AI Chip Export Controls
US AI Chip Export Controls
Quick Assessment
Section titled “Quick Assessment”| Dimension | Assessment | Evidence |
|---|---|---|
| Effectiveness | Moderate (1-3 year delay) | Chinese AI labs report hardware shortages; NVIDIA China revenue dropped from 21% to 13% of total |
| Enforcement | Weak | Approximately 140,000 GPUs smuggled to China in 2024; BIS has only one officer for all of Southeast Asia |
| Durability | Uncertain | China’s $47.5B Big Fund III accelerating domestic alternatives; SMIC producing 7nm chips via multi-patterning |
| Evasion Rate | High | Third-country transshipment via Singapore, Malaysia; NVIDIA Singapore revenue rose from under $1B to nearly $8B quarterly |
| Allied Cooperation | Partial | Netherlands (ASML EUV) and Japan cooperating; South Korea maintaining economic ties with China |
| Safety Relevance | Low-Medium | Primarily serves geopolitical goals; does not address risks from US-developed AI systems |
| Unintended Effects | Significant | DeepSeek achieved GPT-4 parity at 1/10th compute; forced efficiency innovations may accelerate capability development |
Overview
Section titled “Overview”The United States has implemented unprecedented export controls on advanced semiconductors and semiconductor manufacturing equipment, primarily targeting China’s access to AI-enabling hardware. These controls represent the most significant attempt to constrain AI development through hardware governance, with the Biden administration treating semiconductor exports as a national security priority comparable to nuclear technology controls during the Cold War.
The fundamental logic is straightforward: control access to advanced chips, thereby constraining compute availability, which in turn limits the ability to train and deploy frontier AI systems. Unlike other compute governance approaches that focus on monitoring or triggering regulatory requirements, export controls are inherently restrictive—they aim to completely deny access to certain actors rather than regulate usage. This makes them a blunt but potentially powerful instrument for shaping the global AI landscape.
However, the effectiveness and safety implications of these controls remain hotly debated. While they have demonstrably disrupted Chinese AI development in the near term, creating chip shortages and forcing workarounds, they have also accelerated Chinese efforts toward semiconductor self-sufficiency and strained international cooperation on AI safety. The controls primarily serve geopolitical objectives rather than addressing core AI safety challenges, raising questions about whether they ultimately enhance or undermine global AI governance efforts.
Export Control Flow and Enforcement
Section titled “Export Control Flow and Enforcement”Scope and Mechanisms
Section titled “Scope and Mechanisms”Controlled Technologies
Section titled “Controlled Technologies”The export control regime targets three critical components of the AI hardware stack. Advanced logic chips form the core restriction, encompassing NVIDIA’s A100, H100, and H200 series, AMD’s MI250 and MI300 series, and any chip exceeding specific performance thresholds—currently 4,800 TOPS for INT8 operations or 300 TFLOPS for FP16 operations. These thresholds are designed to capture chips capable of training large language models while theoretically allowing less capable hardware for commercial applications.
Semiconductor manufacturing equipment represents the deeper strategic target, as it addresses the production capability rather than just the end products. This includes extreme ultraviolet (EUV) lithography machines exclusively produced by Netherlands-based ASML, deep ultraviolet (DUV) immersion lithography equipment for advanced processes, and specialized deposition and etching tools required for sub-10nm chip fabrication. The equipment controls are potentially more durable than chip controls since manufacturing capabilities are harder to replicate than individual products.
High Bandwidth Memory (HBM) has emerged as a critical bottleneck, as these specialized memory modules are essential for AI accelerators but have limited global production capacity concentrated in South Korean companies like SK Hynix and Samsung. The December 2024 control expansion specifically targeted HBM exports, recognizing that memory bandwidth often determines AI training performance more than raw compute power.
