1. Restricted Nvidia chips remain the blueprint for Chinese AI builds
Chinese developers are still drawing up and constructing artificial intelligence data centers around Nvidia processors that the United States restricts, according to an investigation published on November 13, 2026. The review of more than 150 procurement notices, investment plans and corporate records identified projects worth over 23 billion yuan, or roughly $3.2 billion, many of which name Nvidia's H100, H200 or H800 chips.
The findings point to a gap between policy and practice. Washington has spent four years tightening the flow of advanced AI semiconductors into China, and Beijing has spent the same period urging its technology sector to buy domestic alternatives. Yet when local governments, state-owned groups and private developers sit down to specify the hardware for new computing centers, the Nvidia name keeps appearing in the paperwork.
The chips named in the documents are the workhorses of modern AI. The H100 and H200 are among the processors most widely used to train large language models and to run them once they are deployed, a task known as inference. The H800 is a variant Nvidia built for the Chinese market. All three handle the heavy, repetitive mathematics that AI systems depend on, and all three sit inside the scope of US export controls in some form.
2. What the $3.2 billion figure does and does not show
The headline number needs careful reading. The roughly $3.2 billion represents the total value of the projects examined, not the value of Nvidia hardware inside them. It covers the full cost of each facility, including buildings, electricity supply and the rest of the supporting infrastructure that a large computing center requires.
Just as important, the records do not establish that the developers actually obtained every processor they listed, nor that any export law was broken. A procurement notice or an investment prospectus describes what a project intends to use. It is a statement of design and ambition, not a customs record or proof of delivery.
That distinction matters because Chinese planning documents often reflect what engineers would prefer to build with rather than what they can lawfully buy. A project can name H100 servers in its investment listing while still searching for financing, waiting on hardware or quietly preparing to substitute other chips. The documents therefore measure demand and dependence more clearly than they measure supply.
Read that way, the roughly $3.2 billion is best understood as a measure of how many Chinese AI projects still treat Nvidia as their reference platform, rather than as a tally of restricted chips that have crossed the border.
3. Projects from Fujian to Inner Mongolia name Nvidia hardware
The records span several provinces and range from early proposals to facilities already in operation. In Fujian, a proposed computing center valued at 6.2 billion yuan, or about $873 million, lists H100 servers in its official investment listing. That project has not yet closed its funding and is still looking for financing partners.
In Shaanxi, a state-owned engineering group won a contract linked to a 3 billion yuan computing center whose specifications name both the H100 and the H200. In Jilin, a project valued at 2.26 billion yuan proposes a hybrid approach that would pair Nvidia hardware with Chinese-made processors in the same facility.
A proposal in Inner Mongolia goes further in its detail, specifying 625 H200 servers. In the eastern city of Wuxi, local authorities reported that a cluster built on H800 chips began operating in March 2025.
The mix of stages is telling. Some of these entries describe facilities that exist and are running, while others describe investments that may never be completed. Having an operating center on the books and having a proposal on paper are not evidence of the same degree of access to restricted chips, and the documents should not be read as if every listed processor is sitting in a rack today.
4. Why do Chinese developers still want Nvidia chips?
The short answer is performance combined with software lock-in. Nvidia processors carry out enormous numbers of calculations at the same time, a property known as parallel processing, which makes them especially well suited to training AI models that learn from vast datasets.
The hardware is only half of the advantage. Nvidia's software tools are used so widely across the AI industry that most models, libraries and engineering workflows are built to run on them first. A team that wants to move to a rival chip often has to rewrite or re-tune code, retrain staff and accept a period of lower efficiency. That switching cost can be large, and it rises with the scale of the project.
For a developer planning a multibillion-yuan computing center meant to serve AI customers for years, specifying the platform that the market already knows best is the path of least resistance. The documents suggest that, on paper at least, many Chinese builders are still choosing that path even when the chips themselves are hard to get.
5. How US export controls on Nvidia China chips have evolved
The United States began restricting sales of advanced AI chips to China in 2022, citing national security. Since then the rules have been revised repeatedly as Washington has tried to keep pace with new products and with workarounds.
In January 2026, the Commerce Department shifted export license applications for the H200 to a case-by-case review. That change opened a narrow door rather than a wide one: any sale of the H200 to China still requires a license, and approval is not guaranteed. Shipments of the H100 remain under tighter restrictions.
The policy design leaves Chinese developers in an awkward position. They can keep building around Nvidia equipment that they may struggle to obtain legally, or they can invest in alternative processors, which can mean rewriting software and taking on extra engineering work. The Jilin proposal, which pairs Nvidia hardware with domestic chips, reads like an attempt to hedge between those two options.
For policymakers in Washington, the records raise a different question. If plans worth billions of yuan are still being drawn up around restricted hardware, officials will face pressure to show how they track whether those chips actually arrive, and through which channels.
6. Huawei narrows the gap in China's AI accelerator market
Beijing's push for self-sufficiency is producing measurable results. Research firm IDC estimated that Nvidia accounted for 34% of China's AI accelerator shipments in the first half of 2026, while Huawei held 28%. An AI accelerator is a chip designed specifically to speed up machine learning workloads, the category that includes Nvidia's data center GPUs and Huawei's competing Ascend line.
Those market share figures describe the broader Chinese market. They are not confirmed shipment counts of the specific restricted Nvidia models named in the procurement records, and they include products that are legal to sell into China.
Even so, a six-point gap between the US incumbent and its main domestic rival shows how far the landscape has shifted since export controls began. Chinese buyers that once had little choice beyond Nvidia now have a homegrown supplier with significant scale.
There is also evidence that capable AI models can be built without American hardware. Chinese AI developer Z.AI released an image-generation model trained entirely on Huawei chips, a result that undercuts the assumption that frontier-grade work in China depends on access to Nvidia processors.
7. What to watch next
Several threads in the records are still open. The Fujian computing center, valued at 6.2 billion yuan, has yet to secure its financing partners, so whether it proceeds, and with which processors, remains to be seen. The Inner Mongolia proposal for 625 H200 servers is another test case, since any lawful delivery of those chips would require a US export license under the case-by-case review the Commerce Department introduced in January 2026.
The hybrid design proposed in Jilin is worth following as well. If developers increasingly plan facilities that blend Nvidia hardware with Chinese processors, it would signal a gradual transition rather than a clean break, giving builders a fallback if the restricted chips never arrive.
Market share data will offer another signal. IDC's estimate for the first half of 2026 placed Nvidia only six percentage points ahead of Huawei in Chinese AI accelerator shipments, and future readings will show whether that margin keeps shrinking as domestic supply grows and as more developers follow examples such as Z.AI's Huawei-trained model.
8. What the records mean for Nvidia, Beijing and Washington
For Nvidia, the documents are a reminder of how deeply its products are woven into China's AI planning, even after years of restrictions that have limited what the company can legally sell there. Demand for its platform in the country has not disappeared; much of it appears to have been pushed into planning documents that may or may not be fulfilled.
For Beijing, the records show that the campaign to replace foreign chips is uneven. Domestic suppliers are winning share, and at least one Chinese developer has trained a model end to end on Huawei hardware, yet many local projects still default to Nvidia when they write down what they want to buy.
For Washington, the findings add to the scrutiny already facing officials responsible for export control enforcement. Members of Congress have been pressing for answers on how chip export rules are applied and policed, and evidence that billions of yuan in planned projects still name restricted hardware is likely to sharpen those questions.
The central tension is unresolved. As long as Nvidia's chips and software remain the default choice for AI builders worldwide, Chinese developers will keep designing around them, and US authorities will keep facing the challenge of ensuring that the chips themselves stay out of reach.