Global Compute Governance And AI Infrastructure Regulation

Introduction

Global compute governance and AI infrastructure regulation refers to the emerging legal and competition framework governing access to the physical and digital resources required to develop and deploy artificial intelligence. These resources include GPUs and AI accelerators, cloud-computing capacity, data centres, networking infrastructure, foundation-model infrastructure, electricity, specialised chips, semiconductor manufacturing, and high-speed interconnection services.

Compute has become a strategic economic input. Where a small number of firms control advanced chips, cloud platforms, data-centre capacity, or AI acceleration infrastructure, they may acquire the ability to influence who can develop AI, at what cost, at what scale, and under what contractual conditions.

Competition law therefore increasingly intersects with:

• infrastructure regulation;
• semiconductor policy;
• cloud regulation;
• data governance;
• national-security controls;
• export controls;
• merger control;
• essential-facilities doctrine;
• interoperability;
• access and non-discrimination;
• energy regulation; and
• AI regulation.

The central regulatory question is:

Should access to advanced compute be treated merely as a commercial service, or can sufficiently concentrated compute infrastructure become an essential economic facility subject to competition and regulatory obligations?

1. Meaning of Compute Governance

Compute governance concerns the rules governing the production, allocation, access, pricing, transfer and control of computational capacity.

The relevant infrastructure can be divided into several layers.

A. Semiconductor layer

This includes:

• advanced AI GPUs;
• CPUs;
• TPUs and other accelerators;
• high-bandwidth memory;
• semiconductor manufacturing;
• advanced packaging;
• lithography equipment.

B. Data-centre layer

This includes:

• hyperscale data centres;
• specialised AI clusters;
• cooling systems;
• power infrastructure;
• networking;
• storage;
• physical security.

C. Cloud layer

Cloud providers may control:

• compute rental;
• GPU allocation;
• AI model hosting;
• networking;
• storage;
• model-training environments;
• APIs.

D. Model infrastructure

This includes:

• foundation-model training infrastructure;
• inference infrastructure;
• model-serving systems;
• AI development platforms;
• model registries.

E. Energy layer

Large AI clusters require enormous quantities of electricity. Consequently, electricity-grid capacity, transmission connections and data-centre power contracts can become competitive bottlenecks.

2. Why Compute Has Become a Competition Issue

Traditional competition law often focused on inputs such as:

• raw materials;
• transportation;
• electricity;
• telecommunications;
• financial capital.

Advanced compute increasingly resembles another strategic production input.

An AI developer may need:

Capital → GPUs → cloud capacity → data → networking → electricity → model training → deployment.

If one undertaking controls several of these layers, it can potentially create vertical foreclosure.

For example:

Chip supplier → cloud provider → AI model developer → AI application marketplace

creates substantially different competitive risks from a market in which these functions are independently supplied.

3. Compute Concentration

A major concern is compute concentration.

Suppose:

• Firm A controls 60% of available advanced AI accelerators;
• Firm B controls most hyperscale AI cloud capacity;
• Firm C owns a leading foundation model;
• Firm B also invests in Firm C.

The resulting ecosystem could create:

1. input foreclosure;
2. customer foreclosure;
3. preferential access;
4. discriminatory pricing;
5. tying;
6. exclusivity;
7. self-preferencing;
8. raising rivals' costs;
9. interoperability restrictions.

The problem is particularly acute because AI infrastructure has high fixed costs and economies of scale.

4. Essential-Facilities Theory

One possible legal response is the essential-facilities doctrine.

The basic concept is that a dominant undertaking controlling infrastructure indispensable to competitors may, in exceptional circumstances, be required to provide access.

However, courts generally impose stringent requirements.

Typical questions include:

1. Is the infrastructure genuinely indispensable?
2. Is there a realistic alternative?
3. Does duplication make economic sense?
4. Would refusal eliminate effective competition?
5. Is there an objective justification?
6. Can access be provided without undermining legitimate investment incentives?

Applied to AI:

Could access to a particular AI cloud, GPU cluster, semiconductor technology, or network become indispensable for competition?

The answer will depend heavily on the particular market.

5. The Microsoft/Activision Principle

The European Commission's treatment of digital ecosystems in Microsoft/Activision Blizzard illustrates how competition authorities may examine infrastructure and ecosystem leverage.

The important conceptual issue is not simply market share. It is whether control of one digital layer can be used to disadvantage competitors in another.

