Global Research Compute Inequality And Knowledge Concentration

Introduction

Global research compute inequality refers to the unequal distribution of high-performance computing resources—such as GPUs, TPUs, supercomputers, cloud computing capacity, specialised AI accelerators, high-speed networking and large-scale data infrastructure—among universities, research institutions, governments and private firms.

The problem becomes particularly significant in artificial intelligence and computational science because access to compute increasingly determines who can conduct frontier research, train large models, reproduce experiments, discover scientific knowledge and influence technological standards.

This produces a related phenomenon of knowledge concentration: when a relatively small number of corporations, universities, states or research laboratories possess disproportionate computational resources, they may also acquire disproportionate control over research agendas, datasets, publications, patents, talent and technological standards.

Competition law does not generally treat unequal research resources as unlawful by itself. The legal concern arises when compute concentration is reinforced through exclusionary conduct, discriminatory access, tying, refusal to deal, mergers, intellectual-property restrictions, platform self-preferencing, or control over essential technological infrastructure.

1. Meaning of Research Compute Inequality

Research compute inequality has several dimensions.

A. Hardware inequality

Frontier research may require enormous quantities of:

• GPUs;
• AI accelerators;
• supercomputers;
• high-bandwidth memory;
• networking equipment;
• storage;
• specialised processors.

A university with a small research cluster may therefore be unable to reproduce research performed by an organisation possessing tens of thousands of accelerators.

B. Cloud-compute inequality

Even where researchers do not own hardware, they may depend upon cloud providers.

This creates a distinction between:

physical access → possession of computing infrastructure

and

economic access → ability to purchase sufficient computing capacity at affordable prices.

C. Data-compute interaction

Compute is rarely independent of data.

A research institution possessing:

data + compute + researchers + distribution

can create a cumulative advantage over competitors.

D. Talent concentration

Researchers may migrate toward institutions capable of providing frontier compute.

This creates a feedback loop:

Compute → research → publications/patents → reputation → talent → investment → more compute

2. Knowledge Concentration

Knowledge concentration occurs when the production, control or dissemination of knowledge becomes disproportionately dependent upon a limited number of entities.

It may involve:

1. concentration of research infrastructure;
2. concentration of scientific datasets;
3. concentration of AI models;
4. concentration of research talent;
5. concentration of intellectual property;
6. concentration of scientific publishing;
7. concentration of cloud infrastructure;
8. concentration of research funding.

The result can be a computational knowledge bottleneck.

For example:

If only a few organisations can afford to train the largest scientific AI models, they may determine which scientific questions receive computational attention.

This is more significant than ordinary market concentration because it can affect the direction of scientific development itself.

3. Why Compute Is Becoming a Competitive Input

Traditional competition analysis often focuses on inputs such as:

• raw materials;
• labour;
• capital;
• transportation;
• electricity.

In AI-intensive research, compute itself can become a strategically important input.

The relevant competitive question may therefore become:

Can researchers obtain sufficient computational resources to compete in the production of knowledge?

If a dominant infrastructure provider controls access to scarce accelerators, cloud capacity or specialised computing environments, its conduct can affect downstream research markets.

4. Relevant Competition-Law Theories

A. Essential-facilities theory

Where a computational infrastructure is genuinely indispensable and cannot reasonably be replicated, refusal to provide access may raise essential-facilities concerns.

However, competition law traditionally applies this doctrine cautiously.

The claimant normally needs to demonstrate something approaching:

1. indispensability;
2. lack of realistic alternatives;
3. exclusionary effect;
4. inability to reasonably duplicate the facility;
5. absence of legitimate justification.

B. Refusal to deal

A dominant cloud or infrastructure provider could potentially face scrutiny if it:

• refuses access to critical research infrastructure;
• selectively restricts capacity;
• terminates access to competing researchers;
• discriminates against downstream competitors;
• uses infrastructure control to protect another market.