Enforcement Architecture
Section titled “Enforcement Architecture”The Bureau of Industry and Security (BIS) operates the control system through the Export Administration Regulations (EAR), which apply not only to US-origin items but also to foreign-produced items containing US components above specified thresholds—typically 25% for most technologies but as low as 0% for the most sensitive items. This extraterritorial reach is enabled by the deep integration of US technology in global semiconductor supply chains, from design software (Synopsys, Cadence) to manufacturing equipment components.
The Entity List mechanism requires licenses for any exports to listed Chinese companies, with most applications facing a “presumption of denial.” As of late 2024, over 600 Chinese entities are subject to these restrictions, including major AI companies like SenseTime, Megvii, and compute infrastructure providers. Violations carry substantial penalties: civil fines up to $300,000 per violation and criminal penalties including up to 20 years imprisonment for willful violations.
Implementation Timeline and Evolution
Section titled “Implementation Timeline and Evolution”Control Rounds and Key Provisions
Section titled “Control Rounds and Key Provisions”| Date | Action | Key Provisions | Immediate Impact |
|---|---|---|---|
| October 2022 | Initial controls | A100/H100 restrictions; “US person” rule; performance thresholds | NVIDIA created A800/H800 variants for China; workforce disruptions |
| October 2023 | First expansion | Lower thresholds; A800/H800 banned; Iran/Russia added | Closed “China-compliant chip” loophole; additional entity listings |
| December 2024 | Major update | 140+ entity additions; HBM controls; node-agnostic tools | First country-wide HBM restrictions; 24 new SME categories |
| January 2025 | AI Diffusion Rule | Global performance thresholds; “green zone” for allies | H100/H200 blocked globally except for approved countries |
| April 2025 | Trump freeze | All China-bound AI chips halted; compliance review | $5.5B charge for NVIDIA on H20 inventory |
| August 2025 | Partial reversal | Conditional Nvidia/AMD sales; 15% revenue to government | H20 sales resumed under revenue-sharing arrangement |
| December 2025 | Further loosening | H200 approved for China export | Most powerful chip ever approved for China |
The October 2022 initial rules marked a dramatic escalation from previous technology controls, implementing broad restrictions on AI chips and introducing the controversial “US person” rule prohibiting American citizens and permanent residents from supporting Chinese semiconductor development. These initial controls targeted chips with performance characteristics similar to NVIDIA’s A100, effectively cutting off Chinese access to the most advanced AI hardware. According to the Congressional Research Service↗, the rules forced US citizens and green card holders to leave China’s semiconductor industry effective October 12, 2022.
The October 2023 expansion significantly tightened restrictions by lowering performance thresholds, extending controls to additional countries including Iran and Russia, and adding more semiconductor manufacturing equipment to the restricted list. This update responded to observed evasion attempts and closed loopholes that had allowed continued access through third-party suppliers and modified chip variants.
The December 2024 update marked another escalation, adding over 140 Chinese entities to restriction lists, implementing new controls on HBM, and tightening restrictions on quantum computing technologies. As CSIS analysis↗ noted, Chinese chipmakers had acquired a “huge stockpile” of equipment between October 2022 and December 2024, significantly reducing the immediate impact of new rules.
Measurable Impacts and Evidence
Section titled “Measurable Impacts and Evidence”Documented Disruptions
Section titled “Documented Disruptions”Chinese AI laboratories and companies have publicly reported significant disruptions to their hardware procurement. ByteDance suspended sales of its cloud AI services to external customers, citing chip shortage concerns. Alibaba Cloud implemented usage restrictions on its GPU instances, prioritizing internal applications over external customers. Chinese universities have reported delays in AI research projects due to equipment unavailability, with some researchers shifting to less compute-intensive approaches.
| Metric | Pre-Controls | Post-Controls | Change |
|---|---|---|---|
| NVIDIA China data center revenue share | 21% (FY2023) | 13% (FY2025) | -38% relative decline |
| NVIDIA China-compliant GPU market share | 95% | 0% (for restricted chips) | Complete exclusion from high-end |
| Huawei Ascend chip shipments (2024) | N/A | 200,000 units | Versus NVIDIA’s 1 million H20s to China |
| China’s share of ASML revenue | 49% (mid-2024) | ~20% (end-2025) | Normalization after stockpiling |
| Secondary market premium for restricted chips | 1x | 2-3x list price | Significant supply-demand imbalance |
According to Fortune Asia↗, NVIDIA’s China sales dropped to a “mid-single-digit percentage” of data center revenue by early 2024. CEO Jensen Huang stated that NVIDIA’s market share for restricted AI GPUs in China fell from 95% to “zero” following export controls.
