The National Security Case for Limiting China's Access to Advanced U.S. Compute: Evidence from PLA Procurement Documents

Cole McFaul is a Senior Research Analyst and an Andrew W. Marshall Fellow at the Center for Security and Emerging Technology (CSET). Sam Bresnick is a Research Fellow and an Andrew W. Marshall Fellow at CSET.

On Wednesday, April 22, the House Foreign Affairs Committee will discuss a package of bills aimed at strengthening restrictions on China's access to U.S. AI chips, semiconductor manufacturing equipment, and broader AI technology. These bills constitute a small subset of the dozens of related measures that have been proposed—but not yet passed—by this Congress. The flurry of legislative activity comes at a time when the White House appears open to negotiating with China over access to advanced, U.S.-origin technologies; in December, the Trump administration opened a licensing pathway for the export of Nvidia's H200 GPUs to China. Reports indicate that compute restrictions could be part of negotiations during next month's summit in Beijing between presidents Donald Trump and Xi Jinping.

Many important questions underlie these policy debates, including those concerning economic trade-offs, China's technological indigenization efforts, and the broader U.S.-PRC bilateral relationship, among others. While these are important considerations, this article addresses another critical question: will allowing China to access advanced U.S. compute help the People's Liberation Army (PLA) develop and deploy military AI systems?

Our research suggests the answer is yes. We have analyzed thousands of AI-related procurement documents published by the PLA during the past three years, and this piece elucidating the stakes of the restrictions on U.S. compute is the result of both previously published analysis and ongoing research, the findings of which will be released in the coming months.

In this piece, we detail two distinct mechanisms through which access to advanced U.S. compute can enable the PLA's adoption of military AI systems:

1. The PLA is procuring AI models trained on U.S. compute or distilled from U.S. models.
2. The PLA is acquiring advanced U.S. chips.

This analysis has led to three key takeaways relevant to the ongoing policy discussions:

1. Greater access to advanced U.S. compute will likely strengthen China's military AI capabilities.
2. The PLA will benefit from advances in China's civilian technology sector, including in AI.
3. The PLA's adoption of advanced AI capabilities may be used to undermine U.S. military advantages.

The PLA is procuring AI models trained on U.S. compute and/or distilled from U.S. models

Chinese political and military leaders believe that AI could help the PLA narrow the gap with the U.S. military, or even surpass it altogether. Many Chinese strategists view AI as a critical enabler of future military capabilities that will allow for faster and better decision-making, as well as more precise and efficient operations. They argue that the force that better develops and adopts AI-enabled military systems will gain an advantage in tomorrow's conflicts. Critically, previous CSET research has shown that Beijing's "intelligentization" efforts increasingly rely on technologies developed by the civilian sector—for example, we have reviewed procurement documents soliciting DJI drones, Unitree robotic dogs, and iFlyTek algorithms.

Forthcoming CSET research examines the PLA's procurement of large language models (LLMs) for military applications. Our analysis of procurement documents published by the PLA over the last three years shows that the Chinese military is actively adopting commercially available LLMs. We have reviewed requests for proposal (RFPs) for the deployment of AI models trained by some of China's top AI labs, including DeepSeek, Alibaba, iFlytek, and others. Crucially, many of these models, including DeepSeek's, were trained on U.S.-designed chips, illustrating the direct link between U.S. compute and Chinese military capabilities.

Although this brief is primarily concerned with the role of U.S. compute, we also note that LLMs developed by DeepSeek and other top PRC AI labs were also reportedly distilled from U.S. models. Distillation, a technique where one model is trained on the outputs of another, is an additional channel through which access to U.S. technology can ultimately bolster the PLA's access to advanced AI capabilities. We do not, however, delve into the details of distillation below.

The applications of these models are expansive. Our analysis suggests that the PLA is experimenting with LLMs across a wide range of military tasks, including for data analysis, cyber offense, cognitive domain operations, and coordinating the activities of unmanned vehicles, among other applications.

In August 2025, for example, a PLA unit requested a system that used a DeepSeek model for "multilingual translation, key information extraction, intelligent Q&A, intelligent document generation, flight path analysis, and image comparison analysis." A different document featured a request for a DeepSeek-enabled "unmanned [vehicle] cluster" command and control system. These systems could be useful for analyzing large volumes of data to inform military decision-making and coordinating drone swarms. Previous CSET research has examined China's investments in AI-enabled command, control, communications, computers, cyber, intelligence, surveillance, reconnaissance, and targeting technologies, and these documents further illustrate the PLA's focus on using AI for such tasks.

In March 2025, a military unit in Anhui Province published a procurement notice for a "Cybersecurity Professional Capability Generation Support System," a cyber range that must, according to the document, "achieve intelligent integration with DeepSeek and other large models," with capabilities including "intelligent attack" and "intelligent penetration." While our open-source data offers limited visibility into the PLA's offensive cyber capabilities, this document makes clear that the PLA is exploring the use of Chinese frontier models, including those trained using U.S. chips, to enhance its cyber operations.