The difficult question is whether the infrastructure is merely commercially important or legally indispensable.

C. Discriminatory access

Suppose a dominant cloud provider supplies compute to universities but gives its own affiliated AI laboratory:

• preferential GPU allocation;
• lower prices;
• priority queues;
• superior networking;
• privileged access to new accelerators.

The conduct could raise discrimination and self-preferencing concerns where it disadvantages competing downstream researchers or enterprises.

D. Tying and bundling

Compute may be bundled with:

• cloud storage;
• proprietary AI models;
• software frameworks;
• data services;
• APIs;
• developer tools.

A dominant infrastructure provider might therefore make access to scarce compute conditional upon purchasing another service.

5. The Knowledge-Concentration Feedback Loop

A particularly important concern is the cumulative advantage mechanism.

Stage 1 — Compute concentration

A small number of institutions obtain disproportionate computing capacity.

↓

Stage 2 — Research concentration

They conduct experiments that smaller institutions cannot afford.

↓

Stage 3 — Intellectual-property concentration

Successful research generates:

• patents;
• proprietary models;
• datasets;
• trade secrets;
• publications.

↓

Stage 4 — Talent concentration

Researchers move toward better-funded institutions.

↓

Stage 5 — Capital concentration

Investors and governments direct additional resources toward proven research centres.

↓

Stage 6 — Further compute concentration

The leading institutions purchase even greater computational capacity.

This creates a self-reinforcing technological advantage.

6. Case Laws

The following cases do not all concern AI research compute directly. They provide important legal principles applicable to the emerging problem of compute concentration, infrastructure access, intellectual property and technological bottlenecks.

Case 1 — United States v. Terminal Railroad Association

Supreme Court of the United States

This is one of the foundational American cases concerning control over an indispensable infrastructure facility.

A group of railroad companies controlled access to an important terminal facility, creating barriers for competing railroads.

The Supreme Court found that exclusionary control over the infrastructure could violate competition law.

Relevance to research compute

The case provides an analogy for computational infrastructure.

If a small number of entities control an infrastructure that competitors genuinely cannot reasonably replicate, the legal issue becomes whether access is being used to exclude competition.

The analogy must nevertheless be applied carefully because modern cloud computing normally has multiple providers and alternatives.

Principle

Control over indispensable infrastructure can create competition-law obligations where exclusion substantially forecloses rivals.

7. United States v. AT&T

United States Supreme Court / U.S. antitrust proceedings

The AT&T litigation concerned control over telecommunications infrastructure and the relationship between infrastructure ownership and competitive markets.

The case ultimately resulted in structural separation of major telecommunications activities.

Relevance

The telecommunications analogy is particularly useful for research compute.

Cloud infrastructure increasingly resembles a foundational layer upon which numerous downstream activities depend:

chips → cloud → computing → AI models → applications → scientific research

Where infrastructure ownership enables control over downstream markets, structural remedies may become relevant.

Principle

Vertical integration becomes especially significant when control over an upstream infrastructure can be used to restrict downstream competition.

8. Aspen Skiing Co. v. Aspen Highlands Skiing Corp.

U.S. Supreme Court

Aspen Skiing is a leading refusal-to-deal case.

A dominant ski operator discontinued a cooperative arrangement that had previously allowed customers to purchase an integrated multi-mountain ticket.

The Supreme Court treated the termination of profitable cooperation under particular circumstances as potentially exclusionary conduct.

Relevance to compute

Consider a dominant computational infrastructure provider that historically supplies capacity to independent research organisations but suddenly withdraws access in circumstances suggesting an intention to disadvantage those organisations.

The Aspen Skiing reasoning provides a framework for considering:

• prior cooperation;
• termination;
• economic rationality;
• competitive effects;
• exclusionary purpose.

Limitation

Aspen Skiing is exceptional rather than a general requirement that dominant companies supply competitors.

