How U.S. Chip Export Controls Work and Why They Shape the AI Race
From Entity List designations to end-use verification, the architecture of semiconductor export restrictions defines the boundary between American technological leverage and China's path to self-sufficiency.
U.S. export controls on semiconductors are the primary tool Washington uses to limit China’s access to advanced chips, restricting sales of AI accelerators, lithography equipment, and manufacturing technology through a legal framework administered by the Commerce Department’s Bureau of Industry and Security.
The current restrictions, tightened significantly in October 2022 and expanded in 2023-2024, represent the most aggressive use of Export Controls since the Cold War. They target not just specific companies but entire technological capabilities, creating what policymakers describe as a small yard, high fence around cutting-edge Semiconductors and the tools needed to make them. Understanding how this system works—what triggers restrictions, who enforces them, and how companies navigate compliance—is essential to following the semiconductor decoupling that now defines U.S.-China competition in artificial intelligence.
The Legal Foundation: Export Administration Regulations
Export controls rest on the Export Administration Regulations (EAR), a set of rules governing what products can be sold to whom and under what conditions. The commerce department’s Bureau of Industry and Security (BIS) maintains the Entity List, a roster of foreign companies, research institutions, and government organisations deemed national security risks. Any U.S. company—or foreign company using U.S.-origin technology—must obtain a license to sell controlled items to Entity List members, and these licenses are typically denied.
Products are classified by Export Control Classification Numbers (ECCNs), which determine whether an item requires a license for export. Advanced semiconductors fall under ECCN 3A090, which since October 2022 has included chips exceeding specific performance thresholds: logic chips above 600 mm² die size with computational density above 5.92, or memory bandwidth exceeding 600 GB/s. This threshold effectively captures Nvidia’s A100 and H100 GPUs, AMD’s MI250 accelerators, and similar AI training chips.
The October 2022 rules introduced a ‘foreign direct product rule’ (FDPR) extension, arguably the most consequential shift. Under FDPR, even chips made entirely outside the United States using non-U.S. equipment fall under U.S. jurisdiction if they were produced using any U.S. technology or software. This allows Washington to restrict sales by Taiwan’s TSMC, South Korea’s Samsung, and other foreign foundries to Chinese customers—even when no U.S. components cross borders. The rule’s extraterritorial reach sparked protests from allied governments but has proven enforceable because virtually all advanced chip manufacturing relies on American design software from Synopsys and Cadence.
The Entity List Mechanism
Entity List designation follows a multi-step process. BIS receives intelligence—from U.S. agencies, allied governments, or compliance violations flagged by companies—indicating a foreign entity poses national security concerns. Common triggers include ties to China’s military-civil fusion programs, participation in surveillance technology development, or attempts to evade existing controls. The agency then conducts a review, often consulting the Departments of State, Defense, and Energy.
Once listed, an entity faces a presumption of denial for license applications. Between October 2022 and March 2026, BIS added more than 490 Chinese entities, according to a Center for Strategic and International Studies analysis. High-profile additions include Huawei (May 2019), SMIC (December 2020), and AI chip designer Biren Technology (October 2022). Huawei’s designation alone triggered a cascade effect: its suppliers, from ARM to TSMC, were forced to sever ties, ultimately crippling the company’s smartphone business and pushing it toward domestic alternatives.
Removal from the Entity List is rare but possible. It requires demonstrating changed behaviour—typically cooperation with U.S. authorities, restructuring to eliminate problematic activities, or evidence the original designation was erroneous. No major Chinese semiconductor firm has successfully secured removal since 2019.
Enforcement and Compliance Architecture
BIS lacks the resources to police every shipment. The agency employs roughly 400 export enforcement agents, a figure unchanged since 2015 despite exponential growth in controlled technology trade. This creates reliance on corporate compliance programs. U.S. and allied semiconductor firms maintain internal export control teams, using screening software to cross-check customer lists against Entity List updates published weekly.
