Compute Governance: AI Chips Export Controls Policy

Quick Assessment

Key Links

Overview

Compute governance represents a method for monitoring and regulating AI systems by tracking the hardware required to develop frontier models¹. In the context of AI chip export controlsPolicyUS AI Chip Export ControlsComprehensive empirical analysis finds US chip export controls provide 1-3 year delays on Chinese AI development but face severe enforcement gaps (140,000 GPUs smuggled in 2024, only 1 BIS officer ...Quality: 73/100, it encompasses U.S. government policies aimed at balancing technological leadership with national security by managing access to advanced semiconductors, particularly high-performance GPUs and AI-specific chips used for training large-scale AI systems.

The current policy framework emerged from concerns about China's AI capabilities and potential military applications. AI chips—including NVIDIA A100/H100 series, AMD MI300 series, and Google TPUs—are essential for frontier AI training, which requires tens of thousands of chips operating in parallel². Because these chips are detectable, excludable, and concentrated in U.S.-allied supply chains, they serve as a potential "chokepoint" for AI governance³.

The U.S. has shifted from blanket restrictions to a more calibrated approach. Effective January 15, 2026, the Bureau of Industry and Security published a final rule marking a significant departure from previous policy⁴. Rather than a presumption of denial for all advanced AI chips to China and Macau, the new framework evaluates license applications on a case-by-case basis, provided exports meet rigorous supply, security, and testing conditions⁴. This represents the latest development in an evolving policy landscape that began with initial restrictions in 2022.

History and Policy Evolution

Conceptual Origins

Compute governance as a framework emerged from AI safety and policy research communities around 2022-2023⁵. The approach builds on the recognition that physical hardware is more tractable for regulation than intangible assets like data or algorithms. Organizations like BlueDot Impact and GovAIOrganizationGovAIGovAI is an AI policy research organization with ~40-45 staff, funded primarily by Coefficient Giving ($1.8M+ in 2023-2024), that has trained 100+ governance researchers through fellowships and cur...Quality: 43/100 popularized the framework, emphasizing how AI chips' physical characteristics—detectability, excludability, and concentrated supply chains—make them suitable for regulatory intervention⁶.

The intellectual foundation came from research emphasizing that cutting-edge AI requires massive computational resources. A 2023 report from the Center for Security and Emerging TechnologyOrganizationGeorgetown CSETCSET is a $100M+ Georgetown center with 50+ staff conducting data-driven AI policy research, particularly on U.S.-China competition and export controls. The center conducts hundreds of annual gover...Quality: 43/100 (CSET) titled "AI Chips: What They Are and Why They Matter" highlighted export control opportunities via supply chain dominance⁷. GovAI's analysis "Computing Power and the Governance of AI" outlined the detectability and excludability of chips in data centers as key governance leverage points⁸.

Timeline of Key Policy Developments

Recent Administrative Changes

The Trump administration has implemented substantial revisions to the export control regime. In early 2026, the Department of Commerce rescinded the Biden-era AI Diffusion Rule while simultaneously strengthening certain semiconductor export controls¹⁴. The administration issued guidance warning against using Chinese chips like Huawei Ascend and cautioned about consequences for allowing U.S. chips to train Chinese AI models¹⁵.

A Presidential Proclamation formalized a 25% tariff on advanced AI chip imports intended for customers outside the United States¹³. This was accompanied by controversial revenue-sharing arrangements requiring companies like Nvidia to pay 15% of H20 chip sales and 25% of H200 chip sales to China to the U.S. government¹⁰.

Current Policy Framework

Core Requirements and Mechanisms

The January 2026 final rule establishes strict certification requirements that exporters and intended recipients must satisfy[^4]:

Performance Thresholds: The regulation permits export of chips with total processing performance (TPP) less than 21,000 or total DRAM bandwidth less than 6,500 GB/s—representing chips up to 13 times more powerful than previously allowed¹⁶.