Chinese technology companies have publicly acknowledged the impact while announcing alternative strategies. Baidu has increased investment in its domestic Kunlun chip development, while Alibaba has accelerated development of its Yitian processors. China launched Big Fund III↗ in May 2024 with $47.5 billion in registered capital—the largest phase of China’s semiconductor investment program—bringing total state-backed investment across all three phases to approximately $188 billion.
Circumvention and Adaptation
Section titled “Circumvention and Adaptation”Despite these impacts, substantial circumvention activity has been documented. According to CNAS↗, an estimated 140,000 high-performance GPUs worth billions of dollars were smuggled into China in 2024 alone. The Information↗ identified at least eight Chinese AI chip-smuggling networks, each engaged in transactions valued at more than $100 million.
| Smuggling Indicator | Evidence | Scale |
|---|---|---|
| Third-country transshipment | NVIDIA Singapore revenue exploded from less than $1B (Q4 2022) to nearly $8B (Q3 2024) | 8x increase |
| Large-scale networks | Single orders for 2,400 H100s ($120M) documented | $100M+ per network |
| Enforcement actions | Singapore arrested 3 individuals for $390M diversion | Hundreds of millions seized |
| DOJ prosecutions | October 2024-May 2025: $160M in attempted exports | $50M reached China before disruption |
| Tactics used | Duplicated serial numbers, hidden in prosthetic baby bumps, labeled as “tea” or “toys” | Increasingly sophisticated |
Cloud computing workarounds present a more sophisticated challenge, as Chinese entities can access restricted hardware through US cloud providers’ international services. While major providers have implemented compliance measures, smaller providers and reseller arrangements create ongoing vulnerabilities. Some Chinese companies have established subsidiaries in third countries specifically to access controlled technologies legally.
The stockpiling phenomenon created its own complications. As CSIS reported↗, Chinese companies dramatically increased purchases between announcements and implementation. In the case of HBM, “massive Chinese stockpiling efforts had already begun by early August 2024” after reports that controls were likely. ASML’s China revenue hit 49% of total revenue in mid-2024—an abnormally high figure driven by pre-restriction purchasing—before normalizing to approximately 20% by end of 2025.
Strategic Effectiveness Assessment
Section titled “Strategic Effectiveness Assessment”Time Delays and Technical Impact
Section titled “Time Delays and Technical Impact”Expert assessments of the time delay imposed by export controls range widely, from 6 months to 5 years depending on the specific capability and Chinese response speed. For training models at the scale of GPT-4, most analysis suggests a 1-3 year delay based on reduced hardware access and the need to develop workarounds. However, this estimate assumes continued advancement in US capabilities during the delay period rather than static competition.
Semiconductor manufacturing equipment controls likely impose longer delays than chip controls, as building advanced fabrication capabilities requires 3-7 years even with unlimited resources and access to foreign talent. CSIS analysis↗ notes that SMIC has struggled to achieve consistent 7nm production yields comparable to TSMC’s mature processes. According to American Affairs Journal↗, SMIC uses “double patterning” with DUV equipment to achieve 7nm, but this process is “time consuming and expensive, making it difficult to manufacture at volume scale.” It is “increasingly clear that SMIC will not be able to get to something that can be called a 3 nanometer process using its existing DUV tools.”
The effectiveness varies significantly by application area. Controls appear most effective against large-scale model training requiring thousands of high-end chips, moderately effective against inference applications that can utilize lower-performance hardware, and least effective against research applications that can often proceed with limited hardware resources or cloud access.