The case demonstrates the importance of:

• vertical integration;
• access to infrastructure;
• interoperability;
• ecosystem foreclosure;
• commitments;
• access remedies.

For AI infrastructure, similar reasoning could arise where a cloud provider integrates:

compute + cloud + operating environment + AI models + applications.

6. Google Android

Google Android Case

The European Commission's Google Android decision is highly relevant to AI infrastructure because it demonstrates how control over an important technological ecosystem can be combined with contractual restrictions.

The Commission examined Google's use of:

• tying;
• contractual restrictions;
• exclusivity-related arrangements;
• ecosystem control.

Relevance to AI compute

An AI cloud provider could theoretically condition access to scarce GPU capacity on:

• use of its proprietary AI platform;
• adoption of its model;
• exclusive deployment;
• restrictions on multi-cloud deployment.

Such arrangements could create infrastructure-to-model foreclosure.

The Android case therefore provides a useful conceptual framework for analysing AI infrastructure ecosystems.

7. Google Shopping

Google Shopping Case

The Google Shopping litigation demonstrates the importance of self-preferencing where a platform controls an important gateway.

Although the case did not concern AI compute, its broader significance lies in the relationship between:

infrastructure/gateway control + downstream commercial activity.

An AI infrastructure provider could potentially favour its own:

• AI models;
• inference services;
• applications;
• agents;
• developer tools.

For example, a dominant cloud provider could theoretically provide its own foundation models with:

• better GPU availability;
• preferential latency;
• cheaper internal compute;
• preferential networking;
• privileged technical integration.

Such conduct could raise self-preferencing concerns.

8. Google Search (Shopping and Search)

The wider Google Search jurisprudence demonstrates how control over an important digital gateway can create competitive advantages in adjacent markets.

The AI equivalent could involve a provider controlling:

Compute → model hosting → AI search → AI assistants → applications.

Competition authorities may therefore examine whether infrastructure access creates downstream advantages that rivals cannot realistically reproduce.

9. Intel

Intel Abuse-of-Dominance Case

The European Commission's Intel decision concerned conditional rebates and exclusionary conduct.

The case is particularly relevant because it involved a technology-input market rather than a conventional consumer market.

Intel demonstrated that dominant suppliers of technologically important inputs can potentially use commercial arrangements to exclude rivals.

Application to AI accelerators

A dominant AI-chip supplier might theoretically employ:

• loyalty rebates;
• conditional discounts;
• exclusivity;
• bundled procurement;
• preferential allocation;
• contractual restrictions.

The competition issue would be whether these practices foreclose competing accelerator suppliers.

10. Qualcomm

Qualcomm Cases

The European Commission's Qualcomm proceedings provide another important precedent concerning strategic technology inputs.

The broader lesson is that competition authorities can investigate exclusionary strategies involving technologically significant components even where the immediate product is not itself a consumer-facing platform.

AI relevance

AI accelerators increasingly function as strategic technological inputs.

Competition questions may therefore concern:

• accelerator availability;
• pricing;
• licensing;
• interoperability;
• technical standards;
• exclusivity;
• supply agreements.

11. Bronner

Bronner v Mediaprint

The Bronner judgment of the Court of Justice of the European Union established a demanding standard for refusal-to-deal claims.

The case is important because it emphasises that not every valuable infrastructure is legally essential.

Three major ideas emerge:

1. indispensability;
2. elimination of effective competition;
3. absence of objective justification.

AI infrastructure application

A company should not automatically obtain access to another firm's GPU cluster merely because that cluster is:

• cheaper;
• faster;
• technologically superior;
• more convenient.

The claimant would ordinarily need to establish genuine indispensability.

This prevents competition law from becoming a general price-regulation mechanism.

12. IMS Health

IMS Health

The IMS Health case is another central essential-facilities precedent.

It concerned access to a proprietary information structure.

The Court recognised exceptional circumstances in which refusal to license/access a protected infrastructure could constitute abuse.

AI relevance

The case is useful for analysing situations involving:

• proprietary AI infrastructure;
• closed technical standards;
• specialised datasets;
• model interfaces;
• compute orchestration systems.

If a particular infrastructure becomes indispensable for effective competition, access obligations may become legally conceivable.

13. Magill

Magill

The Magill line of jurisprudence established exceptional circumstances under which refusal to license intellectual property may constitute abuse.