Supply and Volume Constraints: Exports to China and Macau are capped at 50% of the number of chips shipped to U.S. customers for domestic use¹⁶. Exporters must provide attestations of sufficient U.S. supply through auditable data demonstrating no diversion from U.S. or allied users¹⁷.

Security and Testing: Advanced AI chips must undergo review by a qualified independent testing laboratory in the United States before each export shipment¹⁸. Intended recipients must adopt and demonstrate compliance frameworks to the U.S. government ensuring chips will not undermine U.S. security interests¹⁸.

Know Your Customer (KYC): Exporters must conduct due diligence to confirm chips will not be used to train AI models for weapons of mass destruction or military intelligence¹⁹. Location verification features in chip shipments are being explored to prevent illegal diversion to adversaries¹⁹.

Presumption of Denial Maintained: For China-owned data centers located outside China and for chips exceeding the established thresholds, the presumption of denial remains in effect¹⁶.

Legislative Proposals

Congress is considering several bills to strengthen or modify export controls[^20]:

GAIN AI Act (Guaranteeing Access and Innovation for National Artificial Intelligence): Would require chips to be prioritized for U.S. customers before export to certain countries. Includes a 15-day public notice period allowing U.S. buyers to express purchase interest²⁰.

Chip Security Act (introduced May 2025 by Sen. Tom Cotton, R-AR): Bipartisan legislation mandating "chip security mechanisms" such as export certificates, periodic location verification via server check-ins, and software-based authentication tools. Exporters would be required to report credible evidence of diversion, tampering, or unauthorized use²¹.

AI Overwatch Act (introduced December 2025 by Chairman Mast): Would require congressional review of export licenses for advanced AI chips to China, with a 30-day notification period before licenses may be issued—similar to oversight of military sales²².

STRIDE Act (Semiconductor Technology Resilience, Integrity, and Defense Enhancement): Would require State Department coordination with partner countries on semiconductor export controls²².

Strategic Objectives

The overarching U.S. policy aims to promote American export of "full-stack AI export packages"—hardware, models, software, applications, and standards—to allied nations while simultaneously tightening enforcement on American adversaries²³. The approach reflects a shift from the Biden Administration's AI Diffusion Rule toward sustaining and enhancing America's global AI dominance²⁴.

International Coordination and Allied Responses

The U.S. has sought to coordinate export controls with allies, though implementation varies significantly. The AI Action Plan calls for proactive, government-coordinated efforts to promote exports to allied nations while tightening enforcement against adversaries²⁰. Key allies including the EU, Netherlands, Germany, Japan, South Korea, and Taiwan have implemented list-based controls on AI chips and semiconductor manufacturing equipment²⁵.

However, allied authorities are generally narrower than U.S. capabilities under the Export Control Reform Act (ECRA) and Export Administration Regulations (EAR)²⁵. The Netherlands and Japan entered a trilateral deal with the U.S., though analysts note it is weaker than the U.S. regime, with ongoing challenges in aligning on technical thresholds like 16/14 nanometer logic chips²⁶.

A governance analysis recommends international harmonization of export controls, with the U.S. engaging partners to establish coordinated approaches²⁷. Strengthening institutional capacity requires U.S. attachés deployed to key transshipment hubs and information-sharing mechanisms with allies to disrupt diversion networks²¹.

Enforcement Challenges and Institutional Capacity

Resource Constraints

The Bureau of Industry and Security faces significant resource constraints in enforcing export controls²¹. Analysis reveals that "diversion of cutting-edge AI chips to China has exposed the limits of traditional export controls," with billions of dollars in illicit flows blunting the impact of 2022-2023 restrictions²¹.

BIS is under-resourced for monitoring global supply chains, lacking sufficient staff, tools, and presence in key transshipment hubs²¹. Recommended improvements include expanding BIS staffing, deploying supply chain analytics capabilities, monitoring end-use patterns across high-risk jurisdictions, and developing new export controls on semiconductor manufacturing sub-systems²⁸.

