AI-Driven Arbitrage Systems and Market Structure Centralization

1. Introduction

AI-driven arbitrage systems are automated or semi-autonomous systems that identify and exploit price, liquidity, information, or execution differences across markets. Traditional arbitrage may involve buying an asset where it is cheaper and simultaneously selling it where it is more expensive. AI systems can perform this process continuously by analysing enormous quantities of market data, predicting price movements, routing orders, adjusting strategies, and reacting within milliseconds.

The competition-law concern arises when the same technology moves beyond independent arbitrage and becomes a mechanism through which market information, pricing decisions, execution infrastructure, or access to liquidity becomes concentrated in one platform or a small number of AI-controlled intermediaries.

The central legal question is therefore:

When does AI-enabled arbitrage remain a legitimate efficiency-enhancing activity, and when can control over algorithms, data, execution infrastructure, or market access contribute to exclusion, coordination, or structural centralisation?

There is still relatively limited reported case law dealing expressly with autonomous AI arbitrage. Consequently, existing cases concerning algorithmic pricing, information exchange, digital platforms, hub-and-spoke coordination, high-frequency trading, and market access provide the principal legal framework.

2. Meaning of AI-Driven Arbitrage

An AI-driven arbitrage system may perform five interconnected functions:

1. Market scanning – monitoring multiple exchanges, platforms, dealers or geographic markets.
2. Price prediction – forecasting short-term price movements using machine learning.
3. Opportunity detection – identifying price discrepancies.
4. Automated execution – buying and selling automatically.
5. Continuous learning – modifying trading strategies based upon observed market responses.

For example:

Exchange A: Asset = ₹100
Exchange B: Asset = ₹102

An AI system could simultaneously buy at ₹100 and sell at ₹102, subject to transaction costs and execution risk.

The competition implications become more complicated when the AI system also controls:

• access to market data;
• order-routing;
• liquidity;
• execution priority;
• pricing recommendations;
• competitor information;
• market-making infrastructure;
• trading APIs; or
• the technological infrastructure used by competing traders.

3. Arbitrage Is Not Inherently Anti-Competitive

Arbitrage can produce substantial competitive benefits.

Positive effects

AI arbitrage may:

• eliminate unjustified price differences;
• improve liquidity;
• increase market efficiency;
• improve price discovery;
• reduce transaction costs;
• identify inefficient markets;
• accelerate information incorporation;
• increase cross-market competition;
• reduce opportunities for discriminatory pricing.

The European Commission has expressly recognised that algorithms can generate efficiencies, reduce costs and barriers to entry, and independently monitor competitors' prices.

Therefore, the mere use of AI or automated trading is not sufficient to establish an antitrust violation.

The legal concern arises from the architecture and competitive consequences of the system.

4. How Arbitrage Can Produce Market Structure Centralisation

AI-driven arbitrage may progressively produce a highly concentrated market structure through several mechanisms.

A. Data concentration

Superior AI systems require enormous quantities of:

• real-time prices;
• order-book information;
• transaction data;
• customer behaviour;
• liquidity information;
• historical market data.

A firm possessing the largest dataset can potentially train a more accurate model.

This produces a data-feedback loop:

More users → more data → better AI → better arbitrage → more users → still more data

Eventually, competing firms may find it difficult to reproduce the incumbent's informational advantage.

B. Computational concentration

Advanced arbitrage requires:

• high-performance computing;
• specialised chips;
• low-latency networks;
• cloud infrastructure;
• colocated servers;
• sophisticated models.

High computational costs can create barriers to entry.

Thus, competition may shift from:

Who has the best trading strategy?

to:

Who controls the technological infrastructure necessary to deploy the best strategy?

C. Liquidity concentration

A successful AI arbitrage platform may attract increasing trading volume.

More volume generates:

• deeper liquidity;
• better execution;
• more market data;
• lower transaction costs.

This can create a liquidity network effect.

Competitors may then be unable to obtain comparable execution quality without connecting to the dominant system.

D. Information centralisation

An intermediary may receive information from numerous market participants.

If that information is subsequently incorporated into an AI model, the intermediary effectively becomes an information hub.

