According to Reuters' report of November 13, AI computing center projects across China that plan to use GPUs subject to U.S. export controls, such as NVIDIA's H100 and H200, total more than $3 billion. The figure suggests that demand for controlled GPUs persists in China. However, the amount includes project costs listed in procurement documents and investment plans. It does not represent the value of GPUs actually delivered. Looking at the five projects for which Reuters cited specific amounts, about 60% are still at the concept or solicitation stage, and no GPU deliveries have been confirmed for any of them. The numbers are better read as an indication of the scale of planned projects than as a record of actual procurement.

AD

NVIDIA GPU-based AI computing center plans are under way across China

The projects Reuters reviewed span multiple provinces and autonomous regions, from coastal areas to the interior, and include some involving state-owned enterprises. In Xianyou County, Fujian Province, an AI computing center with a total investment of 6.2 billion yuan (about $873 million) is being planned, and the project is currently seeking commercial partners. According to the Fujian provincial investment promotion website, the first phase is expected to deliver roughly 4,000 PFLOPS of sparse computing performance, and all of its compute servers are planned to use NVIDIA H100s.

In Shaanxi Province, a subsidiary of the state-owned engineering giant Sinomach won a contract related to an AI computing center worth 3 billion yuan. The GPUs specified for use are the H100 and H200. In Jilin Province, a 2.26 billion yuan computing center is planned, combining the H100 with Chinese-made AI chips. About 80% of the required funding is expected to come from bank loans.

In Inner Mongolia, a 1.5 billion yuan project plans to deploy 625 H200 servers. According to Chinese media reports in June, construction has begun on two buildings that make up the first phase. In Gansu Province, a 1.16 billion yuan computing center construction contract was signed in July, with a plan to deploy about 3,072 NVIDIA GPUs, counting the H100 and H200 together.

There are also cases where computing centers using NVIDIA GPUs are already in operation. In March 2025, the government of Wuxi, Jiangsu Province, announced that a computing cluster using NVIDIA H800s had begun operating. In June 2026, Shanghai Stock Exchange-listed Damingcheng disclosed that a related entity in Fujian had built an H800 cluster and signed a computing services contract initially worth 759 million yuan. That contract was later terminated early, however. It shows that building the equipment and securing ongoing demand for it are separate matters.

About 60% of the five projects with reported values are at the concept stage

The table below summarizes the status of the five projects for which Reuters cited specific amounts, as of the time of reporting. The figures mix total investment, construction contract values and overall project scale, and none of them is the purchase price of the GPUs themselves.

Project Reported amount Status
Xianyou County, Fujian 6.2 billion yuan (total investment) Seeking commercial partners
Shaanxi 3 billion yuan (scale of computing center) Related contract won
Jilin 2.26 billion yuan (total investment) Listed in a government investment brochure
Inner Mongolia 1.5 billion yuan (project scale) Construction of two first-phase buildings reportedly started; GPU delivery unconfirmed
Gansu 1.16 billion yuan (construction contract value) Construction contract won; GPU delivery unconfirmed

Status reflects what can be confirmed from Reuters' reporting. Damingcheng's computing services contract is not included. The Fujian and Jilin projects are classified as concept/solicitation stage, and the remaining three as contract/construction stage.

The five projects total 14.12 billion yuan (about $1.99 billion). Of this, the Fujian and Jilin projects, which are at the concept or solicitation stage, account for 8.46 billion yuan, or 59.9% of the total. For the remaining three projects, worth 5.66 billion yuan, what can be confirmed is limited to contract signings and the start of construction; no GPU deliveries have been confirmed.

For the dollar conversion, we used the rate Reuters gave for the Fujian project: 6.2 billion yuan = $873 million. At an assumed rate of 157 yen to the dollar, the five-project total of about $1.99 billion comes to roughly 310 billion yen.

These five projects, however, are only a part of the more than $3 billion total Reuters reported. Individual amounts and progress for the remaining projects could not be confirmed, and the overall total also includes facilities already in operation. The roughly 60% share at the concept or solicitation stage therefore cannot be applied to the entire $3 billion-plus pipeline.

The fact that project documents list NVIDIA GPU models is also not evidence that procurement has been finalized. The Fujian computing center plan, for example, is a project listed on a government investment promotion site to attract companies. Although it states that H100s will be used, this reflects the current proposed equipment configuration and does not mean a procurement contract has been concluded. The investment being sought is also 6.2 billion yuan, the same as the total investment, and no investors are named.

The Jilin project is similarly at the planning stage, with about 80% of the required funding expected from bank loans. Even when a GPU model is written in the specifications, financing, export licensing, equipment orders and delivery each require separate procedures. Amounts in an investment plan cannot be treated as the value of equipment whose procurement is settled.

The 1.16 billion yuan Gansu contract was likewise won by a consortium led by a subsidiary of the state-owned Shaanxi construction company, and it covers construction work. The materials Reuters reviewed contain no information on GPU suppliers or delivery dates. Even if construction of a computing center is progressing, that does not mean the GPUs to be installed inside have been delivered.

AD

Facilities already in operation and the U.S. export licensing system

It has also been confirmed that facilities using controlled NVIDIA GPUs are already running in China. Besides the H800 cluster in Wuxi, universities continue to make small purchases. According to Reuters, Shenzhen University signed a contract in September 2026 to order three H100s, and Northwestern Polytechnical University, which is on the U.S. export control Entity List, bought two H200s the following October. These are examples of demand translating into concrete contracts and purchases.

Being subject to U.S. export controls also does not mean a blanket ban on every transaction. On January 15, 2026, the U.S. Commerce Department's Bureau of Industry and Security (BIS) amended its rule on exports of advanced semiconductors to China, changing the license review policy for some products, including NVIDIA's H200, from a presumption of denial to case-by-case review.

The covered products are advanced computing items with a total processing performance (TPP) below 21,000 and total DRAM bandwidth below 6,500 GB per second. The rule gives the H200 and AMD MI325X as examples, and equivalent or lower-performance products that meet the conditions are also eligible for case-by-case review.

Obtaining an export license, however, requires meeting multiple conditions. The total number of such products shipped to China and Macau is capped at 50% of the total shipped for end use in the United States. Exporters must also avoid delaying supply to U.S. customers and must not divert production capacity for equivalent or better chips to China. Other requirements include pre-shipment testing by a third-party laboratory in the United States and customer verification measures to prevent unauthorized remote access by restricted parties.

Transactions that violate other export controls, such as those involving military or nuclear-related uses, are not permitted. If an application meets these conditions and is approved, covered products such as the H200 can be exported to China. Note that while the rule cites the H200 as a specific example, it does not name the H100. Whether the projects reported this time have obtained export licenses also cannot be determined from publicly available procurement documents.

Given this track record of operating facilities and the licensing framework, it is not unnatural for Chinese AI computing centers to plan to adopt NVIDIA GPUs. Operating facilities and purchase contracts do exist. But those are separate from the five projects whose amounts were tallied here, and the fact remains that no GPU deliveries have been confirmed for those five.

Going forward, the clearest indicators of whether plans have turned into actual equipment procurement will be the delivery status of the 625 H200 servers planned in Inner Mongolia and the roughly 3,072 NVIDIA GPUs planned in Gansu. If their delivery or acceptance is announced, it would confirm that planned demand has translated into actual equipment installation.