By
Stuart Kerr, Technology Correspondent, LiveAIWire

Five companies control the infrastructure on which the global AI
economy runs. Their cloud platforms host the training clusters, the inference
endpoints, the data pipelines, and the application layers that billions of
people and millions of businesses depend on daily. The concentration is not
incidental: it reflects the economics of AI development, where the capital
requirements for frontier model training are so large that only a handful of
organisations can sustain them. The question is whether this concentration
constitutes a new form of digital feudalism — and whether AI will entrench
it or eventually disrupt it.

The feudal analogy has limits, but its proponents make a serious
point. In a feudal system, those who control the land extract rent from those
who work it. In the AI economy, those who control the compute, the data, and
the frontier models extract value from the application builders, the
enterprise customers, and the end users who depend on their infrastructure.
The terms of that dependency are set by the platform owners, not by those who
depend on them.

How Power Concentrates in the AI Stack

The AI technology stack has a natural tendency toward
concentration at each layer. At the infrastructure level, the capital
requirements for large-scale GPU clusters favour hyperscale cloud providers.
At the model level, the data and compute requirements for frontier model
training favour organisations with existing data advantages and deep capital
reserves. At the application level, network effects favour platforms that
already have large user bases, because user data improves model performance
which attracts more users.

The result is a stack in which competitive moats compound across
layers. A company that leads at the infrastructure level can offer preferential
pricing to allied model developers. A model developer with the
best-performing foundation model can attract the most application builders.
An application platform with the largest user base generates the most
feedback data to improve its underlying models. Each advantage reinforces the
others, creating dynamics that are difficult for new entrants to overcome
without equivalent capital.

Research from the Stanford AI
Index has consistently documented the concentration of AI research
and development capacity among a small number of institutions, both corporate
and national. The most capable frontier models are produced by a group of
companies that can be counted on two hands. The resources required to join
that group are growing, not shrinking, as model scale continues to
increase.

Small Players, Dependent Ecosystems

The application layer of the AI economy is more diverse than the
infrastructure and model layers, but diversity does not necessarily mean
independence. A startup building an AI-powered legal tool on top of a major
foundation model API is dependent on the pricing, terms of service, and
continued availability decisions of the model provider. If the provider
raises API prices, changes its content policies, or discontinues a model
version, the startup’s product is affected without recourse. The dependency
is structural, not incidental.

This dependency dynamic is not unique to AI — web businesses have
long been dependent on search engine algorithms and social media platform
policies — but AI amplifies it. The foundation model is not just a
distribution channel; it is a core component of the product. When the API
changes, the product changes. The startup has limited ability to switch
providers quickly because the transition cost involves re-engineering
integrations, retraining fine-tuned models, and re-evaluating performance on
the new platform.

What this means for anyone building on AI infrastructure: the
terms on which you depend on platform providers deserve the same scrutiny as
any other critical supplier relationship. The risk of platform dependency is
real and the history of technology platforms suggests that terms tend to
tighten as market position solidifies. The decentralised
AI models being explored through blockchain and federated
architectures represent one response to this dependency risk,
though they face significant technical challenges relative to the performance
of centralised systems.

Open Source as a Counterweight

The open-source AI movement represents the most significant
structural counterweight to platform concentration. Meta’s release of the
Llama model family, Mistral’s open-weight models, and the broader ecosystem
of open-source models available through Hugging Face have created genuine
alternatives to proprietary API dependence for many use cases. An
organisation that can run an open-weight model on its own infrastructure is
not dependent on a hyperscaler’s pricing decisions for that
capability.

The counterweight has limits. Open-source models currently lag the
performance of frontier proprietary models on many benchmarks, particularly
for complex reasoning tasks. Running competitive open-weight models requires
significant GPU infrastructure that many organisations cannot afford at the
scale where they become competitive. And the open-source models that exist
today were trained on data and with compute that was itself concentrated —
the open availability of the model weights does not change the concentration
that produced them.

Research from the Electronic
Frontier Foundation on AI and digital rights has examined open
source as a democratic counterweight to AI concentration, noting that genuine
openness requires not just model weight availability but transparency about
training data, evaluation methodologies, and the governance of the
organisations that produce and maintain models. By those broader criteria,
the open-source AI landscape is more mixed than model weight availability
alone suggests.

Regulatory Responses and Their Limits

Antitrust regulators on both sides of the Atlantic have begun
examining AI market concentration. The UK Competition and Markets Authority
published a foundational model review in 2024 identifying concentration risks
in the AI stack and recommending measures to ensure that the market develops
competitively. The EU’s AI Act and Digital Markets Act together provide a
framework for addressing the most egregious forms of self-preferencing and
platform lock-in, though enforcement is in early stages.

The fundamental challenge for competition regulators is that
market concentration in AI is partly an artefact of genuine scale economies
rather than anticompetitive conduct. Prohibiting concentration that reflects
legitimate efficiency advantages is difficult to justify legally and may
reduce the investment incentives that have produced the AI capabilities being
regulated. The policy goal is ensuring that the benefits of those
capabilities are accessible to a broad range of users and that the market
remains open to competitive entry — objectives that require nuanced
intervention rather than structural remedies designed for earlier technology
markets.

The parallel with the
hidden infrastructure dependencies that underpin the AI economy is
direct: the power relationships embedded in AI infrastructure are not visible
to most users, and their consequences for innovation, pricing, and access are
felt long before they become the subject of regulatory
attention.

Breaking the Web or Reinforcing It

The honest answer to whether AI will break or reinforce digital
power concentration is that current trajectories reinforce it, and the forces
that might disrupt those trajectories are present but not yet dominant.
Open-source models are improving. Specialised hardware is becoming more
accessible. Regulatory attention is increasing. National AI strategies in
Europe, India, and elsewhere are investing in AI infrastructure that is not
dependent on US hyperscalers.

Whether those countervailing forces are sufficient to produce a
genuinely competitive AI economy, or whether the compounding advantages of
the current leaders prove durable enough to consolidate a two or three-player
global AI market, will be determined in the next five years. The decisions
being made now about open-source investment, regulatory intervention, and
national AI infrastructure will shape the answer. The stakes are not just
commercial: who controls AI infrastructure increasingly determines who
controls the economic and informational environment in which everyone else
operates.

The competition between centralised and distributed AI power
structures is also playing out at the national level. China’s approach to AI
development — state-directed investment in national champions, data
localisation requirements, and restricted access to foreign AI services —
represents one model for national AI sovereignty. The EU’s approach of
regulatory standards and strategic investment in European AI infrastructure
represents another. The US approach of private sector leadership with
evolving regulatory oversight represents a third. None of these models has
yet demonstrated that it can combine the innovation dynamism of competitive
markets with the public accountability that the concentration of AI power
requires. The answer to that challenge, when it comes, will shape the global
AI governance architecture for decades. For a deeper examination of who
controls the most powerful AI systems and under what governance,
the parallel with frontier model development is
instructive.

About the Author

Stuart
Kerr is a technology correspondent at LiveAIWire, covering artificial
intelligence, emerging technologies, and their impact on society and
industry.