This creates competition concerns where the system:

• receives competitors' confidential information;
• combines it into common pricing models;
• distributes recommendations to competing firms; or
• uses aggregated information to disadvantage particular participants.

5. The Hub-and-Spoke Problem

One of the most important competition-law theories is the hub-and-spoke model.

Traditional structure

Competitor A
↓
Common Hub
↑
Competitor B

The hub may be:

• a pricing algorithm;
• exchange;
• software provider;
• marketplace;
• data intermediary;
• trading platform.

The danger is that competitors may cease making genuinely independent competitive decisions because the common technological intermediary coordinates their conduct.

The Supreme Court's decision in Interstate Circuit, Inc. v. United States, 306 U.S. 208 (1939) remains an important foundation for analysing hub-and-spoke coordination. The Court accepted that an anticompetitive agreement could be inferred from coordinated conduct even where direct evidence of an express agreement was difficult to obtain.

AI can effectively become the modern technological "hub."

6. Algorithmic Collusion Versus Independent Parallel Behaviour

This distinction is fundamental.

Lawful independent behaviour

Firm A's AI observes market prices and independently adjusts its price.

Firm B does the same.

Both algorithms may eventually produce similar prices.

That fact alone does not necessarily establish an unlawful agreement.

Potentially unlawful coordination

Firm A and Firm B:

• provide competitively sensitive information to a common AI provider;
• knowingly use the same coordination mechanism;
• configure the algorithm to maintain common prices; or
• agree to follow algorithmic recommendations designed to suppress competition.

Here the algorithm may be evidence or an instrument of coordination.

The European Commission's 2023 horizontal-cooperation guidance specifically recognises that algorithms may monitor existing anticompetitive agreements or facilitate coordination and that "collusion by code" can constitute a cartel where competitors coordinate essential competitive parameters.

7. Major Competition-Law Issues

A. Price coordination

An AI system could continuously monitor competitors and modify prices to avoid aggressive competition.

This may produce:

• price convergence;
• reduced discounting;
• supracompetitive prices;
• rapid retaliation against price reductions.

The difficult question is whether such behaviour results from independent adaptation or a legally cognisable form of coordination.

B. Common algorithm problem

If competing companies use the same third-party AI system, the provider may effectively become a common pricing intermediary.

This is particularly problematic where the system receives non-public information from competing businesses.

The risk is not simply the software itself; it is the relationship between the users, the information supplied, and the algorithm's function.

C. Exclusion of smaller competitors

A dominant AI arbitrage platform might:

• provide superior execution to preferred customers;
• restrict API access;
• impose discriminatory latency;
• prioritise affiliated traders;
• refuse interoperability;
• withhold essential market data;
• impose discriminatory fees.

These practices can transform an efficiency technology into an access-control mechanism.

D. Self-preferencing

A platform controlling arbitrage infrastructure might operate its own trading or brokerage service.

It could potentially use its infrastructural position to:

1. observe competitors;
2. identify profitable opportunities;
3. execute through its own affiliate first; and
4. disadvantage independent traders.

This creates a vertical conflict between the platform's role as infrastructure provider and its role as market participant.

8. Important Case Laws

1. United States v. David Topkins

United States v. David Topkins, No. 3:15-cr-00201 (N.D. Cal., 2015) is one of the most important early algorithmic-pricing cases.

Online poster sellers allegedly agreed to coordinate prices and used specific pricing algorithms to implement the agreement. Topkins pleaded guilty to horizontal price fixing. The DOJ treated the algorithm as the technological mechanism used to implement the underlying conspiracy.

Principle

An algorithm does not immunise an otherwise unlawful agreement.

Relevance to AI arbitrage

If competing AI systems are deliberately configured pursuant to an agreement to avoid price competition, the fact that execution occurs automatically does not eliminate antitrust liability.

2. United States v. Aston and Trod Ltd.

United States v. Daniel William Aston and Trod Ltd., No. 3:15-cr-00419 (N.D. Cal., 2015) involved online sellers of posters using pricing algorithms.

The DOJ alleged that competitors discussed prices and agreed to adopt specific algorithms so that their prices would remain coordinated.

Principle

Automated pricing can be the implementation mechanism of a cartel.

