AI Chip Export Controls and Global Compute Inequality

Introduction

AI chips—especially advanced GPUs, AI accelerators, high-bandwidth-memory components, and the semiconductor manufacturing equipment needed to produce them—have become strategically important because large-scale AI development depends on access to substantial computing capacity.

AI chip export controls are government restrictions on the export, re-export, transfer, servicing, or use of advanced semiconductor technology. They may target particular countries, entities, end uses, technologies, or computing-capacity thresholds.

The central legal issue is the tension between:

1. national security and strategic technology protection;
2. freedom of international trade;
3. technology-transfer regulation;
4. developmental access to computing infrastructure; and
5. the risk that differentiated access to AI chips produces global compute inequality.

The United States has progressively expanded controls on advanced computing chips and semiconductor manufacturing equipment directed toward China, including measures addressing circumvention and foreign-produced items.

I. Meaning of AI Chip Export Controls

AI chip export controls can operate through several mechanisms.

1. Product-based controls

A government may identify particular technical characteristics that trigger licensing requirements, such as:

• computing performance;
• interconnect bandwidth;
• memory bandwidth;
• processing speed;
• AI acceleration capability;
• chip architecture; or
• ability to operate at large scale.

This approach attempts to regulate the capability of the chip rather than merely its commercial name.

2. Country-based controls

Controls may distinguish between:

• unrestricted countries;
• countries requiring licenses;
• countries subject to presumptions of denial; and
• comprehensively restricted jurisdictions.

The result is that the same AI accelerator can have different legal availability depending upon its destination.

3. Entity-based restrictions

A company, university, research institution, data centre, or other organization may be placed on an export-control list.

This makes access dependent not merely on the physical destination but also on who receives or uses the technology.

4. End-use controls

Restrictions may apply where chips are intended for:

• military applications;
• supercomputing;
• advanced weapons development;
• surveillance;
• nuclear applications;
• advanced AI development; or
• other controlled applications.

5. Semiconductor-equipment controls

Export restrictions increasingly extend upstream to:

• lithography equipment;
• deposition equipment;
• etching equipment;
• inspection equipment;
• electronic-design-automation tools;
• advanced packaging technology; and
• semiconductor manufacturing components.

Thus, export control can affect the ability to manufacture future AI chips, rather than merely restricting existing chips.

II. Development of the AI Chip Export-Control Regime

The modern regime intensified substantially after 2022.

The October 2022 U.S. rules sought to restrict China's access to advanced computing chips and semiconductor manufacturing equipment. Subsequent rules in 2023 and 2024 refined and expanded those controls.

The legal architecture can therefore be viewed as:

Advanced chip → Manufacturing equipment → Design software → Data centre → Computing capacity → AI capability

Controlling any one of these stages can influence the availability of advanced AI compute.

III. What Is Global Compute Inequality?

Global compute inequality describes unequal access among countries, companies, universities, researchers and other actors to computational resources necessary to develop and deploy advanced AI.

It is different from traditional technological inequality.

Traditional technological inequality might concern:

Who possesses computers?

Compute inequality asks:

Who possesses sufficient advanced computational capacity to train, fine-tune and operate frontier AI systems?

Access depends on:

• AI accelerators;
• data centres;
• electricity;
• cooling infrastructure;
• networking;
• semiconductor supply;
• cloud-computing services;
• capital;
• technical expertise; and
• access to advanced models.

Consequently, export controls can produce indirect computational effects even where they do not expressly prohibit AI research.

IV. How Export Controls Can Produce Compute Inequality

A. Direct reduction of chip availability

If advanced accelerators cannot be exported to a particular jurisdiction, its firms may have fewer high-performance computing resources.

This can affect:

• AI training;
• scientific research;
• pharmaceutical modelling;
• climate modelling;
• autonomous systems;
• robotics;
• financial modelling; and
• large language models.

B. Increased cost of compute

Scarcity can increase the effective cost of computing.

A firm that cannot purchase the newest accelerator may have to use:

• older GPUs;
• less efficient processors;
• domestically developed alternatives;
• cloud resources in other jurisdictions; or
• larger quantities of less capable hardware.

Therefore:

Export restriction → scarcity → higher compute cost → slower AI development

C. Economies of scale

Advanced AI development often rewards large-scale infrastructure.

Suppose:

• Country A can obtain 1 million advanced accelerators;
• Country B can obtain only 50,000;
• Country C can obtain only older-generation accelerators.

Even if all three countries possess engineers and data, their practical AI-development capacity may diverge substantially.

This produces a compute-capital asymmetry.

V. Extraterritoriality and the Global Semiconductor Supply Chain

One of the most important legal issues is the Foreign Direct Product Rule (FDPR).