9. Verizon Communications Inc. v. Trinko

U.S. Supreme Court

Trinko significantly limited the circumstances in which refusal to deal by a dominant firm constitutes an antitrust violation.

The Court emphasised the risks of requiring firms to share infrastructure and recognised that forced sharing can reduce incentives to invest.

Relevance

This case creates the central tension in research-compute regulation:

Should dominant compute providers be required to share scarce infrastructure with researchers?

There are competing considerations.

Access argument

Sharing can:

• broaden research participation;
• prevent technological exclusion;
• increase innovation;
• reduce knowledge concentration.

Investment argument

Compulsory access can potentially:

• reduce incentives to build infrastructure;
• increase regulatory costs;
• create capacity-allocation disputes;
• discourage investment in next-generation computing.

Principle

Competition law must balance access concerns against incentives for infrastructure investment.

10. Microsoft Corp. v. United States

The Microsoft antitrust litigation concerned Microsoft's control over the operating-system ecosystem and its efforts to protect its position against emerging competitive threats.

The case demonstrated how control over a powerful technological platform can be leveraged to disadvantage complementary or potentially disruptive technologies.

Relevance to AI compute

The emerging equivalent may involve a technological stack:

GPU → cloud → operating environment → AI framework → model → application

A company controlling multiple layers could potentially disadvantage researchers or competitors operating at another layer.

Principle

Technological ecosystems can produce exclusionary effects even when the conduct does not resemble traditional price-based competition.

11. European Commission v. Google Android

General Court of the European Union

The Google Android litigation involved Google's contractual arrangements concerning Android, search and mobile applications.

The case illustrates how dominance in one technological layer can potentially be leveraged into adjacent markets through contractual restrictions and ecosystem control.

Relevance to research compute

The principle is applicable where access to computational infrastructure is conditioned on participation in a broader technological ecosystem.

Potential concerns include:

• exclusivity;
• preferential distribution;
• contractual restrictions;
• ecosystem lock-in;
• tying;
• foreclosure of competing technologies.

Principle

Dominance in an upstream technological ecosystem can have competitive consequences in neighbouring markets.

12. European Commission v. Microsoft

Microsoft — tying and interoperability litigation

The EU Microsoft proceedings addressed Microsoft's conduct concerning its dominant operating system and related software products.

A major aspect of the case concerned interoperability and Microsoft's control over technical information necessary for competing products.

Relevance to knowledge concentration

Knowledge concentration is not limited to physical compute.

Technical information can itself become an important competitive input.

For AI research this may include:

• model documentation;
• APIs;
• interoperability information;
• technical specifications;
• benchmarking information;
• training-data documentation;
• safety information.

If dominant infrastructure providers strategically restrict interoperability, researchers may become locked into particular technological ecosystems.

Principle

Control over technical interoperability can reinforce technological dominance.

13. IMS Health GmbH & Co. KG v NDC Health

Court of Justice of the European Union

IMS Health concerned copyright-protected information and access to a data structure necessary for competitors to operate effectively in a market.

The case developed the European approach to exceptional circumstances under which refusal to license intellectual property may raise competition-law concerns.

Relevance to research knowledge

Modern scientific competition increasingly depends on combinations of:

data + software + models + compute + intellectual property

A dominant entity controlling an indispensable scientific dataset or technical standard could potentially create competitive problems where competitors cannot realistically operate without access.

Principle

IP rights do not automatically create competition-law liability, but exceptional circumstances can justify intervention.

14. Magill TV Guide

Joined Cases C-241/91 P and C-242/91 P

The European Court recognised circumstances in which refusal to license copyrighted material could constitute an abuse of dominance.

The case is important because it established a restrictive framework for compulsory licensing.

Relevance to AI research

A similar issue could emerge where a dominant research platform controls:

• uniquely valuable datasets;
• scientific metadata;
• proprietary model interfaces;
• essential technical information.