Relevance

The case demonstrates the distinction between:

Independent algorithmic optimisation
and
algorithmic implementation of coordinated strategy.

3. Eturas UAB and Others v Lietuvos Respublikos konkurencijos taryba

Case C-74/14, Eturas, Court of Justice of the European Union (2016).

Several travel agencies used a common computerised booking system. The system administrator communicated a measure that automatically restricted discounts available through the system.

The CJEU considered whether the use of a common electronic system and knowledge of the restrictive mechanism could constitute a concerted practice.

Principle

Electronic systems can provide evidence of coordination; competition law does not become inapplicable merely because the mechanism is technological.

Relevance

An AI arbitrage platform shared by competing traders could similarly become relevant evidence of coordination where users knowingly participate in a common restrictive mechanism.

4. T-Mobile Netherlands BV and Others v Raad van bestuur van de Nederlandse Mededingingsautoriteit

Case C-8/08, T-Mobile Netherlands (2009).

The CJEU held that a concerted practice can arise even from a single meeting, and recognised the importance of exchanges of competitively sensitive information in determining whether independent competitive decision-making has been compromised.

Principle

Competition law protects the requirement that undertakings independently determine their market conduct.

Relevance

In an AI environment, the relevant "contact" may involve:

• shared data feeds;
• common algorithmic instructions;
• common software settings;
• coordinated model parameters.

The technological form does not necessarily change the underlying competition-law analysis.

5. Interstate Circuit, Inc. v. United States

306 U.S. 208 (1939).

The U.S. Supreme Court addressed coordinated conduct involving multiple distributors and a common commercial strategy. The case became an important foundation for modern hub-and-spoke conspiracy analysis.

Principle

An anticompetitive arrangement may sometimes be established through the structure and circumstances of coordinated conduct rather than a conventional written agreement.

Relevance to AI

A common AI provider may potentially function as the modern technological hub connecting otherwise competing firms.

The analogy must nevertheless be established factually; simply using the same software is not automatically equivalent to a cartel.

6. Samir Agrawal v. Competition Commission of India

The Indian Ola/Uber algorithmic-pricing litigation is particularly important.

The allegation was that cab aggregators' algorithms effectively fixed prices accepted by drivers and therefore facilitated coordination between drivers.

The CCI rejected the initial information for lack of a prima facie agreement, and the matter proceeded through appellate litigation. The Competition Appellate Tribunal's reasoning emphasised that algorithmically determined fares, by themselves, did not establish the necessary collusion or hub-and-spoke arrangement.

Principle

Algorithmic price determination ≠ automatically algorithmic cartelisation.

There must be an appropriate legal basis connecting the algorithmic mechanism with prohibited coordination.

Importance

This is particularly significant for India because it prevents competition law from treating every centrally calculated AI price as inherently unlawful.

7. Manoj K. Sheth v. Competition Commission of India / NSE

The Indian securities-market litigation concerning NSE co-location and algorithmic/high-frequency trading provides another important perspective.

The case concerned allegations that preferential co-location access could give certain market participants faster access to market information and therefore competitive advantages.

The 2026 appellate decision discusses algorithmic and high-frequency trading, co-location, latency, access to market information, and the potential competitive implications of unequal technological access.

Principle

In technologically intensive markets, access to infrastructure and information can itself become a competition issue.

Relevance to AI arbitrage

If an exchange or dominant infrastructure provider gives selected AI traders:

• faster data;
• superior connectivity;
• lower latency;
• preferential execution; or
• better access to order information,

the issue may move beyond ordinary technological competition into potential exclusionary or discriminatory conduct.

8. United States and State Plaintiffs v. RealPage, Inc.

The RealPage litigation is one of the most significant recent developments in algorithmic competition law.

The DOJ alleged that competing landlords supplied non-public, competitively sensitive information to RealPage's pricing software and that the system facilitated alignment of rental prices. The DOJ subsequently obtained proposed and final settlements involving RealPage and participating landlords, including restrictions concerning the use of competitors' sensitive information and algorithmic pricing practices.

Principle

An algorithmic intermediary may create serious competition concerns where it:

• aggregates competitors' sensitive information;
• uses that information to generate pricing recommendations;
• reduces independent decision-making; or
• structurally strengthens its position as the central information intermediary.