Under this mechanism, U.S. controls can in certain circumstances apply to foreign-produced items because of their connection to U.S.-origin technology, software or manufacturing equipment.

The significance is substantial because semiconductor manufacturing is geographically fragmented.

For example:

U.S. design technology

↓

Dutch/Japanese/Korean/Taiwanese equipment and components

↓

Foreign semiconductor fabrication

↓

AI accelerator

↓

Global data centre

A restriction imposed at one point in this chain can affect multiple jurisdictions.

The FDPR has therefore become an important instrument for preventing circumvention through third countries.

VI. Territorial Sovereignty Versus Extraterritorial Regulation

A fundamental legal question is:

How far may one country regulate transactions occurring outside its territory?

Traditional international economic law generally distinguishes territorial jurisdiction from extraterritorial effects.

However, semiconductor controls complicate the issue because:

• the product may contain U.S.-origin technology;
• manufacturing may use U.S.-controlled equipment;
• design software may originate in the United States;
• the transaction may involve U.S. persons; and
• the ultimate end user may be in another country.

Thus, modern export control increasingly regulates technological dependency rather than geographical location alone.

VII. National Security Exception Under International Trade Law

The strongest legal justification for AI-chip restrictions is generally national security.

GATT Article XXI contains an essential-security exception.

The major question is whether a state can simply characterize any technology restriction as a national-security measure or whether international review remains possible.

This issue became particularly important in:

1. Russia – Measures Concerning Traffic in Transit

The WTO Panel considered GATT Article XXI and held that the security exception was not completely beyond legal review.

The decision is important for semiconductor export controls because it demonstrates that national-security measures can raise questions about:

• existence of an emergency;
• essential security interests;
• connection between the measure and the emergency; and
• good faith.

It is therefore relevant to the legal architecture surrounding technology-based national-security restrictions.

VIII. Six Important Case Laws / Legal Authorities

Because there are relatively few reported judicial decisions specifically concerning AI-chip export controls, the following authorities combine direct semiconductor/export-control cases with WTO cases dealing with national-security, semiconductor trade and technology restrictions.

1. United States v. Shih, 9th Cir. (2024)

This is one of the most directly relevant modern semiconductor export-control cases.

The defendant was convicted in connection with exporting monolithic microwave integrated circuits to China without the required authorization.

The Ninth Circuit treated the relevant controls as national-security controls and rejected the argument that their connection with international export-control arrangements prevented them from being treated as national-security restrictions.

Importance

The case establishes that:

• semiconductor-related components can fall within national-security export controls;
• technical classification matters;
• export-control violations can produce criminal liability;
• international control arrangements do not necessarily remove domestic national-security character; and
• compliance obligations can extend beyond conventional weapons.

Relevance to AI chips

Modern AI accelerators contain highly sophisticated semiconductor technology. The reasoning illustrates why governments can treat certain semiconductor technologies as strategic technologies rather than ordinary commercial goods.

2. United States v. Chi Mak, 9th Cir. (2012)

Chi Mak concerned the attempted export of controlled naval technology to China under the Arms Export Control Act and ITAR.

The Ninth Circuit upheld the conviction and recognized the government's important interest in regulating the international dissemination of military technology.

Importance

The case addresses:

• controlled technical information;
• licensing;
• national-security interests;
• willfulness;
• public-domain limitations; and
• constitutional challenges to export controls.

Relevance to AI

AI semiconductor technology increasingly has dual-use characteristics. The Chi Mak reasoning demonstrates how technical information can be controlled even when the information is not itself a physical weapon.

3. United States v. Liang, 9th Cir. (2013)

Liang involved evasion of national-security export controls concerning thermal-imaging technology.

The Ninth Circuit's treatment of national-security controls is significant because it demonstrates that criminal consequences can attach to evasion of licensing requirements even when the controlled technology is not itself a conventional weapon. The Shih court relied upon Liang in discussing the sentencing consequences associated with evasion of national-security controls.

Relevance

The case supports the broader proposition that dual-use technological capabilities may be regulated because of their strategic applications.

That principle is directly relevant to:

• AI accelerators;
• high-performance computing;
• advanced networking;
• semiconductor equipment; and
• AI-enabled military systems.

4. United States v. McKeeve, 1st Cir. (1997)

McKeeve involved an attempt to export computer equipment to Libya without the required authorization.

The First Circuit addressed the consequences of evading national-security-related export controls. Later Ninth Circuit authority has relied on McKeeve when interpreting the relevant sentencing framework.

Importance

The case illustrates that:

computer technology itself can become the subject of national-security export regulation.