The DeepSeek Paradox
Section titled “The DeepSeek Paradox”The case of DeepSeek↗ illustrates both the limitations and unintended consequences of export controls. Despite hardware restrictions, DeepSeek trained models achieving GPT-4 parity using:
- Less-advanced H800 chips (allowed until October 2023) performing “similarly to H100s” for their workloads
- Efficiency innovations reducing memory usage and accelerating calculations without sacrificing accuracy
- Dramatically lower costs: DeepSeek claims V3 training cost $6 million versus $100 million+ for GPT-4, using approximately one-tenth the compute of Meta’s Llama 3.1
As Brookings↗ observed, DeepSeek’s success is “a stark illustration of why U.S. export controls on advanced computing chips, instead of impeding China’s AI progress, may actually be accelerating it” by forcing efficiency innovations. When DeepSeek-R1 launched on January 20, 2025, it surpassed ChatGPT as the most downloaded iOS app in the United States within a week, triggering an 18% drop in NVIDIA’s share price.
Unintended Consequences
Section titled “Unintended Consequences”Export controls have accelerated Chinese investment in domestic semiconductor capabilities, with the 2024 national semiconductor fund representing the largest government technology investment in Chinese history. This “forced innovation” effect could ultimately result in a more capable and independent Chinese semiconductor industry, reducing long-term American leverage. Historical precedents like Japanese semiconductor development following US trade restrictions in the 1980s suggest this outcome is plausible.
The controls have also strained relationships with key allies whose companies face compliance costs and lost revenue. Dutch and Japanese semiconductor equipment manufacturers initially resisted US pressure for aligned restrictions, requiring sustained diplomatic engagement and potential compensation for lost sales. South Korean memory manufacturers face particular tensions given their economic ties to China and strategic relationships with the United States.
International standardization and research collaboration have been complicated by export controls, as Chinese researchers and companies face exclusion from conferences, joint projects, and technical working groups. This fragmentation of the global AI research community could ultimately slow overall progress on safety research and beneficial applications.
Safety Implications Analysis
Section titled “Safety Implications Analysis”Arguments for Safety Benefits
Section titled “Arguments for Safety Benefits”Proponents argue that concentrating advanced AI capabilities within democratic societies provides better governance prospects than diffused global access. Democratic institutions, civil society oversight, and legal frameworks in the US and allied countries may be better positioned to implement AI safety measures than authoritarian systems optimized for rapid deployment regardless of risks. The argument suggests that even a temporary advantage for democratic AI development could be crucial if alignment breakthroughs occur during the window of opportunity.
The time delay argument emphasizes that even short-term delays in Chinese AI development could provide crucial breathing room for safety research and governance development. If AI poses substantial risks that require careful management, slowing the overall pace of the global AI race could be beneficial regardless of which specific actors are delayed. This perspective treats export controls as buying time for the global community to develop better safety measures.
Controls may also prevent proliferation of advanced AI capabilities to non-state actors or adversarial regimes that could pose misuse risks. By constraining the hardware ecosystem, export controls theoretically limit the ability of malicious actors to acquire dangerous AI capabilities through commercial channels or theft of technology.
Arguments Against Safety Benefits
Section titled “Arguments Against Safety Benefits”Critics argue that export controls primarily represent geopolitical competition rather than genuine safety measures, noting that they do nothing to address risks from US-developed AI systems that may pose equally significant challenges. If the primary safety concerns involve rapid capability advancement or insufficient safety research, slowing only Chinese development while US companies continue racing may provide minimal benefit.
The cooperation disruption argument suggests that effective AI governance requires international coordination, particularly between the US and China as the leading AI powers. Export controls fundamentally undermine the trust and technical exchange necessary for coordinated governance measures, potentially making global AI safety coordination more difficult precisely when it becomes most important.
Some analysts worry that export controls could accelerate rather than slow dangerous AI development by encouraging rushed deployment before additional restrictions take effect. Chinese companies may face incentives to quickly operationalize AI systems using existing hardware rather than investing in careful safety research, potentially increasing near-term risks.
