The AI Power Shift: How Intelligence Is Becoming Concentrated

As artificial intelligence advances, the systems that build, fund, and distribute it are consolidating, shaping who can develop it, access it, and ultimately control its direction.

How the Foundations of AI Are Being Built

In September 2023, Amazon announced that it would invest up to $4 billion in Anthropic while committing its cloud infrastructure to support the training and deployment of its models. Around the same period, Microsoft extended its multi-billion-dollar partnership with OpenAI, integrating its models across Azure and embedding them within its enterprise software ecosystem.

These moves reflect a pattern that has been unfolding across the industry, where model development, infrastructure, and distribution are becoming tightly coupled through long-term strategic alignment.

The scale at which these systems are being built has also shifted. Estimates from industry analysis suggest that training frontier models can require investments running into tens or hundreds of millions of dollars, driven by the need for large-scale compute, data pipelines, and sustained operational support.

As the cost base expands, participation at the frontier has become more selective, with fewer organizations able to commit to the resources needed to compete at that level.

The supply of critical infrastructure reinforces this concentration. Advanced AI development depends heavily on specialized GPUs, and over the past few years NVIDIA has become the central provider of this hardware. Demand has risen sharply as companies expand their AI capacity, with reports highlighting how access to these chips has begun to influence which organizations can scale their systems and how quickly they can do so.

When these developments are viewed together, a more structural pattern begins to appear. Artificial intelligence is being built within a system where compute, capital, and distribution are increasingly interconnected, and where access to one layer strengthens control over the others.

Organizations that operate across multiple parts of this stack are able to shape not only the development of the technology, but also how it is accessed, deployed, and integrated into the broader economy.

This structure carries implications that extend beyond the pace of technological progress. Earlier phases of the digital economy evolved in environments where participation was wider at the outset, even as consolidation followed later. Artificial intelligence is developing under conditions where infrastructure requirements, capital intensity, and platform integration are significant from the beginning, which influences who can build at scale and how that capability is distributed.

Within this system, a relatively small group of organizations has come to occupy central positions across multiple layers. Figures such as Sam Altman, Demis Hassabis, Satya Nadella, and Jensen Huang are often seen in discussions about the future of AI, and their presence reflects the roles they play within a system that is becoming more integrated as it scales.

The direction of technological development remains important, and capability continues to advance at a rapid pace. At the same time, the structure through which that progress is being made is becoming more concentrated, shaping access to the technology, and influencing how it is applied. The evolution of artificial intelligence therefore involves both an expansion of capability and a reorganization of control.

The question that follows from this is broader than performance or timelines. It concerns how the distribution of power is changing as these systems are built, and what that means for the organizations, economies, and institutions that depend on them.

Why AI Is Not Decentralizing Like the Internet

Discussions about artificial intelligence often draw implicit comparisons with earlier technological shifts, particularly the rise of the internet, where early openness created space for broad participation before consolidation appeared at later stages.

This sequence shaped expectations about how modern technologies evolve, with the assumption that access expands first, and concentration follows over time. The development of AI is unfolding along a different path, shaped by constraints that were not present in the same way during earlier cycles.

One of the most significant differences lies in the role of infrastructure. The early internet relied on protocols that were widely accessible and relatively inexpensive to adopt, which allowed individuals and smaller organizations to experiment, build, and scale without requiring large upfront investment.

In contrast, modern AI systems depend on specialized hardware, large-scale data processing, and sustained compute capacity, all of which introduce a level of capital intensity that narrows participation at the outset.

The distribution of capability also follows a different pattern. Internet technologies could be deployed independently once infrastructure was in place, allowing new entrants to build services without controlling the underlying layers. AI capabilities are increasingly accessed through platforms that bundle infrastructure, models, and deployment environments together.

Cloud providers play a central role in this process, offering both the compute required to train models and the services through which those models are delivered, which creates a tighter coupling between development and distribution.

This coupling has implications for how competition unfolds. Organizations that control infrastructure are able to support their own model development, while also providing the resources that others depend on. This dual role allows them to influence both supply and access, shaping the ecosystem in ways that extend beyond their own products.

Partnerships between model developers and cloud providers reinforce this structure, aligning incentives across different layers of the stack and reducing the independence that might otherwise exist between them.

