export controls

The United States began restricting Nvidia's advanced AI processors in 2022 and subsequently tightened controls on modified versions like the H800. The Trump administration has also initiated a process to review H200 export licenses for eligible Chinese customers on an individual basis. These measures are intended to slow the growth of Chinese AI capabilities by limiting the hardware available for training large models.

Nvidia has stated that these controls effectively shut it out of the Chinese AI data-center market. Consequently, the company has excluded China data-center computing revenue from its financial forecasts due to regulatory uncertainty. However, the procurement documents show that Chinese customers are still designing their infrastructure around Nvidia's specifications rather than pivoting away from them.

Analysts suggest that the current enforcement strategy may need to evolve to be more effective. Sunny Cheung of the Jamestown Foundation noted that the government should move past looking only at the initial buyers. He argued that the focus should shift to examining who ultimately uses the computing power and for what specific purposes.