This is particularly important for AI because advanced AI computing hardware increasingly has strategic significance independent of its immediate military application.

5. United States – Measures on Certain Semiconductor and Other Products and Related Services and Technologies, WTO DS615

This is the most directly relevant international trade dispute.

China initiated WTO consultations challenging U.S. measures concerning:

• advanced computing semiconductor chips;
• supercomputer items;
• semiconductor manufacturing equipment;
• related services; and
• related technologies.

China invoked provisions including GATT Articles I, X and XI, TRIMS, TRIPS and GATS. The United States maintained that the measures involved national-security issues not susceptible to WTO dispute settlement.

Importance

DS615 directly raises the central legal conflict:

Can restrictions on advanced computing technology be challenged under international trade law when the restricting state characterizes them as national-security measures?

Relevance to compute inequality

This dispute is especially important because the regulated products are precisely the technologies that determine access to advanced computing.

6. China – Value-Added Tax on Integrated Circuits, WTO DS309

In DS309, the United States challenged China's preferential VAT treatment for domestically produced or designed integrated circuits.

The dispute concerned alleged discrimination between domestic and imported IC products.

China and the United States subsequently reached a mutually agreed solution, with China agreeing to modify or revoke the relevant VAT-refund measures.

Importance

The case demonstrates the other side of semiconductor competition:

export controls are not the only mechanism capable of altering global semiconductor access.

Governments can also use:

• taxation;
• subsidies;
• domestic preferences;
• procurement;
• investment incentives; and
• industrial policy.

These measures can affect the distribution of semiconductor production capacity and consequently global compute.

7. Russia – Traffic in Transit, WTO

Although not a semiconductor case, this WTO authority is particularly important for understanding national-security exceptions.

The Panel considered whether Article XXI of GATT could be reviewed and established a framework for examining the invocation of the security exception rather than treating it as completely self-judging.

Relevance to AI chips

If a semiconductor export restriction is justified exclusively through national security, the case provides an important conceptual framework for asking:

1. What security interest is being protected?
2. Is there a relevant international emergency?
3. Is the measure genuinely connected with that interest?
4. Is the measure applied in good faith?

IX. Competition-Law Dimension

AI chip export controls also create competition concerns.

1. Supply concentration

If only a small number of firms can manufacture the most advanced AI accelerators, export restrictions can amplify their strategic importance.

The market can therefore exhibit:

• supplier concentration;
• capacity constraints;
• switching costs;
• technological lock-in; and
• dependence upon particular architectures.

2. Geographic fragmentation

Different countries may develop separate AI-computing ecosystems.

For example:

Ecosystem A

Advanced chips → advanced manufacturing → frontier AI

versus

Ecosystem B

restricted chips → domestic substitutes → lower computing efficiency

This can produce technological bifurcation.

X. AI Chip Export Controls and Developing Countries

Global compute inequality is particularly significant for developing economies.

A country may lack:

• domestic semiconductor fabs;
• advanced packaging;
• GPU manufacturing;
• high-end cloud infrastructure;
• sufficient electricity;
• capital for data centres;
• advanced AI researchers; and
• access to frontier accelerators.

Export controls can therefore interact with pre-existing infrastructure inequality.

The result can be:

Semiconductor inequality → compute inequality → AI-development inequality

This does not mean every export control necessarily causes such inequality. The effect depends on alternative supply sources, licensing policies, cloud access, domestic capacity and the availability of less advanced chips.

XI. Cloud Computing Creates a New Legal Problem

A major development is that physical possession of a chip is no longer necessarily required.

A researcher may access advanced computing through:

Cloud provider → remote data centre → AI accelerator

Therefore, controlling physical chip exports without regulating access to cloud computing can create an enforcement gap.

Conversely, imposing extensive restrictions on foreign cloud access can broaden export controls beyond traditional goods regulation.

This raises questions concerning:

• jurisdiction;
• remote access;
• data sovereignty;
• services regulation;
• model-weight transfers;
• foreign cloud providers; and
• circumvention.

XII. Circumvention and Transshipment

Export restrictions can generate incentives for third-country transactions.

Potential mechanisms include:

• shell companies;
• intermediaries;
• false end-user certificates;
• transshipment;
• cloud-based access;
• re-export;
• disguised data-centre ownership; and
• diversion through jurisdictions with weaker controls.

This explains why modern regimes increasingly emphasize end-user and end-use controls, not merely customs declarations.

U.S. authorities have repeatedly strengthened controls to address diversion and circumvention.

XIII. The January 2025 AI Diffusion Framework

A significant development was the January 2025 U.S. AI Diffusion Framework, which attempted to structure global access to advanced AI computing through differentiated country treatment.