The principle is particularly relevant to AI because AI infrastructure combines:

• patents;
• software;
• interfaces;
• proprietary architectures;
• technical standards.

The existence of intellectual-property rights does not automatically immunise conduct from competition law.

Nevertheless, the threshold remains exceptionally high.

14. Microsoft I

Microsoft I

The EU's Microsoft case concerning interoperability is highly relevant to AI infrastructure regulation.

Microsoft's control over important software interfaces created concerns regarding competitors' ability to interoperate effectively.

AI infrastructure analogy

Comparable questions may arise where dominant AI infrastructure providers restrict:

• model portability;
• API compatibility;
• cloud migration;
• GPU orchestration;
• container portability;
• data movement;
• model deployment.

This connects compute governance with interoperability regulation.

15. Microsoft II / Cloud Ecosystems

Modern cloud markets create a different but related concern.

A cloud provider can potentially combine:

Compute + storage + data + networking + software + AI models.

This creates opportunities for:

• bundling;
• tying;
• technical restrictions;
• switching costs;
• data-egress charges;
• interoperability limitations.

These practices can make customers technically dependent on a provider even where competing cloud services theoretically exist.

16. Amazon Marketplace

Amazon Marketplace

The European Commission and other authorities have examined Amazon's dual role as:

1. infrastructure/platform provider; and
2. downstream competitor.

The central issue is the potential use of platform advantages to benefit the platform's own activities.

AI infrastructure analogy

A hyperscaler could simultaneously operate:

• a cloud;
• an AI accelerator marketplace;
• a foundation-model business;
• an AI application store.

This creates potential conflicts of interest.

The regulatory concern becomes:

Can the infrastructure provider fairly allocate scarce compute to competing AI developers while simultaneously competing with those developers?

17. Apple App Store

Apple App Store Litigation

The Apple App Store disputes illustrate the competition problems associated with control over an important technological distribution gateway.

Although app distribution is different from compute, the structural principle is relevant:

Control over an unavoidable technological gateway can create downstream competitive power.

An AI cloud infrastructure provider could become a comparable gateway if AI developers cannot economically obtain sufficient compute elsewhere.

18. NVIDIA and AI Accelerator Concentration

The rise of advanced AI has created a strategic competition question surrounding accelerator supply.

A dominant accelerator ecosystem can potentially benefit from:

• hardware advantages;
• software ecosystems;
• developer tools;
• compatibility;
• optimisation libraries;
• network effects.

The competitive advantage therefore may not arise solely from the physical GPU.

It can arise from:

GPU + software stack + developer ecosystem + cloud availability + switching costs.

Competition authorities may increasingly evaluate the entire AI compute stack rather than individual hardware products.

19. Cloud Computing as an Essential Input

Cloud infrastructure raises a particularly difficult issue.

A cloud provider may possess:

• enormous capital resources;
• global data centres;
• proprietary networking;
• AI-specific hardware;
• specialised software.

New entrants may technically be able to build their own infrastructure but economically be unable to replicate it.

This creates a distinction between:

Technical duplicability

The infrastructure can theoretically be reproduced.

Economic duplicability

The infrastructure can realistically be reproduced at commercially viable cost.

Competition law increasingly has to confront this distinction.

20. Compute Scarcity and Raising Rivals' Costs

A dominant infrastructure provider could potentially disadvantage rivals by:

• reserving scarce GPU capacity;
• prioritising affiliated businesses;
• imposing discriminatory access conditions;
• delaying capacity allocation;
• imposing excessive switching costs;
• restricting interoperability.

The resulting effect may be:

raising rivals' costs rather than directly excluding them.

This is especially important in AI markets because compute shortages can substantially increase the time and capital required to train a model.

21. AI Infrastructure and Merger Control

Merger control may become one of the most important tools of compute governance.

Authorities may examine acquisitions involving:

• chip manufacturers;
• cloud providers;
• AI startups;
• model developers;
• data-centre operators;
• networking firms;
• AI software companies.

A transaction may raise concerns even when the target has relatively low current revenues.

Why?

Because the target may possess:

• critical technology;
• strategic talent;
• unique model architecture;
• specialised compute;
• future competitive potential.

This is the innovation-competition / nascent-competition problem.

22. Killer Acquisitions in AI Infrastructure

Traditional turnover thresholds can fail to capture strategically important AI transactions.

A small AI infrastructure startup may have:

• low revenue;
• enormous technological significance.