International Coordination Challenges
Section titled “International Coordination Challenges”Allied Participation
Section titled “Allied Participation”| Country | Key Companies | Cooperation Level | Key Constraints |
|---|---|---|---|
| Netherlands | ASML | High | EUV banned; DUV licensing required since September 2024 |
| Japan | Tokyo Electron, Nikon | High | Security alliance; equipment restrictions aligned with US |
| South Korea | Samsung, SK Hynix | Partial | HBM controls strain China economic ties; fab investments in China |
| Taiwan | TSMC | High | Sub-7nm exports to China halted November 2024 |
| Germany | Zeiss, TRUMPF | Moderate | ASML component suppliers; some reluctance |
Dutch participation through ASML’s EUV restrictions represents the most critical international cooperation, as no alternative suppliers exist for the most advanced lithography equipment. According to TrendForce↗, the Netherlands expanded controls on September 7, 2024 to require licenses for ASML’s 1970i and 1980i DUV machines—with the Dutch government taking “full control” rather than deferring to US authority. ASML must now also apply for licenses to service, provide spare parts, and software updates for previously sold immersion lithography systems.
Japanese cooperation through major semiconductor equipment manufacturers like Tokyo Electron has been similarly crucial but required sustained diplomatic engagement. Japan’s participation was facilitated by its security alliance with the United States and concerns about Chinese military capabilities, but Japanese companies continue to seek narrow interpretations of restrictions to minimize commercial impact.
South Korean participation remains incomplete and problematic, as Samsung and SK Hynix dominate global HBM production essential for AI systems. The December 2024 controls extended HBM restrictions to South Korean firms operating in China—creating significant compliance costs for Samsung’s Xi’an NAND flash facility and SK Hynix’s Dalian operations. Taiwan’s TSMC was directed by the Commerce Department↗ to halt advanced sub-7nm chip exports to China starting November 11, 2024.
Limits of Extraterritorial Reach
Section titled “Limits of Extraterritorial Reach”While US technology integration gives American export controls substantial global reach, technological sovereignty initiatives in various countries aim to reduce this dependence. China’s domestic semiconductor development, while currently unsuccessful, represents a long-term challenge to US control effectiveness. European digital sovereignty initiatives and investment in independent technology stacks could similarly reduce American leverage over time.
Alternative supply chains are emerging in areas where US technology dominance is less complete. Chinese companies have increasingly turned to domestic alternatives for design software, manufacturing equipment, and materials where technically feasible. According to CNAS↗, Chinese startup Yuliangsheng has reportedly developed a DUV immersion tool for AI-capable chips that is being tested by SMIC. While current alternatives are typically inferior, sustained investment and improvement could reduce control effectiveness over multi-year timeframes.
The emergence of new technology paradigms could also limit control effectiveness. Quantum computing development, neuromorphic chips, and other alternative computing approaches may not depend on the same semiconductor technologies currently controlled, potentially creating new pathways for advanced AI development outside existing control regimes.
Enforcement Resource Constraints
Section titled “Enforcement Resource Constraints”The CSIS Mismatch report↗ highlights a fundamental gap between strategic ambitions and enforcement capacity:
| Resource | Current Level | Adequacy |
|---|---|---|
| BIS total staff | Fewer than 600 employees | Overseeing trillions in dual-use exports |
| BIS budget | Under $200 million | Policing global smuggling networks |
| Southeast Asia officers | 1 officer | Covering region with primary transshipment routes |
| Red flags added (2024) | 8 new indicators | Enhanced but still reactive |
As CSIS notes, “BIS’s limited in-region capacity—there is only a single export-control officer assigned to cover all of Southeast Asia—stands in stark contrast to the billion-dollar smuggling operations it seeks to monitor and disrupt.” The December 2024 controls added a new “knowledge standard” requiring companies to demonstrate due diligence, but this places the burden on exporters rather than expanding government enforcement capacity.