Relevance

This is highly relevant to AI-driven arbitrage because an AI platform with access to the trading strategies and data of numerous competitors could potentially become a centralised market-information infrastructure.

9. The Centralisation Feedback Loop

One of the most significant structural risks is the following:

More participants

↓

More transactions

↓

More proprietary data

↓

Better AI model

↓

More accurate arbitrage

↓

Better execution

↓

More users

↓

Higher market share

↓

Greater data and liquidity advantage

This can create a self-reinforcing competitive moat.

The problem becomes especially significant if rivals cannot obtain equivalent:

• data;
• liquidity;
• computing power;
• APIs;
• execution speed;
• training datasets.

10. AI Arbitrage and Essential-Facility-Type Concerns

A dominant AI infrastructure provider may become strategically important where competitors depend upon it for:

• real-time market data;
• liquidity;
• order routing;
• clearing interfaces;
• trading APIs;
• execution infrastructure;
• benchmark information.

A refusal to provide access is not automatically unlawful. Competition law generally requires additional conditions such as dominance, indispensability, exclusionary effect, lack of objective justification, or equivalent requirements depending upon the jurisdiction.

The important point is that technological infrastructure can become competitively significant in exactly the same way as physical infrastructure.

11. Information Asymmetry

AI arbitrage systems can create two opposing situations.

Situation 1 — Efficiency-enhancing information

AI discovers previously dispersed information and incorporates it into prices.

This may improve:

• price discovery;
• liquidity;
• market efficiency.

Situation 2 — Strategic information concentration

A dominant intermediary obtains confidential information from many competitors and uses it to improve its own trading strategy.

This can create:

information asymmetry → superior AI → greater market share → more information → even greater AI advantage.

This second structure presents greater competition concerns.

12. Algorithmic Arbitrage and Market Manipulation

Competition law must also be distinguished from securities/market-abuse regulation.

An AI arbitrage system could potentially engage in conduct involving:

• spoofing;
• layering;
• wash trading;
• artificial liquidity;
• quote manipulation;
• latency exploitation;
• misleading signals.

Such conduct may implicate securities legislation even where competition law is not the principal legal framework.

Therefore, an AI trading system can simultaneously raise questions under:

1. Competition law
2. Securities law
3. Market-abuse regulation
4. Data-protection law
5. Consumer law
6. Financial-sector regulation

13. Tacit Algorithmic Coordination

The most difficult theoretical issue occurs when:

• A uses AI independently;
• B uses AI independently;
• neither communicates directly;
• both algorithms monitor the market;
• each learns that aggressive price competition produces retaliation;
• both gradually converge on higher prices.

Economic research has demonstrated that reinforcement-learning algorithms can, under certain experimental conditions, learn supracompetitive pricing strategies without explicit communication.

Recent economic work likewise examines models in which algorithm selection and repeated interaction can generate supracompetitive outcomes without direct communication.

Legal difficulty

Economic coordination does not automatically equal legal coordination.

Competition law in many jurisdictions still requires some legally relevant form of:

• agreement;
• communication;
• concerted practice;
• common understanding; or
• exclusionary conduct.

Consequently:

Economic collusion and legally provable collusion are not necessarily identical.

14. AI as a Market-Structure Controller

The most important conceptual shift is that AI can move from being merely a trading tool to becoming a market-structure controller.

Stage 1 — Tool

AI identifies arbitrage opportunities.

Stage 2 — Optimiser

AI decides how and when to execute trades.

Stage 3 — Intermediary

Other firms depend upon the AI platform for execution.

Stage 4 — Information hub

The platform receives information from numerous competitors.

Stage 5 — Gatekeeper

The platform controls access to data, liquidity or execution.

Stage 6 — Market shaper

The platform's decisions influence the competitive conditions under which all other participants operate.

This final stage presents the greatest structural competition concerns.

15. Competition-Law Framework

A. Agreement or concerted practice

Authorities may examine whether AI systems facilitate:

• price fixing;
• output coordination;
• information exchange;
• market allocation;
• bid coordination;
• common pricing rules.

