China's Military Steals Nvidia Chips Despite AI Ban

China officially banned its institutions from using Western AI models and restricted purchases of Nvidia chips (specialized processors that power large AI systems, originally designed for graphics rendering but repurposed as the primary hardware for training AI). It then proceeded to steal them anyway. That is the core finding from Georgetown University's Center for Security and Emerging Technology — CSET, a think tank dedicated to the national security implications of emerging technologies — whose senior fellow delivered that conclusion directly to the U.S. Congress in April 2026.

The research lands at a precise inflection point: the Pentagon simultaneously announced deals with 7 major tech companies to deploy AI inside classified military systems. As the U.S. accelerates forward on military AI, CSET's data shows China's military apparatus is acquiring the same foundational technology through three separate channels that export controls have not closed.

China's Nvidia AI Ban — What Export Controls Are Missing

The contradiction at the heart of CSET's findings: China publicly banned access to foreign AI models and restricted Nvidia GPU (graphics processing unit — the chip type that became the primary hardware for AI training after proving able to run millions of parallel calculations simultaneously) purchases. These measures were framed as both ideological, keeping foreign AI influence out of Chinese systems, and practical, protecting domestic AI development from dependence on Western infrastructure.

In practice, CSET Senior Fellow Andrew Lohn testified before the U.S.-China Economic and Security Review Commission (an independent government body that monitors the national security implications of U.S.-China trade and economic relations) in April 2026 that the bans have not stopped acquisition:

"Despite banning the models and restricting chip purchases, they are certainly taking active steps to acquire the technology. That includes activities to acquire expertise, hardware, and models."

Lohn identified three distinct acquisition channels still operating in parallel:

• Expertise recruitment: Targeting U.S.-trained AI researchers, particularly those with knowledge of frontier model development (building cutting-edge AI systems with the highest available capabilities), through financial incentives and joint research partnerships
• Hardware acquisition: Obtaining Nvidia chips and high-bandwidth memory (the specialized fast-access RAM that large AI training runs require) through third-party intermediaries who circumvent export restrictions
• Model weight extraction: Acquiring model weights (the millions of trained numerical parameters that define exactly how an AI system responds and reasons) through cyberattacks or insider access to AI company infrastructure

China's Military AI Blueprint: 14 PLA Competitions Decoded

To map where China's military AI investment is actually directed, CSET researchers analyzed a dataset of 14 PLA (People's Liberation Army — China's combined military force encompassing army, navy, air force, rocket force, and cyber units) technology competitions running from January 2023 through December 2024. Unlike most China threat assessments that rely on secondhand reporting, this study used primary documents: actual PLA procurement announcements and Chinese-language news sources analyzed directly.

Three findings from the 14-competition dataset reshape conventional threat assessments:

• Multi-domain integration is the stated priority: 10 of 14 PLA challenges deliberately span at least two operational domains simultaneously — combining air, cyberspace, space, and information warfare rather than treating them as separate capability silos. This matches PLA doctrine for what Chinese military strategists call "intelligentized warfare."
• Drone warfare dominates the agenda: 5 of 14 challenges explicitly reference UAV swarm countermeasures (systems designed to intercept, jam, or destroy coordinated fleets of autonomous drones) or offense-defense applications. Over half of 18 nationwide Chinese technology challenges analyzed across 2023–2025 focused on unmanned aerial vehicles and counter-UAV applications.
• Military-civil fusion is measurable, not theoretical: Competition participants included defense state-owned enterprises (government-controlled arms manufacturers), military universities including the National University of Defense Technology, civilian research universities like Peking University, and commercial private tech firms — all feeding directly into PLA capability pipelines.

Scale context matters: a single competition held jointly in Nanjing and Qingdao attracted more than 100 participating teams. These events are not research conferences — they are compressed technology procurement pipelines that transform academic AI research into operational military tools on timelines that outpace traditional defense acquisition cycles.