Current Status and Trajectory
Section titled “Current Status and Trajectory”Near-Term Outlook (1-2 Years)
Section titled “Near-Term Outlook (1-2 Years)”Export controls are likely to become more comprehensive and stringent as the Biden administration and Congress view them as a successful policy tool. Additional countries may be added to restrictions, performance thresholds may be lowered further, and new categories of dual-use technology may be controlled. The December 2024 expansion suggests this trajectory will continue regardless of specific Chinese responses.
Chinese circumvention efforts will likely become more sophisticated, with increased use of third-country intermediaries, cloud workarounds, and alternative supply chains. However, the scale of circumvention is unlikely to fully offset the restrictions given the concentrated nature of advanced semiconductor production and US supply chain integration.
International coordination may face increasing strain as economic costs mount for allied countries’ companies. European and Asian governments may seek to limit cooperation if domestic industry impacts become severe, particularly if Chinese retaliation targets their exports or investments.
Medium-Term Projection (2-5 Years)
Section titled “Medium-Term Projection (2-5 Years)”Chinese domestic semiconductor capabilities will likely improve substantially, though probably not to the point of matching the most advanced international production. SMIC and other Chinese foundries may achieve consistent 7nm production and begin 5nm development, reducing but not eliminating dependence on foreign suppliers. Chinese design capabilities and chip architectures optimized for AI workloads may also advance significantly.
The effectiveness of current controls will likely diminish as Chinese alternatives mature and circumvention networks become more established. However, the most advanced capabilities—particularly those requiring cutting-edge manufacturing processes and specialized equipment—may remain restricted for longer periods.
Alternative approaches to AI development that require less advanced hardware may gain prominence, including more efficient algorithms, specialized chip architectures, and novel computing paradigms. These developments could reduce the overall importance of the specific technologies currently controlled while maintaining significant capability gaps.
Long-Term Uncertainties
Section titled “Long-Term Uncertainties”The ultimate trajectory depends heavily on Chinese success in developing independent semiconductor capabilities, which remains highly uncertain despite massive investment. Semiconductor manufacturing involves extraordinarily complex supply chains and tacit knowledge that may be difficult to replicate quickly, but historical examples of technology transfer and indigenous innovation suggest it is achievable over longer time horizons.
Geopolitical developments could dramatically alter the export control landscape, from potential negotiated limitations on AI development to escalating technology competition that extends controls to additional sectors. The intersection of AI capabilities with military applications may drive further restrictions regardless of safety considerations.
Technological breakthroughs in either AI efficiency or alternative computing approaches could fundamentally change the strategic landscape, potentially making current export controls obsolete or requiring entirely new approaches to technology governance.
Key Uncertainties and Research Gaps
Section titled “Key Uncertainties and Research Gaps”Quantitative Impact Assessment
Section titled “Quantitative Impact Assessment”Reliable metrics for measuring export control effectiveness remain limited, as Chinese AI development occurs largely behind closed doors and companies have incentives to either overstate impacts (when seeking policy changes) or understate them (when minimizing strategic vulnerabilities). Independent assessment of Chinese AI capabilities relative to US systems is challenging but essential for evaluating control effectiveness.
The relationship between hardware constraints and AI capability development is complex and poorly understood. Different applications have varying hardware requirements, and algorithmic improvements can partially compensate for hardware limitations. Better models for predicting the AI capability impact of specific hardware restrictions would improve policy design and assessment.
Economic impact measurement faces similar challenges, as the costs and benefits of export controls extend far beyond direct semiconductor trade to include effects on innovation, alliance relationships, and long-term competitive positioning. Comprehensive cost-benefit analysis requires models that capture these broader implications.
Strategic and Safety Questions
Section titled “Strategic and Safety Questions”Whether the time purchased by export controls is being used productively for safety research and governance development remains an open question. If US AI companies continue rapid development and deployment without commensurate safety measures, the strategic value of slowing competitors may be limited from a safety perspective.
