Reuters finds $3.2 billion of planned AI data centers specifying Nvidia chips Washington restricts
A set of compute projects still hunting for financing and still tendering have already written their machine models into public documents, and those models are the ones Washington will not sell.
Three key facts
Reuters published an investigation on November 13 after reviewing more than 150 procurement notices, investment plans and corporate records. It identified AI data center projects worth over 23 billion yuan, about $3.2 billion, that explicitly specify Nvidia's H100, H200 or H800 chips, the processors used to train and run advanced AI models. One proposed computing center in Fujian is valued at 6.2 billion yuan, about $873 million, and its official investment listing names H100 servers, though it is still seeking financing partners.
The number needs careful reading. The $3.2 billion is total project investment covering buildings, power and other infrastructure. The documents do not establish that all the processors were obtained, and they are not evidence that export laws were broken. Elsewhere, a state-owned engineering group in Shaanxi won a contract tied to a 3 billion yuan center naming H100 and H200 chips; a 2.26 billion yuan project in Jilin proposes mixing Nvidia hardware with Chinese processors; Inner Mongolia has a proposal listing 625 H200 servers; and authorities in Wuxi reported an H800 cluster began operating in March 2025.
On policy, the United States has restricted advanced AI chip sales to China since 2022. In January 2026 the Commerce Department moved H200 export applications to case-by-case review, though licenses are still required, and H100 shipments remain more tightly restricted. Meanwhile, an IDC report cited by Reuters estimated Nvidia held 34 percent of China's AI accelerator shipments in the first half of 2026, against 28 percent for Huawei.
Wang Dou's Take
Clear up the easiest misreading first. The $3.2 billion is not chips bought under the table. It is the construction budget for these projects. A building still raising money that says it will house H100s, and a building that actually houses them, are worlds apart.
The real information is somewhere else. Planners would rather write H100 servers into a project document that has not secured financing, for a model they most likely cannot get. Why? Because tender documents are written for review panels, investors and local officials, and H100 clears approval and tells a better story than a domestic card. That is not evidence of smuggling. It is evidence of expectations.
Then look at the market shares: 34 percent against 28 percent. Washington has restricted sales for four years, and Nvidia is still number one in this market. Controls changed the supply structure. They have not yet changed habits, and the CUDA software ecosystem is harder to replace than the silicon. Z.AI has already trained an image model entirely on Huawei chips, so the alternative road exists. Its toll is just not yet low enough for most to take it.
Source: Decrypt