The flow of knowledge within the field has also evolved. Earlier phases of the internet benefited from a strong culture of open standards and shared protocols, which allowed innovations to propagate quickly across the ecosystem.

AI research continues to produce open work, but the most advanced systems are often developed within organizations that operate under competitive and strategic constraints. Decisions about what to release, what to restrict, and how to provide access are influenced by considerations that extend beyond research alone.

These factors combine to produce a system in which access to capability is mediated through a smaller set of channels. Participation is not absent, but it is structured differently, with meaningful influence concentrated among organizations that operate across multiple layers of the stack. Smaller players can build on top of these systems, but their ability to shape the underlying direction of the technology is more limited.

The result is an ecosystem that evolves with a different balance between openness and control. Capability continues to expand, and new applications emerge across domains, yet the conditions under which that expansion occurs are shaped by infrastructure, capital, and platform integration from the beginning.

Understanding this structure helps explain why the trajectory of AI feels distinct from earlier waves, even as it draws on many of the same underlying ideas about technological change.

The Big Guns Building the Stack: Where Corporate Power Is Expanding in AI

The concentration of power in artificial intelligence becomes more visible when examined through the structure of the companies operating at its center. What appears on the surface as competition between model providers is, in practice, a system in which a small number of firms are extending their influence across multiple layers of the AI stack, combining infrastructure, model development, and distribution into integrated positions.

Let’s consider how Microsoft has approached this shift. Its partnership with OpenAI is often described in terms of product integration, but the relationship runs deeper. Microsoft provides cloud infrastructure through Azure, supports the training and deployment of models, and embeds those capabilities directly into widely used enterprise tools such as Office and developer environments.

This creates a loop in which infrastructure supports models, and models drive demand back into infrastructure and software ecosystems. The influence of the company extends across all three layers, allowing it to shape both how AI is built and how it is used.

A similar pattern can be observed within Google DeepMind, where research, infrastructure, and product distribution are closely aligned within a single organization. Google’s control over its own compute resources, combined with its integration of AI into search, cloud services, and productivity tools, allows it to operate across the stack without relying on external partners to the same extent. This internal alignment reduces dependency and increases the ability to move quickly from research to deployment.

Amazon occupies a different position, but one that is equally significant. Through its investment in Anthropic and its role as a dominant cloud provider, Amazon has positioned itself as both an infrastructure provider and a key participant in model development. Its strategy reflects an understanding that control over compute and deployment environments can be as influential as control over the models themselves.

Across these cases, the pattern is consistent. Companies are not competing solely at the level of individual products. They are building positions across a stack in which each layer reinforces the others. Control over compute enables model development. Model development drives demand for infrastructure. Distribution channels ensure adoption and embed these systems within existing workflows.

This form of vertical integration alters how competition operates. Smaller firms and independent developers can build applications on top of these systems, but their ability to influence the underlying direction of the technology is constrained by the layers they do not control. Access is available, but it is mediated.

The result is an ecosystem that resembles a set of interconnected platforms rather than a field of independent actors. Innovation continues to occur across different levels, but the foundational layers on which that innovation depends are increasingly concentrated within a small group of organizations.

Understanding this structure helps explain why discussions about AI often return to the same set of companies. Their prominence is not simply a function of visibility or market share. It reflects their position within a system where control is distributed across layers, and where operating across multiple layers provides a level of influence that is difficult to replicate from the outside.

Governments Enter the Race: AI as Strategy, Not Just Technology

As artificial intelligence systems have become more central to economic and technological competition, governments have begun to engage with them in ways that go beyond regulation. The shift is visible in how AI is increasingly treated as infrastructure with strategic importance, rather than as a category of software that can be governed through conventional policy frameworks.

In the United States, export controls on advanced semiconductor technologies have introduced a new layer of constraint on the global development of AI systems. Restrictions on the sale of high-performance chips to China are intended to limit access to the compute required for training advanced models, effectively linking national security considerations to the availability of core infrastructure.

These measures do not operate in isolation. They interact with a supply chain that is already concentrated, reinforcing the role of a small number of companies that produce and control the hardware on which AI development depends.

In Europe, the focus has taken a different form, with the passage of the EU AI Act introducing a regulatory framework that classifies AI systems based on risk and imposes obligations on how they are developed and deployed.

