China’s Multibillion-Dollar AI Data Center Plans Still Specify Restricted Nvidia Chips

Table of Contents

• You might want to know
• Main Topic
• Key Insights Table
• Afterwards...

You might want to know

• What does the reported $3.2 billion figure actually represent—and does it confirm that restricted Nvidia chips were purchased?
• Why are Chinese developers still specifying Nvidia hardware while domestic alternatives are expanding?

Main Topic

Chinese companies and public-sector organizations continue to plan artificial intelligence data centers around Nvidia processors, even though U.S. export controls restrict access to some of the company’s advanced chips. A Reuters investigation identified more than $3.2 billion in Chinese AI projects that specify Nvidia hardware, based on a review of more than 150 procurement notices, investment plans and corporate records. The projects collectively represent more than 23 billion yuan ($3.2 billion) in planned or reported investment.

The figure requires careful interpretation. It is the overall value of the projects, not a confirmed total for Nvidia chip purchases. Project budgets can include land, buildings, electrical systems, cooling equipment, networking and other infrastructure. The documents reviewed by Reuters do not establish that every specified processor was acquired, delivered or installed. Nor do they establish that any export law was broken. The reported $3.2 billion describes project investment, not verified spending on restricted chips.

Several records illustrate the range of plans. In Fujian province, an official investment listing describes a proposed computing center valued at 6.2 billion yuan ($873 million) and specifies servers using Nvidia H100 chips. The project was still seeking financing partners, so its listing demonstrates an intended hardware configuration rather than a completed purchase or operating facility.

In Shaanxi, a state-owned engineering group won a contract connected to a 3 billion yuan computing center that names Nvidia H100 and H200 chips. In Jilin, a separate 2.26 billion yuan project proposes combining Nvidia hardware with processors made in China. These examples show that project plans can include a mix of foreign and domestic technology, but the records alone do not establish which chips ultimately enter service.

Another proposed development in Inner Mongolia specifies 625 H200 servers. By contrast, authorities in Wuxi reported that an H800 computing cluster began operating in March 2025. These cases have different evidentiary weight: a proposal or procurement specification describes an intention, while a report of an operating facility indicates that a system has started functioning. Even an operational report, however, does not by itself resolve how the equipment was sourced or whether all its components are covered by the same restrictions.

The distinction matters because Nvidia’s H100, H200 and H800 processors are associated with advanced computing used to train and run AI models. Nvidia chips can perform thousands of calculations simultaneously, a capability that makes them valuable for demanding workloads. Their appeal is not limited to hardware performance. Nvidia’s widely used software ecosystem can also make it more straightforward for developers to build and operate AI systems on its platforms.

That software familiarity can create switching costs. Moving a project to another processor may require developers to adapt or rewrite software, test systems, retrain staff and make additional engineering decisions. For a data center planned around a particular hardware and software stack, changing chips can therefore affect budgets, schedules and technical design. Developers face a practical choice: continue planning around Nvidia equipment that may be difficult to obtain, or adopt alternatives that could require substantial adjustments.

U.S. controls have shaped that choice since the United States began restricting sales of advanced AI chips to China in 2022, citing national security concerns. The source article reports that, in January 2026, the Commerce Department moved H200 export applications to case-by-case review. Sales still require licenses, and H100 shipments remain more tightly restricted. The review process does not amount to automatic approval: each application remains subject to licensing decisions.

This policy environment helps explain why project documents can name restricted products without proving that those products were delivered. Planning and procurement records may reflect a preferred design, a sought-after configuration or a project that remains dependent on financing and approvals. Restrictions can also change which equipment is available over time. Consequently, the gap between a specification and a completed data center can be significant.

At the same time, the project records suggest that Nvidia remains embedded in Chinese AI infrastructure planning. Developers may continue to list its processors because of performance, software compatibility or established workflows, even when access is uncertain. The persistence of those specifications is an indication of demand and technical preference, but it should not be treated as a shipment tally or as evidence that controls have been circumvented.

China’s domestic chip sector is also gaining ground. An IDC report cited by Reuters estimated that Nvidia accounted for 34% of China’s AI accelerator shipments in the first half of 2026, while Huawei accounted for 28%. Those figures refer to the broader AI accelerator market. They are not a confirmed count of restricted Nvidia models, and they should not be confused with the value of the project investments identified by Reuters.

The market-share estimates indicate competition between Nvidia and Chinese suppliers, while also showing that the transition is not necessarily complete. Huawei’s reported share points to the growing role of domestic processors. Nvidia’s reported share, meanwhile, suggests that the company remained a substantial participant in China’s AI accelerator market during the period covered by the estimate.

One example of a domestic alternative comes from Chinese AI developer Z.AI, which has released an image-generation model trained entirely on Huawei chips. The example demonstrates that at least some AI development can proceed without American hardware. It does not, by itself, establish that every Nvidia-dependent workload can be transferred easily, or that domestic systems match Nvidia platforms across all applications and operating requirements.

For companies planning large computing facilities, the decision is therefore broader than choosing a processor. It involves considering hardware availability, software compatibility, system performance, project financing, regulatory requirements and the long-term support of the chosen platform. A design based on a difficult-to-access processor may face delays or require revision. A design based on domestic alternatives may reduce exposure to export restrictions but require engineering work and changes to existing software.

The findings also have implications for U.S. officials. The records add to scrutiny of how advanced-chip export rules are implemented and enforced. They underline the difference between projects that name controlled equipment and actual, verified shipments. For policymakers, this distinction is important: investment plans can help reveal market demand and intended technical choices, but they cannot independently confirm whether restricted chips reached a particular buyer or facility.

Overall, the Reuters review offers a snapshot of a market navigating both technological preference and regulatory uncertainty. Chinese organizations continue to include Nvidia processors in some AI infrastructure plans, while domestic alternatives are developing and export licensing remains consequential. The documents point to continued Nvidia relevance in project planning, but they do not prove that all listed chips were obtained or that restrictions were violated.

Key Insights Table

Afterwards...

The next phase of China’s AI infrastructure buildout will depend on more than the chips named in project plans. Financing, licensing decisions, hardware availability, software compatibility and the capabilities of domestic processors will all influence which proposals become operating facilities. As projects move from planning to deployment, clearer evidence about completed purchases and installed systems will help distinguish intended configurations from actual chip access.

For developers, continued uncertainty makes flexibility valuable. Some may keep Nvidia-based designs while evaluating alternatives; others may prioritize domestic processors and invest in the engineering needed to adapt their systems. For regulators and observers, the central task will be to separate investment announcements and procurement specifications from verified shipments. Those distinctions will be essential to understanding how export controls affect China’s AI sector and how quickly domestic alternatives can meet its needs.