The fact that the information is protected by intellectual property does not automatically resolve the competition question.

Principle

Intellectual-property exclusivity and competition policy may conflict where exceptional circumstances create substantial foreclosure.

15. Bronner v Mediaprint

Case C-7/97

Bronner concerned access to a newspaper distribution system controlled by another undertaking.

The CJEU applied a strict test for compulsory access to infrastructure.

Relevance to research compute

Bronner is particularly useful for analysing whether a large cloud or supercomputing system should be treated as indispensable infrastructure.

The claimant would need to establish more than:

“This infrastructure is cheaper or better.”

The relevant question is closer to:

“Can effective competition realistically exist without access to this infrastructure?”

Principle

Superior infrastructure is not necessarily an essential facility. Genuine indispensability is critical.

16. Implications for Global Research Competition

The cases collectively reveal several important principles.

17. Compute as an "Essential Facility" — Difficult but Important

Not every scarce GPU cluster should be treated as an essential facility.

A legally serious claim would require analysis of:

1. Indispensability

Is the specific computing resource genuinely indispensable?

2. Replicability

Can researchers obtain equivalent resources from:

• another cloud provider;
• national supercomputers;
• universities;
• public research facilities;
• alternative accelerator architectures?

3. Capacity scarcity

Is the relevant resource genuinely capacity-constrained?

4. Competitive foreclosure

Does denial prevent meaningful competition?

5. Legitimate justification

Could allocation restrictions be justified by:

• cybersecurity;
• safety;
• export controls;
• energy constraints;
• research integrity;
• national-security requirements?

18. Public Research Infrastructure

Governments may respond to compute inequality by establishing publicly funded infrastructure.

Possible models include:

National AI compute centres

Universities receive subsidised access to national GPU clusters.

Research-compute grants

Researchers receive computing credits instead of conventional research grants alone.

Shared supercomputing facilities

Multiple institutions share infrastructure.

Open scientific model programmes

Governments fund openly accessible models and datasets.

International compute pools

Multiple states create shared research-computing infrastructure.

These policies can promote competition without necessarily imposing compulsory access obligations on private firms.

19. Competition Between States

Compute inequality also has a geopolitical dimension.

Countries possessing:

• advanced semiconductor industries;
• major cloud providers;
• energy infrastructure;
• supercomputers;
• AI research institutions

can obtain significant advantages in scientific development.

This creates a potential division between:

compute-rich jurisdictions

and

compute-dependent jurisdictions.

The resulting inequality may influence:

• AI research;
• biotechnology;
• climate modelling;
• materials science;
• drug discovery;
• defence technology;
• quantum computing;
• economic forecasting.

20. Subsidies and Competition Neutrality

Public investment in research compute can itself create competition concerns.

Suppose a government provides enormous subsidies to one domestic company for developing AI infrastructure.

Potential consequences include:

• crowding out foreign competitors;
• preferential access;
• state-supported market dominance;
• distortion of cloud markets;
• barriers to international research collaboration.

Therefore, public compute policy should distinguish between:

research infrastructure subsidies

and

commercial competitive subsidies.

Transparent eligibility rules are important.

21. International Competition-Law Coordination

Research compute markets are inherently cross-border.

A single AI research organisation may have:

• chips manufactured in one country;
• cloud infrastructure in another;
• researchers across several countries;
• datasets originating globally;
• customers worldwide.

Consequently, competition authorities may need cooperation concerning:

1. merger control;
2. cloud concentration;
3. semiconductor bottlenecks;
4. data access;
5. AI infrastructure;
6. interoperability;
7. cross-border subsidies;
8. export controls.

22. Risk of Scientific Gatekeeping

The deepest concern is not merely economic.

If compute becomes a prerequisite for frontier research, dominant infrastructure owners may indirectly determine:

which researchers can participate in knowledge production.

This can create scientific gatekeeping.

