The views expressed are solely the author’s and do not reflect those of Perry World House, the University of Pennsylvania, or Carnegie Corporation of New York.
The Trump administration’s approach to technology competition with the People’s Republic of China is less aggressive than that of its predecessor. Spooked by Beijing’s rare earth material export controls, Washington has pursued a tech détente. The administration’s strategy may end up sacrificing long-term technological advantage for short-term bilateral stability.
The most worrisome aspect of the Trump administration’s tech policy reversal is its approach to artificial intelligence (AI) semiconductor export controls. The case for relaxing restrictions on cutting-edge AI semiconductors rests on a reassuring story: keep Beijing supplied, and it will remain hooked on U.S. technologies. Opponents of restrictions argue they will accelerate China’s drive toward semiconductor self-sufficiency. But this framing misses the mark. In reality, Beijing’s strategy is akin to having its cake and eating it too: maintaining access to high-end US semiconductors while simultaneously strengthening its domestic chip design and manufacturing capabilities.
Over a decade ago, China’s leadership identified advanced semiconductors as a strategic vulnerability and mobilized various chip-focused industrial strategies. To overcome its reliance on foreign companies for technologies across the semiconductor supply chain—including chip design software, cutting-edge lithography tools, and other semiconductor manufacturing equipment—Beijing has poured capital into its domestic semiconductor industry, and continues to do so to the tune of tens of billions of dollars. While the returns on these investments have been uneven, their scale and ambition are undeniable. Beijing’s intent is clear: it aims to innovate its way out of foreign dependence.
China is determined to address critical chokepoints where foreign companies have dominated chip design and production. In some cases, it has made progress. For example, Chinese companies are increasingly competitive in manufacturing mature-node semiconductors. They have also made advances in high-bandwidth memory chips, which are important for training and running large language models (LLMs). And Zhipu AI, a Chinese startup, recently released a model trained on domestically produced chips.
Elsewhere, however, China’s progress has been limited. Beijing has been unable to reduce dependencies on foreign firms for advanced lithography equipment and continues to lag in semiconductor design. Moreover, Chinese companies are forced to rely on older tools to produce serviceable AI chips using a highly inefficient process. Claims that Huawei is catching up with Nvidia are incorrect. The company will reportedly produce only 750,000 of its newest AI chips this year, amounting to less than 1 percent of the computing capacity produced by its U.S. competitors. Though China would prefer to power its AI ambitions with domestic chips, it does not produce enough high-quality semiconductors to do so, and even its best products lag those of foreign firms.
China’s goal is to be the world’s AI leader by 2030. Its chip shortages are the most significant threat to that ambition. If China wants to remain competitive with the United States in AI, then its companies will continue using leading U.S. compute resources, especially for AI training and high-performance applications, at least in the short term.
There is ample evidence of China’s continued reliance on advanced U.S.-designed chips. DeepSeek made international headlines in early 2025 by releasing a leading model trained on Nvidia semiconductors. Leading Chinese tech companies such as ByteDance and Alibaba continue to access advanced U.S. chips in third-country data centers to train their large language models. And while Nvidia’s Jensen Huang has repeatedly argued that the Chinese military is not interested in using his company’s products, there is robust evidence suggesting otherwise. The Trump administration’s nominee to lead the National Security Agency recently told Congress that the People’s Liberation Army (PLA) is “aggressively” seeking U.S. chips to fuel its modernization.
At the same time, Beijing is trying to reduce its dependence on U.S. technologies. It recently instructed domestic data center operators to source over 50 percent of their semiconductors from Chinese producers. Chinese fabs have been notified they should source 50 percent of their chipmaking tools from domestic manufacturers. Reporting indicates that Chinese officials have told domestic AI labs to use Chinese chips for AI inference, while they consider granting access to recently de-controlled Nvidia H200s for “advanced AI research and development.”
These initiatives clearly illustrate China’s dual-track semiconductor approach. The Chinese government continues to invest heavily in domestic semiconductor capacity and encourage AI developers to substitute home-grown technologies where they can. At the same time, Chinese companies are trying to ensure near-term access to the frontier, largely via high-end US compute. While the Chinese government has not yet approved imports of H200s, it aims to remain competitive technologically, militarily, and economically while pursuing self-sufficiency. Rather than keeping Beijing dependent, relaxing, or failing to enforce, export controls makes it easier for China to pursue its two-pronged approach.
At this point, imposing strict export controls will do little to further China’s ambitious self-sufficiency aims, but they would impede Beijing’s AI development pathway. China knows that the United States could restrict its access to advanced chips at any time. A temporary pause does not change that expectation, so Beijing will continue to support the domestic semiconductor industry with subsidies and other measures.
The right way to understand Beijing’s semiconductor strategy is as a portfolio approach. China is doing two things at once: hedging by investing in domestic capacity while Chinese companies maintain access to advanced foreign technologies. Now is the wrong time to drop restrictions in favor of short-term profits.
AI and the People’s Liberation Army
The stakes for decisions surrounding export controls are high, as the very chips that power advanced LLMs are being used, both directly and indirectly, to boost China’s military modernization and preparation for future conflicts. Chinese defense strategists believe that intelligentization, the third phase of China’s military modernization, will see AI enable autonomous operations and accelerate targeting cycles. Key to China’s vision of future wars is gaining information dominance, or controlling the flow, quality, and utility of information while denying an adversary the ability to do the same. AI is becoming critical to fusing and analyzing battlefield information, so it is no surprise that the PLA is investing heavily in this and related technologies.