Another RFP calls for a DeepSeek model for use with "psychological attack" and propaganda systems. Previous work has shown the PLA's interest in using AI to enhance cognitive domain operations, including via the generation of high-end deepfakes. The PLA seeks to use these capabilities to manipulate and degrade its adversaries' decision-making and thus gain military advantage. LLMs could prove especially useful for these and other kinds of influence operations.

While some have argued that U.S. export controls will not affect Beijing's military modernization ambitions, these documents reveal that efforts to constrain China's access to U.S. compute are still valuable for safeguarding U.S. national security interests. To our knowledge, China's leading commercial frontier AI labs—not the Chinese government—are developing the country's most capable AI models. Our findings, which show that the PLA is deploying those models for military use cases, therefore suggest that slowing China's development of frontier AI through compute restrictions will ultimately also hamper the PLA's use and adoption of advanced AI capabilities.

The PLA is acquiring advanced U.S. chips

The second mechanism through which U.S. compute can enable the PLA's adoption of AI is through the direct use of U.S.-origin chips. In our review of PLA procurement documents, RFPs for direct acquisitions of chips were the most common avenue through which the PLA tried to access U.S. compute. These RFPs included requests for both controlled and non-controlled chips.

Some have argued that the PLA is not interested in using U.S. chips for military activities. Notably, Nvidia's Jensen Huang has said that the PLA "simply can't rely" on U.S. chips due to security concerns, and that China does not "need Nvidia chips, certainly, or American tech stacks, in order to build their military." Our analysis, which relies on publicly available procurement documents published by the PLA, does not allow us to assess the scale of the PLA's use of U.S.-origin chips. Many of the procurement documents in our dataset include requirements to use domestic technology, reflecting Beijing's continued push for self-sufficiency. Beijing has also recently issued guidance that aims to limit Chinese AI labs' use of foreign chips, and DeepSeek and Alibaba have committed to increase their utilization of domestic chips for inference, for example. Other companies, such as Zhipu, are reportedly using Huawei's chips for LLM training.

At the same time, the documents in our dataset demonstrate that the PLA continues to request advanced U.S. chips for military applications. In May 2025, a PLA research laboratory requested sixteen Nvidia H100 GPUs, which are export controlled, to support research on electromagnetic-characterization modeling and infrared-characterization modeling using DeepSeek-70B and Alibaba's Qwen-32B LLMs. An RFP published in July 2025 requested a high-performance server equipped with Nvidia H20s in order to simulate a "demonstration and verification system," which could be useful for a wide variety of tasks, including weapons system modeling.

While most of the documents we have reviewed involve the procurement of AI chips for on-premises deployment, several RFPs include requests for access to cloud computing resources. One July 2024 RFP called for 10 million core hours of physical supercomputer cluster resources for use by a PLA unit, with an explicit requirement that the supplier "be able to provide NVIDIA V100 or A100 GPU compute resources." That a PLA unit was soliciting access to controlled U.S. hardware through an intermediary illustrates the many pathways military end users will pursue to gain access to these chips.

The procurement documents analyzed in this brief are just a few examples of the hundreds of RFPs soliciting advanced, U.S.-designed AI hardware between 2023 and 2026. Forthcoming research will more comprehensively examine the PLA's requests for compute resources.

Conclusion

In this piece, we refrain from analyzing specific ongoing legislative initiatives. Instead, we aim to provide more evidence of the national security imperatives for constraining China's ability to access advanced computing technologies.

To be sure, our analysis of publicly available PLA procurement documents does not allow us to make comprehensive claims about the PLA's adoption of military AI systems. Similarly, U.S. policymakers are currently making decisions with incomplete information about whether and to what extent U.S. AI compute will support the PLA in the short or long term.

Despite this uncertainty, we offer the following takeaways for U.S. policymakers:

1. Greater access to advanced U.S. compute will likely strengthen China's military AI capabilities, either through the PLA's deployment of AI systems trained using U.S.-origin technology or through the direct acquisition and use of advanced U.S. chips. China is working to develop a range of AI-enabled military capabilities, including decision-support systems and unmanned vehicle swarms, among myriad other applications.
2. The PLA will benefit from advances in China's civilian technology sector, including in AI. The PLA can access dual-use technologies, such as AI, both through contracts with nontraditional vendors and through the adoption of open-source frontier models, like DeepSeek's.
3. The PLA's adoption of advanced AI capabilities may be used to blunt U.S. military advantages. For example, the Chinese military is already experimenting with AI systems to improve its maritime domain awareness and space capabilities, potentially degrading perceived U.S. advantages at sea and in space.

CSET's analysis leaves little doubt that the PLA is moving quickly to develop and deploy military AI systems. Strengthened restrictions on China's access to U.S. technologies may carry costs for industry, but inaction risks far graver consequences for U.S. national security interests.

As always, we're glad to help - visit our support hub to contact us, book live support with an ETO staff member, or access the latest documentation for our tools and data. 🤖