When violations occur, penalties are severe. In 2023, Seagate Technology paid $300 million to settle charges it shipped hard drives to Huawei after the 2019 designation. The fine, one of the largest in export control history, sent a deterrent signal: even passive non-compliance—continuing existing contracts without active smuggling—triggers consequences.
Enforcement increasingly targets ‘transhipment hubs’ where goods are relabelled to disguise final destinations. Hong Kong, Singapore, and Malaysia have seen rising BIS activity. The agency now requires ‘end-use statements’ for sensitive shipments, obligating buyers to certify products will not be resold to restricted parties. Companies shipping to distributors in these jurisdictions face enhanced scrutiny, including on-site verification visits.
The U.S. is not alone in export controls—32 nations coordinate restrictions through the Wassenaar Arrangement, a multilateral regime governing conventional arms and dual-use technology. However, Washington’s semiconductor controls go further than Wassenaar commitments, creating tension with allies who see unilateral U.S. action as undermining their commercial interests. The Netherlands’ reluctant agreement to restrict ASML lithography exports reflects this friction.
Workarounds and the Cat-and-Mouse Dynamic
Export controls are not static barriers but contested boundaries. Chinese firms have developed three primary workaround strategies. First, ‘design-around’—engineering chips that stay just below performance thresholds. Nvidia’s H800 and A800 chips, variants of the H100 and A100 with reduced chip-to-chip interconnect speeds, briefly satisfied this approach until BIS closed the loophole in October 2023.
Second, procurement via shell companies and third-country intermediaries. U.S. authorities have documented networks routing chips through Dubai, Taiwan, and Southeast Asia before arriving in China. A New York Times investigation in March 2024 found that restricted Nvidia chips were openly listed on Chinese e-commerce platforms, often at 2-3x markup, indicating active grey markets.
Third, accelerated domestic substitution. Restrictions have created captive demand for Chinese alternatives. Companies like Biren Technology, Cambricon, and Huawei’s Ascend division now supply AI accelerators that, while lagging U.S. performance by 1-2 generations, meet many domestic use cases. This dynamic creates a paradox: export controls slow China’s access to cutting-edge chips but accelerate the development of indigenous alternatives that may eventually erode U.S. market share globally.
| Specification | Nvidia H100 (Restricted) | Nvidia H800 (Compliant, 2023) | Huawei Ascend 910B (Domestic) |
|---|---|---|---|
| FP16 Performance | 2000 TFLOPS | 1979 TFLOPS | ~400 TFLOPS |
| Memory Bandwidth | 3352 GB/s | 3352 GB/s | 1200 GB/s |
| Chip-to-Chip Interconnect | 900 GB/s | 400 GB/s | ~300 GB/s |
| Manufacturing Node | TSMC 4nm | TSMC 4nm | SMIC 7nm |
The Geopolitical Stakes
Export controls are not merely trade policy—they are strategic instruments designed to maintain U.S. advantage in AI, quantum computing, and next-generation military systems. The logic is straightforward: advanced chips enable advanced AI, and advanced AI underpins everything from autonomous weapons to signals intelligence. By denying China access to frontier compute, Washington aims to preserve a technological lead that translates into military and economic power.
However, effectiveness is contested. A Center for a New American Security report in 2025 argued controls have bought 3-5 years of delay in China’s AI development, particularly for large language models requiring massive GPU clusters. Critics counter that controls accelerate Chinese self-sufficiency, potentially creating a bifurcated technology ecosystem where U.S. firms lose the world’s largest semiconductor market permanently.
The controls also impose costs on U.S. companies. Nvidia derived roughly 20% of revenue from China before October 2022 restrictions took effect, a market now largely inaccessible except through compliant chip variants. AMD, Intel, and smaller fabless designers face similar pressures. The long-term calculation remains uncertain: whether short-term revenue loss is offset by preserved competitive advantage, or whether restrictions ultimately hollow out U.S. market dominance by ceding ground to Chinese alternatives.