A recent report from Georgetown University’s Center for Security and Emerging Technology examining thousands of publicly available PLA procurement documents offers unusually direct evidence of what the PLA’s investments in AI look like in practice, and why access to advanced US-designed chips remains strategically consequential. In short, the Chinese military is attempting to adopt AI across command and control, communications, computers, cyber, intelligence, surveillance, reconnaissance, and targeting (C5ISRT), and for improved autonomy. The PLA is not only seeking AI-enabled capabilities, but also the semiconductors, often those designed by U.S. companies, that make those capabilities viable.
Operationally, the PLA is requesting AI-enabled autonomy and across all domains. There are requests for the quick iteration and prototyping of technologies for piloting unmanned combat vehicles to computer vision-enabled detection and identification of targets on land, at sea, in the air, and in space, as well as interest in drone swarms that can coordinate attacks. Procurement requests also point to the PLA’s broad robotics portfolio, including for robotic dogs and humanoids, and to large quantities of small, cheap drones, echoing the logic of massed, expendable systems.
A second point of emphasis is on using AI to gain decision advantage. The PLA appears intent on mobilizing AI systems to ingest, augment, fuse, and analyze huge volumes of data to accelerate and improve decision-making in wartime. While many militaries are developing AI decision support systems (AI-DSS), the PLA may be particularly interested in them as a means of compensating for a weak, inexperienced officer corps. Chinese strategists distrust the PLA’s chain of command and worry it could be outpaced in a fast-moving conflict. Many of the documents feature requests for AI-DSS at the operational and strategic levels. Moreover, there are requests for planning and situational awareness-like tasks, as well as for anticipating an adversary’s moves. The operationalization of AI-DSS alongside AI-enabled deepfake systems, requests for which there were several, could lead to miscalculations and unintended escalations.
Finally, the documents feature myriad requests for emerging technologies for cyber, information, and cognitive warfare, and efforts to erode U.S. advantages under the sea and in space. The PLA appears to be seeking AI solutions to detect network intrusions, harden communications, and enhance cyber operations and influence tools to manipulate perceptions and decision-making, and for training missions on virtual battlefields and modeling competitor behavior. In space, the PLA is requesting satellite-targeting algorithms. And at sea, the documents specify an interest in autonomous underwater vehicles and sensor networks, potentially aimed at tracking U.S. submarines.
Many of the requests are for comparatively small sums and specify short timelines. Furthermore, they are for a range of AI applications across domains, indicating that the PLA is intent on quickly and cheaply determining which technologies work. By layering AI and other emerging technologies on top of its extant platforms and processes, the PLA appears to believe that AI-enabled improvements will compound over time, thus leading to military advantage.
This is the operational context in which chips matter. Many of the C5ISRT tasks that require AI, including fusing sensor feeds, running (and training) LLMs, and enabling decision support systems all require advanced compute. The PLA’s ambition to broadly deploy its AI tools across domains and missions may be constrained by lack of access to cutting-edge hardware. That said, several AI applications, particularly those related to computer vision, do not necessarily require advanced chips, so it is unreasonable to believe that a return to stricter export controls would imperil China’s military AI ambitions across the board.
That said, greater restrictions and better enforcement would adversely impact the Chinese military. There are two ways that advanced U.S.-designed semiconductors could benefit the PLA. The first is through the PLA’s direct acquisition of these products. Again, there is clear evidence that the Chinese military is seeking to acquire U.S. chips, including those that have been subject to export controls, for use in military systems. There is no indication, however, that U.S. companies are supplying the semiconductors. Third-party distributors are likely providing the technologies to the PLA. Second, U.S. chips can support PLA modernization through their use in training Chinese LLMs that the military then adopts. For example, DeepSeek, which trains its models on U.S. hardware, appears repeatedly in PLA procurement documents. It follows, then, that increased access to U.S. technologies will further enable China’s accelerating military modernization.
There is no question whether advanced chips can contribute to China’s military capabilities. It is clear that they can, and the U.S. government should aim to reduce the role U.S. technologies play in fueling the PLA’s modernization. Limiting the flow of advanced chips is a useful way to maintain a compute advantage over China and the PLA.
Of course, export controls are not a panacea for the China tech threat. They are porous, as evidenced by largescale and widespread smuggling operations. Evasion will continue through transshipment networks, shell companies, and gray-market brokerages. Alongside stricter export controls, the United States must boost enforcement capabilities to limit flows of these advanced technologies into China.
Even with tighter restrictions, Chinese organizations will seek to design and manufacture domestic chips, train models, and adapt to hardware constraints. U.S. planners should therefore treat export controls as a delaying tactic that buys the Pentagon time. Concurrently, the U.S. military must compete with the PLA not only industrially and technologically, but also operationally. This involves hardening U.S. forces against the AI-enabled sensing and surveillance capabilities China is pursuing, and adapting to a future where countries flood the information environment with AI-enabled mis- and disinformation and manipulate data streams to combat the use of AI-DSS. It also means experimenting with AI-enabled technologies and capabilities with an eye toward quickly and effectively adopting the technology.
