The U.S. government has spent three years building an export-control wall around America’s most advanced artificial intelligence technology, yet federal prosecutors keep uncovering holes. A string of espionage indictments, GPU smuggling arrests, and congressional scrutiny of Chinese AI firms suggests that Washington’s trillion-dollar bet on keeping its AI edge is being tested by determined adversaries using theft, transshipment networks, and creative workarounds.
The Export-Control Wall and Its Limits
The formal effort to deny China access to cutting-edge AI hardware began when the Bureau of Industry and Security issued an advanced computing rule in October 2022, codified in the Federal Register as 87 FR 62186. That regulation was the first modern U.S. export-control framework to explicitly target advanced computing integrated circuits, supercomputers, and semiconductor manufacturing end uses in connection with Chinese military modernization and AI-enabled surveillance. By tying licensing decisions to performance metrics like interconnect bandwidth and computing power, the rule tried to draw a line between legitimate commercial chip sales and national security risks, effectively turning high-end GPUs into controlled strategic goods.
The policy has not stayed static. On January 15, 2025, BIS published an Interim Final Rule known as the Framework for Artificial Intelligence Diffusion, which attempted to move beyond hardware chokepoints and address the global spread of advanced AI capabilities themselves, including tighter treatment of certain “Tier 2” countries. Yet the agency later walked that approach back, with a BIS announcement rescinding the rule while promising a replacement focused on strengthening chip-related controls. The reversal underscores how unsettled the policy debate remains: officials are torn between pushing controls deeper into the AI stack and avoiding overreach that could damage U.S. chipmakers’ global market share and innovation incentives.
Stealing the Blueprints: Espionage Inside Big Tech
Export controls can slow the flow of physical chips, but they cannot easily stop a well-placed insider from walking out with designs. Federal prosecutors allege that is what happened between May 2022 and May 2023, when Linwei “Leon” Ding, a Chinese national employed as a software engineer at Google, allegedly copied proprietary files from the company’s internal systems to personal accounts. A superseding indictment filed in the Northern District of California claims Ding exfiltrated more than 1,000 files related to Google’s AI infrastructure, including information about TPU and GPU cluster orchestration, data center scheduling, and custom networking components.
According to the Department of Justice’s public statement on the case, the alleged trade secrets cover the kind of infrastructure know-how that enables training and deployment of large-scale AI models, rather than any single algorithm or model artifact. That distinction matters. If such “plumbing” blueprints reach a foreign competitor backed by state resources, the payoff is not merely a cloned chatbot but the ability to replicate an entire industrial-scale training ecosystem. Institutional knowledge of how to knit together accelerators, interconnects, and software stacks at hyperscale is far harder to replace than a diverted shipment of chips, yet current policy is still geared primarily toward border controls. The Ding indictment highlights a blind spot: safeguarding AI leadership now requires robust corporate security and insider-threat programs that complement government export rules.
Smuggling Chips Through Southeast Asia
Where insiders target blueprints, smugglers go after the hardware itself. In a separate case, U.S. citizens and Chinese nationals were charged in an alleged scheme to move restricted GPUs into China, in violation of export laws governing advanced AI technology. The Justice Department alleges that the defendants illicitly procured and shipped controlled NVIDIA accelerators, including A100 units, and attempted to export newer H100 and H200 systems as well, all while concealing the true end users. According to the criminal complaint, the network relied on transshipment through Malaysia and Thailand, using falsified paperwork, sham contracts, and front companies to disguise that the ultimate destination was the People’s Republic of China.
The alleged operation illustrates a structural weakness in the enforcement architecture. Even as Washington tightens performance thresholds and blacklists specific entities, the global supply chain for high-end GPUs runs through a web of distributors, integrators, and logistics hubs with uneven compliance cultures. A shell company in Kuala Lumpur or Bangkok can submit an order that appears routine, then quietly reroute the hardware to a Chinese data center once the boxes clear customs. The inclusion of both legacy A100 and next-generation H100 and H200 products in the indictment suggests that demand for the most capable accelerators has not abated in the face of stricter rules. If anything, rising model sizes and training budgets increase the economic incentive to exploit gray-market channels, challenging U.S. agencies to build deeper partnerships with foreign customs authorities and to monitor secondary markets more aggressively.
DeepSeek, NVIDIA, and the Congressional Spotlight
As law enforcement pursues individual smugglers and suspected spies, lawmakers have begun to scrutinize how U.S. technology companies manage their own exposure to Chinese AI firms. One focal point is DeepSeek, a Chinese developer of large language models that has drawn attention in Washington for its rapid technical progress and opaque ownership structure. The House Select Committee on the Strategic Competition Between the United States and the Chinese Communist Party has circulated findings suggesting that DeepSeek’s training runs may rely heavily on U.S.-designed accelerators, raising questions about how effectively export controls are constraining advanced AI development in China. While the committee’s report is not itself an enforcement action, it signals growing congressional interest in mapping the full ecosystem of Chinese AI labs, cloud providers, and hardware suppliers.
NVIDIA, whose GPUs sit at the center of most large-scale AI training clusters worldwide, has acknowledged in its regulatory filings that U.S. export controls pose a material business risk. In its most recent annual report to the Securities and Exchange Commission, the company warned that evolving restrictions on sales of advanced accelerators to China could affect revenue, supply chain planning, and product design choices, particularly for data center products tailored to AI workloads. The filing notes that tighter licensing requirements and performance-based thresholds have already forced NVIDIA to adjust certain offerings for the Chinese market. By flagging export rules as a significant risk factor in its SEC disclosure, NVIDIA underscores how deeply national security policy is now intertwined with the commercial trajectory of the AI hardware industry.
Can Washington Really Contain AI?
These episodes collectively raise a hard question: can any one country realistically contain the diffusion of advanced AI capabilities through export controls alone? The Ding indictment shows that proprietary know-how can be siphoned off from inside leading firms, bypassing border checks entirely. The GPU smuggling case demonstrates how determined actors can exploit jurisdictional seams and transshipment hubs to move restricted hardware despite formal bans. Congressional scrutiny of DeepSeek and other Chinese AI players, meanwhile, hints at a sprawling ecosystem in which U.S.-origin chips, cloud services, and software frameworks may still play a central role, even when direct exports are curtailed.
For U.S. policymakers, the implication is that maintaining an AI edge will require a layered strategy rather than a single “wall.” That means pairing hardware controls with stronger corporate security practices, more rigorous vetting of research partnerships, and closer coordination with allies whose ports and distributors sit along key supply routes. It also means accepting trade-offs: overly broad or unstable rules, like the short-lived AI diffusion framework, risk undermining the very innovation base that gives the United States leverage. As adversaries probe for weaknesses (from insider access at tech giants to shell companies in Southeast Asia), Washington’s challenge is to adapt its tools quickly enough to matter, without choking off the global collaborations and markets that fuel American AI leadership. How effectively it balances those competing imperatives will shape not just the security of today’s models and chips, but the trajectory of the next decade of AI development.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