Its acquisition by a hyperscaler could eliminate a future infrastructure competitor.

Competition authorities may therefore examine:

• future innovation;
• pipeline products;
• technological capability;
• access to compute;
• potential competition.

23. National-Security Regulation

Compute governance is not exclusively competition law.

Advanced compute has become a national-security issue.

Governments may regulate:

• semiconductor exports;
• advanced AI chips;
• data-centre construction;
• foreign investment;
• cloud services;
• cross-border compute;
• technology transfers.

This creates a major tension:

Competition policy seeks open markets, while national-security policy may deliberately restrict access.

24. Export Controls and Competition

Export restrictions on advanced AI accelerators can affect competition by limiting which firms or countries can access advanced compute.

They may:

• protect national security;
• slow technological diffusion;
• reshape global supply chains;
• increase scarcity;
• increase prices;
• encourage domestic alternatives.

Competition authorities therefore must distinguish between:

Legitimate governmental restrictions

and

Private restrictions that exploit government-created scarcity.

25. Data Centres and Energy Regulation

AI infrastructure is increasingly connected to energy markets.

A hyperscale AI data centre may require:

• large electricity connections;
• transmission capacity;
• cooling water;
• land;
• fibre networks.

Consequently, competition concerns may arise when data-centre operators obtain preferential access to:

• electricity;
• grid connections;
• public land;
• tax incentives;
• renewable-energy contracts.

This raises the possibility of infrastructure subsidy distortion.

26. State Aid and AI Infrastructure

Governments increasingly subsidise semiconductor and AI infrastructure.

State support may include:

• grants;
• tax credits;
• loans;
• infrastructure subsidies;
• energy incentives;
• research funding.

Competition-law questions include:

1. Does the subsidy favour particular firms?
2. Does it distort competition?
3. Is it technology-neutral?
4. Does it create excess concentration?
5. Does it establish a government-backed infrastructure monopoly?

The EU state-aid framework is therefore potentially important to AI infrastructure governance.

27. DMA-Style Infrastructure Regulation

The Digital Markets Act demonstrates a movement from traditional ex-post competition law toward ex-ante obligations for powerful digital gatekeepers.

AI infrastructure may eventually generate analogous regulatory thinking.

Possible obligations could include:

• interoperability;
• transparency;
• non-discrimination;
• portability;
• switching rights;
• restrictions on self-preferencing;
• access to technical interfaces.

The crucial distinction is:

Competition law asks whether conduct is unlawful; infrastructure regulation can establish obligations before the harm fully occurs.

28. Cloud Switching and Data Egress

One of the most important emerging issues is cloud lock-in.

A customer may face significant costs when moving:

• datasets;
• trained models;
• applications;
• containers;
• inference workloads;
• security configurations.

Even if another cloud offers lower prices, migration may be commercially impractical.

This can create artificial switching costs.

The competitive effect is:

Customers become captive not because the provider is necessarily superior, but because leaving becomes too expensive.

29. Multi-Cloud Competition

Multi-cloud deployment can reduce dependency.

However, infrastructure providers may create obstacles through:

• incompatible APIs;
• proprietary software;
• data-egress costs;
• contractual restrictions;
• technical dependencies;
• preferential internal services.

Competition policy may therefore increasingly treat portability and interoperability as competitive parameters.

30. AI Compute as a Bottleneck Facility

The strongest regulatory case arises where compute becomes a genuine bottleneck.

A bottleneck facility exists where:

a critical upstream resource is controlled by a limited number of suppliers and downstream competitors cannot reasonably reproduce it.

Potential examples include:

• advanced AI accelerators;
• high-bandwidth memory;
• specialised interconnects;
• AI-optimised data centres;
• scarce power connections;
• specialised cloud clusters.

But bottleneck status should be established through economic evidence rather than assumed merely because a technology is important.

31. Six Core Legal Tests for AI Compute Regulation

Authorities could evaluate AI infrastructure using six questions.

Test 1 — Market definition

What is the relevant market?

Possible markets include:

• AI accelerators;
• general-purpose GPUs;
• AI cloud computing;
• model-training compute;
• inference compute;
• specialised AI networking.

Test 2 — Market power

Authorities may examine:

• market share;
• capacity;
• switching costs;
• entry barriers;
• intellectual property;
• software ecosystems.

Test 3 — Indispensability

Can rivals reasonably obtain equivalent compute elsewhere?