Implementation Difficulties

The onerous compliance requirements—including certifications regarding U.S. and global supply, end-user conditions, and provable enforceable controls—may be difficult for companies to meet²⁹. A January 14, 2026 letter from Representative Molenaar noted that the current global shortage of DRAM items poses challenges for meeting licensing requirements²⁹.

Critics argue that volume caps and the presumption of denial for China-owned data centers abroad are unenforceable without substantially enhanced monitoring capabilities³⁰. Loose implementation fails to achieve security goals, while strict implementation would effectively block most exports³⁰.

Criticisms and Controversies

Strategic Incoherence

The Council on Foreign Relations characterized the January 2026 regulation as "strategically incoherent," arguing it simultaneously acknowledges that exporting advanced AI chips to China poses serious national security risks while creating pathways to permit their sale³⁰. The policy authorizes "very large numbers of AI chip exports to China without enforceable guardrails that prevent their misuse," potentially allowing China's AI industry to close the gap with the United States and build the largest data centers in the world³⁰.

The revised thresholds permit chips 13 times more powerful than previously allowed, with critics contending this enables China's AI advancement and military applications despite weak compliance mechanisms³¹.

Legal Challenges

Legal scholars argue the revenue-sharing arrangement is illegal under the Export Control Reform Act (ECRA), which explicitly prohibits charging fees for export licenses³². The 25% tariff may also violate the Constitution's Export Clause, which prohibits taxes or duties on exports³². The revenue-sharing demands—requiring companies to pay 15-25% of China sales revenues to the U.S. government—lack congressional authorization and constitute unauthorized taxation³².

Companies and states potentially have standing to challenge these measures in court, as they exceed executive authority granted by Congress³².

Industry and Operational Concerns

Know Your Customer requirements demand sensitive information from non-U.S. data centers, potentially disrupting operations and innovation outside the United States²⁰. The compliance burden—including identity assurance, physical access restrictions, remote-access controls, foundry capacity protections, and third-party technical testing—creates significant operational challenges³³.

Some industry voices express concern that excessive regulation risks stifling nascent AI development, though this must be balanced against national security imperatives³⁴. The historical fear of "killing AI" through over-regulation persists in certain quarters³⁴.

Enforcement and Evasion

Traditional export controls have proven inadequate due to chip diversion via transshipment through third countries²¹. Infrastructure-as-a-Service (IaaS) providers—cloud computing services—offer a potential workaround to physical chip controls, as they enable indirect access to controlled computational resources³⁵. This suggests the need for diplomatic engagement with allies hosting major cloud providers to coordinate governance approaches³⁵.

Connection to AI Safety and Governance

The relationship between chip export controls and AI safety remains somewhat indirect in available policy analyses. The underlying governance logic relates to capability containment: restricting advanced chip access to potential adversaries aims to limit deployment of large-scale AI systems, which some analysts argue creates a critical window for maintaining technological leadership²⁶.

However, this window may be temporary. Research notes that hardware and algorithmic advances continuously reduce the number of chips needed per model capability, providing only a time-limited safety benefit³⁶. The efficacy of chip controls erodes as efficiency gains allow more powerful models to be trained with fewer resources³⁶.

Proposed mechanisms include international registries for AI chips, cryptographic enforcement on chips to restrict tasks or verify licenses, reporting requirements for large training runs, and "secure, governable chips" with hardened security modules to block unauthorized use³⁷. These on-chip governance mechanisms could theoretically support AI safety objectives, though implementation faces substantial technical and diplomatic challenges.