For example, a research project may be scientifically promising but commercially unattractive. A compute-constrained institution may be unable to undertake it, while a private company can easily run thousands of experiments.

Consequently, market incentives may influence the direction of scientific discovery.

23. Reproducibility Problem

Compute concentration can also undermine scientific reproducibility.

A published result may technically be public while its underlying experiment requires:

• millions of dollars of compute;
• proprietary datasets;
• specialised accelerators;
• private cloud infrastructure.

Other researchers may therefore be unable to reproduce it.

This creates a distinction between:

publication openness

and

reproducibility openness.

Competition policy can become relevant where proprietary infrastructure is deliberately used to prevent competitors from reproducing or challenging technological claims.

24. Research Compute and Merger Control

Mergers can intensify knowledge concentration.

Consider a hypothetical transaction involving:

major cloud provider + AI research laboratory + semiconductor accelerator company.

The combined firm could control:

compute + models + data + distribution + researchers.

Traditional market-share analysis may underestimate the competitive implications.

Merger authorities may therefore need to examine:

• innovation pipelines;
• access to compute;
• researcher mobility;
• model development;
• data access;
• cloud switching;
• accelerator compatibility;
• future technological markets.

25. Remedies

Possible competition-law remedies include:

A. Non-discriminatory access

Require infrastructure providers to apply transparent allocation criteria.

B. Interoperability

Require technical compatibility between competing systems.

C. Cloud portability

Allow researchers to move workloads between providers.

D. Compute-credit programmes

Provide subsidised access for universities and smaller laboratories.

E. Structural remedies

In exceptional cases, separate infrastructure from downstream AI activities.

F. Transparency obligations

Require disclosure of:

• capacity allocation;
• pricing;
• eligibility criteria;
• access restrictions.

G. Merger remedies

Require divestiture or continued independent access to:

• models;
• datasets;
• research teams;
• compute capacity.

26. A Proposed Global Regulatory Framework

A comprehensive framework could operate through five layers.

Layer 1 — Compute measurement

Governments should measure:

• accelerator ownership;
• cloud capacity;
• supercomputer availability;
• research allocation;
• geographic distribution.

Layer 2 — Competition monitoring

Competition authorities should monitor:

• cloud concentration;
• GPU allocation;
• exclusive contracts;
• vertical integration;
• acquisitions.

Layer 3 — Research access

Public-interest research should receive transparent access to major infrastructure.

Layer 4 — Interoperability

Researchers should not be unnecessarily locked into a particular computational ecosystem.

Layer 5 — International coordination

Competition authorities should coordinate investigations involving global compute infrastructure.

27. Key Legal Tension

The fundamental legal tension can be expressed as:

Innovation incentives
vs.
open access to computational infrastructure

Too little intervention may produce:

compute concentration → knowledge concentration → technological dominance.

Too much intervention may produce:

compulsory sharing → reduced investment → lower infrastructure development.

The appropriate legal approach therefore requires targeted intervention rather than automatic compulsory access.

Conclusion

Global research compute inequality represents an emerging intersection of competition law, digital infrastructure regulation, intellectual property, research policy and technological sovereignty.

The most important development is that computational capacity is becoming a productive input into knowledge itself. When compute is highly concentrated, control over infrastructure can translate into control over the ability to conduct frontier research.

The cases involving Terminal Railroad, AT&T, Aspen Skiing, Trinko, Microsoft, Android, IMS Health, Magill and Bronner demonstrate the principal legal tools available: essential-facilities doctrine, refusal-to-deal principles, interoperability obligations, tying analysis, discrimination theories and exceptional access to protected technological resources.

The future regulatory question will therefore not simply be:

“Who has the largest AI model?”

It will increasingly be:

“Who controls the computational infrastructure necessary to create the next generation of knowledge?”

That question makes research compute a potentially significant competition-law and global knowledge-governance issue, particularly where compute concentration combines with control over data, talent, intellectual property and distribution.