Competition in AI infrastructure by Marc Bourreau

Competition in Artificial Intelligence Infrastructure

Marc Bourreau, Telecom Paris & CERRE

OECD Rountable – 5 December 2025

Introduction

• Artificial intelligence (AI)

• Crucial for competitiveness in tech and non-tech sectors

• … but this requires AI to be

• Widely available and affordable

• High quality

• Open to innovation

• Concerns about access to key inputs across the AI stack:

• Talent, data, advanced chips (accelerated compute)…

AI chips

• A highly concentrated market, dominated by Nvidia

• Strong economies of scale and scope (R&D, manufacturing, know-how, …)

• Complementarity between Nvidia’s GPUs and its CUDA software platform

indirect network effects

 strong tendency toward market concentration

• Yet, a contestable market?

• New AI chips under development by other chip manufacturers, hyperscalers, and AI developers

• Room for differentiation?

• Differentiated chips (in terms of cost, energy consumption, …) developed by smaller players

Concerns in AI chip market

• Potential competition concerns

• Excessive prices

• Supply restrictions

• Unfair contractual terms

• Discriminatory practices

• Policy intervention?

• The market is highly concentrated but remains very dynamic

• Key trade-off: competition for the market vs. competition in the market

• Broad, heavy-handed intervention seems premature at this stage

Accelerated compute

• Access to accelerated compute for AI developers:

• Either through advanced chips and/or cloud computing services

• Cloud computing market: oligopoly + fringe

• Hyperscalers (Amazon, Microsoft, Google)

• Fringe of general cloud service providers and smaller specialized providers

Concerns in cloud AI market

• Barriers to switching accelerated compute providers

• Technical barriers

• As by-products of differentiation or innovation

• Learning effects

• Commercial barriers

• Egress fees

• Credits or discounts

 interoperability requirements (e.g., Data Act 2023 in Europe) as possible remedy?

• Partnerships between compute providers and AI developers

• But usually non-exclusive

• Vertical integration along the value chain

• Risks of tying, self-preferencing, practices to discourage switching…

• Justifies a certain degree of regulatory scrutiny

Conclusion

• Strong economies of scale and scope  strong tendency toward concentration in AI chip and accelerated compute markets

• But also a very high pace of innovation  strong dynamic competition

• Overarching market intervention seems premature

• … but scrutiny across the AI stack remains important

• Interoperability/portability provisions may facilitate switching

• … but should be applied carefully to avoid freezing innovation