The search results frame export controls primarily as geopolitical and economic policy rather than as direct AI safety measures focused on misuse risksCruxAI Misuse Risk CruxesComprehensive analysis of AI misuse cruxes with quantified evidence across bioweapons (RAND bio study found no significant difference; novice uplift studies show modest gains on in silico tasks), c...Quality: 65/100 or existential risk mitigation. A more comprehensive connection to AI alignmentApproachAI AlignmentComprehensive review of AI alignment approaches finding current methods (RLHF, Constitutional AI) show 75%+ effectiveness on measurable safety metrics for existing systems but face critical scalabi...Quality: 91/100 research priorities would require examining how compute restrictions affect development timelines for advanced AI systems and the window available for safety research.

Hardware-Level Governance Mechanisms

Hardware-enabled governance (HEG) represents an emerging approach to compute governance that embeds verification and enforcement capabilities directly into AI chips. Key developments include:

• Hardware-Enabled GovernanceApproachHardware-Enabled GovernanceRAND analysis identifies attestation-based licensing as most feasible hardware-enabled governance mechanism with 5-10 year timeline, while 100,000+ export-controlled GPUs were smuggled to China in ...Quality: 70/100 mechanisms that could verify chip location, monitor workloads, or enforce usage restrictions through on-chip firmware
• The FlexHEGProjectFlexHEG (Flexible Hardware-Enabled Guarantees)FlexHEG is a nascent but technically serious proposal to embed tamper-resistant governance processors into AI accelerators, enabling cryptographically verifiable enforcement of international AI com... project, led by Yoshua Bengio with $4.1M in funding, developing a Guarantee Processor architecture for international AI governance verification
• The Chip Security ActPolicyChip Security ActThe Chip Security Act (H.R. 3447/S. 1705) would require location verification mechanisms on advanced AI chips to prevent export control evasion, representing a concrete compute governance intervent... and Stop Stealing Our Chips ActPolicyStop Stealing Our Chips ActThe Stop Stealing Our Chips Act (Senate Apr. 2025, House Dec. 2025) creates an SEC-modeled whistleblower incentive program at BIS to enforce AI chip export controls against smuggling to China and o..., bipartisan U.S. legislation targeting chip smuggling through location verification and whistleblower incentive mechanisms
• Open-source hardware foundations like CHIPS AllianceOrganizationCHIPS AllianceOpen-source hardware development organization under the Linux Foundation. Hosts the Caliptra project (open-source silicon Root of Trust) with contributors including AMD, Google, Microsoft, NVIDIA, ...'s Caliptra Root of Trust, with contributions from AMD, Google, Microsoft, and NVIDIA

For detailed analysis of these mechanisms, their threat models, and implementation timelines, see Hardware Mechanisms for International AI AgreementsAnalysisHardware Mechanisms for International AI AgreementsA comprehensive technical taxonomy of hardware-based AI verification mechanisms—location attestation, TEEs, compute metering, interconnect limits, chip tracking, and remote disablement—mapped to sp....

Key Uncertainties

Several critical uncertainties surround the effectiveness and trajectory of AI chip export controls:

Enforcement Efficacy: Whether the U.S. can effectively prevent diversion of chips through transshipment networks and cloud computing workarounds remains unclear. The scale of illicit flows documented suggests significant enforcement gaps²¹.

Allied Coordination: The degree to which key allies will harmonize their export control regimes with U.S. standards is uncertain, with current implementations showing substantial variation²⁵.

Technological Trajectories: How rapidly algorithmic efficiency and hardware improvements will erode the effectiveness of chip-based controls is unknown. If efficiency gains are rapid, the governance window may be shorter than anticipated³⁶.

Legal Sustainability: Whether revenue-sharing arrangements and tariff structures will survive legal challenges remains to be determined³².

Policy Stability: The significant shifts between administrations (Biden to Trump) raise questions about the long-term stability and coherence of the export control framework¹⁴.

Impact on Innovation: The net effect of controls on U.S. AI innovation and competitiveness is unclear, with competing claims about whether restrictions help or hinder American leadership³⁴.

China's Response: How effectively China can develop domestic alternatives to U.S. chips, either through indigenous innovation or procurement from non-U.S. sources, will significantly determine policy outcomes³⁸.

Sources