The net effect of export controls on global AI safety cooperation is unclear and likely depends on the specific mechanisms for international coordination that emerge. Controls may undermine cooperation by reducing trust and technical exchange, or they may enhance cooperation by maintaining US influence and providing incentives for negotiated agreements.
The durability of current restrictions depends on factors including Chinese technological advancement, allied cooperation, and the evolution of AI technology itself. Understanding these dependencies is crucial for long-term policy planning and assessment of current investments in export control infrastructure.
Alternative Policy Pathways
Section titled “Alternative Policy Pathways”The relationship between export controls and other AI governance approaches remains underexplored. Controls might complement monitoring and threshold-based regulations by providing enforcement mechanisms, or they might undermine such approaches by reducing the cooperation necessary for implementation.
The potential for negotiated alternatives to export controls, such as international AI development agreements or monitoring regimes, deserves more analysis. Understanding the conditions under which China and other countries might accept such alternatives could inform both near-term policy and long-term strategic planning.
How much do export controls contribute to AI safety?
Sources & References
Section titled “Sources & References”Government and Policy Documents
Section titled “Government and Policy Documents”- U.S. Export Controls and China: Advanced Semiconductors↗ — Congressional Research Service comprehensive analysis
- Commerce Strengthens Export Controls↗ — BIS December 2024 press release
- U.S. Authorities Shut Down Major China-Linked AI Tech Smuggling Network↗ — DOJ enforcement action
Think Tank Analysis
Section titled “Think Tank Analysis”- Understanding the Biden Administration’s Updated Export Controls↗ — CSIS December 2024 analysis by Gregory C. Allen
- The Limits of Chip Export Controls in Meeting the China Challenge↗ — CSIS effectiveness assessment
- Where the Chips Fall: U.S. Export Controls Under the Biden Administration↗ — CSIS comprehensive timeline
- Mismatch of Strategy and Budgets in AI Chip Export Controls↗ — CSIS enforcement capacity analysis
- Countering AI Chip Smuggling Has Become a National Security Priority↗ — CNAS smuggling assessment
- Technology to Secure the AI Chip Supply Chain↗ — CNAS hardware-enabled mechanisms proposal
- DeepSeek Shows the Limits of US Export Controls on AI Chips↗ — Brookings analysis
- China’s AI Chip Deficit: Why Huawei Can’t Catch Nvidia↗ — Council on Foreign Relations
Industry and Market Analysis
Section titled “Industry and Market Analysis”- How Chinese company DeepSeek released a top AI reasoning model despite US sanctions↗ — MIT Technology Review
- The Evolution of China’s Semiconductor Industry under U.S. Export Controls↗ — American Affairs Journal
- China Creates $47.5 Billion Chip Fund to Fuel Self-Resilience↗ — Bloomberg Big Fund III coverage
- Netherlands Expands Export Control over ASML’s DUV Machines↗ — TrendForce
- NVIDIA China Revenue Analysis↗ — Fortune Asia
Academic and Research Perspectives
Section titled “Academic and Research Perspectives”- DeepSeek, Huawei, Export Controls, and the Future of the U.S.-China AI Race↗ — CSIS
- The Rise of DeepSeek: What the Headlines Miss↗ — RAND Corporation
- China’s Quest for Semiconductor Self-Sufficiency↗ — Alan Turing Institute CETaS
Related Approaches
Section titled “Related Approaches”- Compute Thresholds — Regulatory triggers rather than restrictions
- Compute Monitoring — Visibility into who’s training what
- International Regimes — Multilateral coordination
AI Transition Model Context
Section titled “AI Transition Model Context”Export controls affect the Ai Transition Model through multiple factors:
| Factor | Parameter | Impact |
|---|---|---|
| Transition Turbulence | Racing Intensity | May reduce or intensify competition depending on response |
| Civilizational Competence | International Coordination | Strains cooperation needed for global AI governance |
| Transition Turbulence | AI Control Concentration | Concentrates frontier capabilities in US and allies |
Export controls primarily serve geopolitical objectives; their AI safety implications are contested and depend on whether time delays are used productively for safety research.