Test 4 — Conduct

Is the provider engaging in:

• tying;
• bundling;
• exclusivity;
• discrimination;
• self-preferencing;
• refusal to deal?

Test 5 — Effects

Does the conduct:

• increase costs;
• reduce innovation;
• delay entry;
• reduce consumer choice;
• restrict AI development?

Test 6 — Justification

Is the conduct objectively justified by:

• security;
• capacity constraints;
• reliability;
• investment incentives;
• legitimate technical requirements?

32. Regulatory Models

Three broad models are possible.

Model A — Traditional competition law

Authorities intervene only when dominance and abusive conduct are established.

Advantage: preserves market incentives.

Disadvantage: intervention may come too late.

Model B — Ex-ante infrastructure regulation

Systemically important compute providers receive obligations concerning:

• access;
• interoperability;
• transparency;
• switching.

Advantage: prevents structural dependency.

Disadvantage: risks excessive regulation of innovation.

Model C — Hybrid model

A hybrid system could classify certain compute infrastructure as systemically important digital infrastructure and impose proportionate obligations.

This is likely to be the most flexible approach.

33. Global Regulatory Divergence

Different jurisdictions may approach compute governance differently.

United States

Emphasis tends to combine:

• antitrust;
• national security;
• semiconductor policy;
• export controls;
• sectoral regulation.

European Union

Greater emphasis may be placed on:

• competition law;
• DMA;
• AI regulation;
• data regulation;
• state aid;
• strategic autonomy.

United Kingdom

The UK model increasingly combines:

• competition law;
• digital-markets regulation;
• cloud and infrastructure competition;
• national-security considerations.

China

The framework combines:

• competition law;
• industrial policy;
• cybersecurity;
• data governance;
• strategic technology policy.

This divergence can itself become a competition issue.

34. Cross-Border Compute Governance

AI infrastructure is inherently international.

A model may be:

• trained in one country;
• hosted in another;
• supplied with chips from a third;
• operated through a cloud provider headquartered elsewhere;
• serving users globally.

Therefore, conflicting regulations can produce:

• compliance duplication;
• regulatory arbitrage;
• fragmented infrastructure;
• localisation;
• higher costs.

International cooperation will increasingly become necessary.

35. Case-Law Principles at a Glance

36. Emerging Legal Risks

The most significant future risks include:

1. Compute hoarding

Large firms reserve scarce capacity and leave insufficient capacity for rivals.

2. Infrastructure tying

Cheap compute is conditional upon purchasing other AI services.

3. Cloud lock-in

Technical and contractual restrictions make switching prohibitively expensive.

4. Vertical foreclosure

Cloud providers disadvantage competing AI models.

5. Accelerator foreclosure

Dominant chip suppliers restrict competing architectures.

6. Data-centre exclusion

Large incumbents secure scarce power and land resources.

7. Regulatory capture

Dominant infrastructure providers influence technical standards and regulatory design.

8. State-supported concentration

Government subsidies unintentionally strengthen already dominant infrastructure firms.

37. Future Competition-Law Framework

A mature global framework could contain:

Compute transparency

↓

Disclosure of capacity, allocation and major contractual restrictions

↓

Non-discrimination

↓

Equivalent customers receive comparable access conditions

↓

Interoperability

↓

Cloud and AI systems can operate across providers

↓

Portability

↓

Models, data and workloads can migrate

↓

Merger scrutiny

↓

Strategic AI infrastructure acquisitions receive heightened examination

↓

Essential-facilities intervention

↓

Access remedies only where genuine indispensability is established

↓

International coordination

↓

Competition authorities cooperate across jurisdictions.

Conclusion

Global compute governance is evolving from a purely technical issue into a major competition-law and economic-governance question.

The central concern is not simply that AI requires enormous computing power. It is that control over compute can become control over participation in the AI economy itself.

The jurisprudence of Bronner, IMS Health, Magill, Microsoft, Intel, Qualcomm, Google Android, Google Shopping and Amazon provides useful legal foundations for analysing these emerging problems, even though most of these cases predate the present AI-compute market.

The key future question will be whether advanced compute should be treated as:

an ordinary commercial input, a strategically important infrastructure service, or—in exceptional circumstances—an essential facility.

A balanced regime should avoid automatically imposing access obligations on every important AI infrastructure provider. Instead, regulators should focus on indispensability, durable market power, foreclosure effects, switching costs, interoperability, and objective justification.