Pentagon Signs 7 AI Deals Amid China's Military AI Acquisition

While China builds capability through the back channels described above, the U.S. Department of Defense is accelerating through corporate partnerships at the front. In May 2026, the Pentagon announced agreements with 7 major tech companies to integrate AI directly into classified military systems — environments handling sensitive intelligence data, targeting coordination, and command infrastructure that cannot run on commercial cloud services by definition.

CSET Interim Executive Director Helen Toner framed the core military appeal of these deployments:

"A lot of modern warfare is based on people sitting in command centers behind monitors, making complicated decisions about confusing, fast-moving situations. AI systems can be helpful in terms of summarizing information or looking at surveillance feeds and trying to identify potential targets."

Classified military AI deployments require security clearances (government-issued authorizations to access sensitive national security information, requiring background investigations that can take years), air-gapped networks (computing environments physically disconnected from the public internet to prevent data exfiltration), and hardware that has passed DoD supply chain verification. The simultaneous clearance of 7 companies signals a DoD-level policy shift — the U.S. military's AI integration timeline has been substantially compressed.

The China-U.S. AI Arms Race: Who Is Actually Winning?

The most sobering disclosure in CSET's research is not a statistic — it is an admission about the limits of current knowledge. When pressed on who is winning the AI arms race in cyberspace, Lohn's Congressional testimony was direct:

"It is not clear yet who benefits between attackers and defenders. The real-world evidence so far shows offense mostly experimenting, while defenders are starting to be overwhelmed from too much help that could potentially be turned against them."

The structural problem: AI-powered cyberattack tools (automated programs that scan networks for vulnerabilities, generate novel exploits, and adapt attack strategies faster than human security teams can respond) are evolving faster than defensive detection and response frameworks can scale. The same AI tools deployed in security operations centers (SOCs — the centralized teams monitoring an organization's network 24/7 for threats) to help defenders triage alerts can be reverse-engineered and weaponized by adversaries who have already acquired the underlying model capabilities through the three channels described above.

CSET's parallel concern extends to orbital infrastructure. In a Newsweek op-ed, researchers Kathleen Curlee and Brian Golden argued that "principles alone are not enough to protect space" — the same logic applies to AI infrastructure broadly. Voluntary commitments and export control principles have not stopped the hardware smuggling, talent recruitment, and model weight extraction that Lohn documented before Congress. Space-based AI systems now carry the same vulnerability profile.

3 Actions CSET Recommended to Congress on China's AI Threat

CSET's research is built for policy implementation, not just academic citation. Lohn's April 2026 Congressional testimony translated the analysis into three concrete legislative and regulatory recommendations:

• Mandatory federal cyber standards: Establish binding baseline security requirements for AI systems deployed in critical infrastructure — not voluntary frameworks that organizations can opt out of based on cost-benefit calculations, but enforceable minimums with consequences for non-compliance
• AI talent retention programs: Create dedicated federal initiatives to keep top AI researchers working in the United States, directly addressing the expertise acquisition channel China is already running at scale through university recruitment and research funding incentives
• Conditional federal AI funding: Tie federal AI research investment to enforceable security safeguard requirements — ensuring that taxpayer-funded AI advances cannot themselves become future acquisition targets through the same three channels currently being exploited

The policy community is paying attention. CSET's "Keeping Top AI Talent in the United States" report generated 125 points and 105 comments on Hacker News — unusually high engagement for a policy research paper, indicating that the technical community already recognizes this as an active, present-tense problem rather than a hypothetical future concern.

The 7 Pentagon AI partnerships represent the acceleration lever. CSET's research is the argument for simultaneously closing the acquisition gap. Both doors are currently open: the U.S. military is integrating AI into classified systems at unprecedented speed, while China's military is acquiring the same foundational technology through channels that have not been sealed. The CSET analysis delivers a direct brief to policymakers — acceleration and security cannot be treated as separate problems with separate timelines.

Access the primary research: CSET's full PLA Challenges and Competitions report is freely downloadable at cset.georgetown.edu. The Andrew Lohn Congressional testimony is publicly available through the U.S.-China Economic and Security Review Commission. For AI automation context and guides, explore the AI learning guides on this site.

Related Content — Get Started | Guides | More News