The framework contemplated different categories of countries and introduced mechanisms involving computing-capacity allocations and controls on advanced AI chips and AI infrastructure.

This illustrates a shift from:

"Do not export this chip to Country X."

toward:

"Manage the global distribution of advanced AI computing capacity."

That is a much broader regulatory concept.

XIV. Legal Issues Created by Global Compute Allocation

1. Non-discrimination

Differentiating countries can raise questions under:

• GATT Article I;
• GATT Article XI;
• GATS;
• TRIMS;
• national-security exceptions; and
• applicable bilateral or regional agreements.

2. Proportionality

A central policy/legal question is whether restrictions are appropriately connected to the security objective.

Possible alternatives include:

• licensing;
• end-use monitoring;
• quantity restrictions;
• controlled data centres;
• trusted-user programs;
• technical safeguards; and
• post-export verification.

3. Extraterritoriality

Controls affecting foreign-produced chips or foreign data centres raise questions concerning the permissible territorial reach of domestic export-control legislation.

4. Due process

Companies may need to understand:

• why a product is controlled;
• what technical threshold applies;
• how licensing works;
• how an entity is designated; and
• what appeal or review mechanisms exist.

Uncertainty can itself impose significant compliance costs.

XV. Relationship Between Export Controls and Technological Sovereignty

Export controls can encourage countries to pursue technological sovereignty.

A country subject to restrictions may respond by developing:

• domestic AI accelerators;
• domestic semiconductor fabs;
• indigenous EDA software;
• advanced packaging;
• domestic cloud infrastructure;
• alternative architectures; and
• domestic AI models.

Consequently, export controls may have two simultaneous effects:

Short-term effect

Reduced access to restricted technology.

Long-term effect

Potential acceleration of domestic technological substitution.

The ultimate effect is therefore not legally or economically predetermined.

XVI. Global Compute Inequality as a Competition Concern

The competition-law dimension can be conceptualized through five forms of inequality:

The most significant concern is the final transition:

hardware inequality → infrastructure inequality → innovation inequality

XVII. Possible Legal and Regulatory Responses

A balanced framework could include:

1. Risk-based licensing

Licensing based upon:

• chip capability;
• end user;
• end use;
• destination;
• computing scale; and
• security safeguards.

2. Trusted data-centre regimes

Approved foreign data centres could receive advanced hardware subject to:

• customer verification;
• monitoring;
• cybersecurity requirements;
• physical security; and
• restrictions on onward transfer.

3. Multilateral coordination

Instead of unilateral export-control regimes, states could develop common rules concerning:

• AI accelerators;
• semiconductor equipment;
• cloud compute;
• model weights;
• technical standards; and
• circumvention.

4. Development-oriented access

International mechanisms could distinguish between:

• military applications;
• commercial AI;
• academic research;
• healthcare;
• climate research; and
• humanitarian applications.

5. Transparency

Export-control authorities could publish clearer criteria concerning:

• performance thresholds;
• licensing standards;
• country categories;
• entity restrictions;
• enforcement priorities; and
• review procedures.

XVIII. Key Legal Principles Emerging From the Case Law

The authorities collectively demonstrate several principles:

1. Advanced technological information can be legally controlled.
2. Semiconductor components can fall within national-security export regimes.
3. Dual-use technology can receive national-security treatment.
4. Willful export-control violations can attract criminal sanctions.
5. International trade obligations may intersect with national-security exceptions.
6. National-security claims are not necessarily completely immune from international legal scrutiny.
7. Semiconductor taxation and domestic preferences can also distort international semiconductor markets.
8. Modern export controls increasingly operate through supply chains rather than territorial borders alone.

XIX. Conclusion

AI-chip export controls represent a transformation from traditional export regulation into strategic governance of computing capacity.

The legal problem is no longer simply whether a particular semiconductor may cross a border. It increasingly concerns:

Who can access the computational infrastructure necessary to participate in the AI economy?

The cases involving Shih, Chi Mak, Liang and McKeeve demonstrate the domestic legal foundations for controlling strategically significant technology. The WTO authorities—particularly US–Semiconductors (DS615), China–Integrated Circuits (DS309), and Russia–Traffic in Transit—illustrate the international-law questions surrounding semiconductor trade, discrimination and national-security exceptions.

The concept of global compute inequality therefore connects export-control law with international trade law, competition law, technology sovereignty and development policy. The central challenge for future regulation will be to protect legitimate national-security interests while avoiding unnecessary fragmentation of the global AI ecosystem and ensuring that access to computational capacity does not become concentrated exclusively among a limited group of states and firms.