Relevant principles arise from Topkins, Trod, Eturas and T-Mobile Netherlands.

B. Abuse of dominance

A dominant AI arbitrage infrastructure provider could potentially face scrutiny for:

• discriminatory access;
• refusal to supply;
• self-preferencing;
• tying;
• exclusionary API restrictions;
• discriminatory execution;
• exploitative data practices.

C. Information exchange

Particular attention should be paid to whether an AI system receives:

• current prices;
• future pricing intentions;
• inventory information;
• margins;
• trading strategies;
• non-public demand information.

The more competitively sensitive the information, the greater the potential competition concern.

D. Market definition

Authorities may need to distinguish between:

• AI arbitrage software;
• trading infrastructure;
• market-data services;
• execution services;
• brokerage;
• exchange services;
• liquidity provision;
• financial information services.

A firm may be dominant in one layer without necessarily being dominant in the entire financial ecosystem.

16. Remedies

Potential remedies can include:

Structural remedies

• divestiture;
• separation of trading and infrastructure operations;
• interoperability obligations.

Behavioural remedies

• non-discriminatory API access;
• prohibition on competitor-data use;
• firewalls;
• transparency requirements;
• restrictions on sensitive-data aggregation.

Algorithmic remedies

• independent algorithm audits;
• model governance;
• logging requirements;
• explainability regarding critical pricing functions;
• restrictions on real-time competitor-data inputs.

Data remedies

• data minimisation;
• purpose limitation;
• separation of customer data;
• restrictions on using competitors' confidential information for model training.

The DOJ's RealPage settlements illustrate the increasing use of restrictions directed specifically at algorithmic systems and competitively sensitive information.

17. Compliance Framework for AI Arbitrage Firms

An AI arbitrage operator should maintain:

1. Independent pricing protocols
2. Competitor-data restrictions
3. Documented model objectives
4. Algorithmic audit trails
5. Access-control mechanisms
6. Information firewalls
7. Human escalation procedures
8. Testing for discriminatory execution
9. Monitoring for unexplained price convergence
10. Periodic competition-law review

Particular care is required where the same AI provider serves multiple competing firms.

18. Key Distinction

19. Emerging Legal Test

A useful analytical framework for regulators is:

AI Arbitrage Competition Test

Step 1 — Identify the AI function

Is it:

• arbitrage;
• pricing;
• execution;
• market making;
• information aggregation;
• infrastructure?

Step 2 — Identify market power

Who controls:

• data;
• liquidity;
• computing;
• APIs;
• execution;
• customers?

Step 3 — Examine information flows

Does the system receive confidential information from competitors?

Step 4 — Examine decision independence

Are firms making independent decisions or following a common algorithmic mechanism?

Step 5 — Examine exclusion

Can rivals obtain comparable access?

Step 6 — Examine effects

Has the system produced:

• foreclosure;
• price coordination;
• reduced liquidity;
• discriminatory access;
• increased entry barriers;
• market concentration?

Step 7 — Examine efficiencies

Are there legitimate benefits such as:

• lower costs;
• improved liquidity;
• faster execution;
• better price discovery?

Step 8 — Select proportionate remedies

The objective should be to preserve legitimate AI-driven efficiency while preventing the technological infrastructure from becoming a mechanism for suppressing competition.

20. Conclusion

AI-driven arbitrage is not inherently anti-competitive. Its traditional function—detecting and eliminating price discrepancies—can improve liquidity, price discovery and market efficiency.

The principal competition-law risk emerges when arbitrage technology becomes centralised market infrastructure.

The most significant risks are:

• concentration of market data;
• concentration of liquidity;
• common algorithmic decision-making;
• exchange or API gatekeeping;
• discriminatory execution;
• use of competitors' sensitive information;
• algorithmic coordination;
• self-preferencing;
• technological barriers to entry; and
• feedback loops that reinforce an incumbent's market position.

The cases of Topkins, Trod, Eturas, T-Mobile Netherlands, Interstate Circuit, Samir Agrawal, the NSE co-location litigation, and RealPage collectively demonstrate an important principle: competition law focuses on the competitive relationship and the conduct facilitated by technology, rather than treating the algorithm itself as either inherently lawful or inherently unlawful.