China’s People’s Liberation Army is concentrating its artificial intelligence spending on a narrow band of military applications rather than trying to match the United States across every domain, according to research from Georgetown University’s Center for Security and Emerging Technology. That selective strategy is unfolding as multiple open-source analyses and U.S. policy measures point to a U.S. advantage in computing hardware and supporting infrastructure, raising questions about whether targeted bets can offset structural disadvantages. The gap matters now because both militaries are racing to integrate AI into operations that could shape any future conflict in the Indo-Pacific.
What is verified so far
The clearest window into PLA procurement priorities comes from CSET’s analysis of Chinese and American military AI purchases, which used procurement records from a snapshot period in 2020. That research sorted contracts into seven application areas: autonomous vehicles, intelligence surveillance and reconnaissance (ISR), logistics, electronic warfare, simulation and training, command and control (C2), and automatic target recognition (ATR). The data showed that the PLA clustered its buying around a subset of those categories, while U.S. defense spending spread more evenly across all seven, according to CSET’s procurement comparison.
A separate CSET report examined what the PLA itself signals it wants from AI by analyzing military wish lists and research priorities. The CSET study on PLA AI priorities provides canonical metadata, including authors, date, and DOI, and links to related primary documents in the same research series. Together, these two datasets offer the most granular publicly available picture of how Beijing is directing military AI dollars.
On the hardware side, a West Point analysis described a large infrastructure mismatch, citing over 4,000 data centers in the United States compared to roughly 400 in China. That same analysis described the U.S. as holding a “commanding lead” in installed infrastructure and investment volume, a gap reinforced by export controls that Washington first imposed in October 2022.
Those controls are administered by the Bureau of Industry and Security, which maintains a dedicated rule page covering advanced computing and semiconductor manufacturing items restricted for export to China. The BIS page points to controlling Federal Register notices and subsequent updates, making it the authoritative reference for the scope of restricted hardware. By restricting access to certain high-end chips and related semiconductor manufacturing equipment, the rules can limit the computing power available to China for AI training and inference at scale.
Quantitative research on AI supercomputer trends, published as a preprint, supports the broader claim about lopsided compute distribution. The arXiv analysis examined the scale of leading systems and the share of global AI supercomputer performance held by each country, with the United States holding a leading share in the dataset the authors assembled. While the paper is not peer-reviewed government data, its explicit methodology and dataset make it a useful primary technical source for tracking hardware concentration.
The Pentagon’s annual report to Congress on military and security developments involving China rounds out the official baseline. The Department of Defense publishes releases and materials related to that assessment on its site, and the report is routinely cited as an authoritative reference in think-tank analyses of PLA modernization and emerging-technology priorities.
What remains uncertain
Several significant gaps limit how confidently analysts can assess the real balance of military AI power. The CSET procurement data relies on a 2020 snapshot, meaning it predates both the tightening of U.S. export controls and Beijing’s subsequent push to accelerate domestic chip production. No publicly available primary dataset tracks PLA AI contracts after that period with the same rigor, so any claim about current Chinese military spending patterns carries an inherent lag.
Direct official statements from the PLA about its AI strategy are scarce in English-language open sources. Most Western assessments, including the CSET reports and the Pentagon’s annual China review, infer priorities from procurement records, published doctrine, and academic output rather than from on-the-record Chinese military declarations. That means the picture of what Beijing actually intends could differ from what outside researchers observe in contract filings.
The arXiv supercomputer analysis, while methodologically transparent, stops at estimates available through 2024. Rapid developments in Chinese AI, including the emergence of competitive open-source large language models and efforts to work around hardware restrictions, suggest the compute gap may be shifting in ways that current datasets do not yet capture. Without updated primary research on AI supercomputer deployments inside China, any present-tense statement about the exact size of the hardware gap should be treated with caution.
There is also a conceptual blind spot in the dominant Western framing. Most analyses measure advantage by counting data centers, chip performance, and investment dollars. But the PLA’s selective approach hints at a different theory of competition: that targeted software innovation and dual-use civilian-military technology transfers could allow China to extract outsized military value from a smaller hardware base. If Beijing succeeds in optimizing open-source AI algorithms for specific battlefield tasks such as ISR or electronic warfare, raw compute totals may overstate the practical American lead. No current quantitative study has tested that hypothesis rigorously.
How to read the evidence
The strongest evidence in this debate comes from primary procurement records and official government documents. The CSET procurement comparison and the PLA wish-list analysis, both built on original contract data and documented research priorities, offer verifiable counts and categories rather than impressionistic judgments. The BIS rule page and the Pentagon’s annual report similarly carry institutional weight because they reflect formal U.S. government positions backed by classified and unclassified intelligence.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