The intent is to shape how AI is integrated into society, addressing concerns around safety, transparency, and accountability. At the same time, the framework reflects an understanding that the deployment of AI systems carries implications that extend beyond individual applications, influencing markets, institutions, and public trust.

On the other hand, China’s approach combines elements of both strategy and control. State-backed initiatives have focused on accelerating domestic AI capabilities while keeping oversight over how these systems are developed and used. Policies aimed at guiding generative AI services, along with broader investment in infrastructure and research, position AI as a component of national development and technological self-sufficiency.

These approaches differ in emphasis, but they converge in one respect. Artificial intelligence is being treated as a domain in which control over infrastructure, capability, and deployment carries strategic weight. Governments are not only responding to the development of AI systems. They are shaping the conditions under which those systems evolve.

This interaction between state and industry adds another layer to the structure already outlined. Companies operate across compute, models, and distribution, while governments influence access to key resources, set regulatory boundaries, and, in some cases, act as participants through funding and policy direction.

The system that emerges is neither purely market-driven nor entirely state-controlled. It is a hybrid in which power is distributed across institutions that operate at different levels but are increasingly interdependent.

The implications of this shift extend beyond competition between firms. They affect how AI capabilities are distributed across regions, how standards are defined, and how the benefits of these systems are allocated. Access to infrastructure, alignment with regulatory frameworks, and proximity to centers of investment all begin to influence who can build and deploy AI at scale.

In this context, the concentration of power in AI cannot be understood solely in terms of corporate strategy. It reflects a broader alignment between technological capability, economic resources, and state-level priorities, each reinforcing the others as the system develops.

What If Power Does Decentralize

The concentration of power in artificial intelligence is not uncontested, and there are credible forces working in the opposite direction. Open-source models, falling compute costs over time, and the possibility of new infrastructure approaches all introduce the potential for a broader distribution of capability. The question is not whether decentralization is possible in principle, but how far it can extend within the constraints already shaping the system.

One of the most visible countercurrents is the growth of open-source AI. Organizations such as Meta have released models that can be downloaded, modified, and deployed without relying on centralized APIs, allowing developers and smaller firms to build applications with greater control over their own systems.

The release of models like LLaMA has expanded access to advanced capabilities and created an ecosystem of tools, fine-tuned variants, and independent deployments that operate outside the largest commercial platforms.

This shift has practical implications. It lowers the barrier to experimentation, enables customization, and reduces dependence on a small number of providers for certain categories of work. For many applications, particularly those that do not require frontier-level performance, open models provide sufficient capability to support real-world use cases.

Over time, improvements in efficiency and optimization could further extend this accessibility, allowing more organizations to operate with a degree of independence.

There are also longer-term factors that could influence the structure of the system. Advances in hardware design, increased competition in chip manufacturing, and the development of more efficient training techniques all have the potential to reduce the cost of building and deploying AI systems. If these trends continue, the concentration driven by high capital requirements may ease, creating space for a wider range of participants.

At the same time, decentralization faces structural limits that are not easily removed. The most advanced models continue to require significant compute resources, and the gap between frontier systems and smaller-scale alternatives remains meaningful in many domains.

Infrastructure at scale involves not only hardware, but also energy, data management, and operational expertise, all of which introduce layers of complexity that extend beyond model architecture.

Distribution also plays a role. Even when models are open, the channels through which they reach users are often controlled by platforms that integrate them into broader ecosystems. Access to users, enterprise adoption, and integration into workflows are shaped by systems that already operate at scale, which means that control over distribution continues to influence how widely different approaches can spread.

This creates a dynamic in which decentralization and concentration coexist. Open systems expand access at the application layer, while the underlying infrastructure and frontier capabilities remain concentrated. Smaller players can build, adapt, and innovate within this space, but the ability to influence the direction of the technology at its core is shaped by resources that are not evenly distributed.

From this perspective, decentralization is not absent. It is layered. It operates within a system where certain forms of access are expanding, while others remain constrained. The balance between these forces will influence how the ecosystem evolves, and whether the current concentration represents a temporary phase or a more persistent structure.

Understanding this interplay is important because it prevents the analysis from settling into a single conclusion. The system is not moving entirely toward concentration or entirely toward openness. It is developing through the interaction of both, with each shaping the limits of the other as technology continues to advance.

Power, Capability, and Control: The Distribution of Intelligence

Across the development of artificial intelligence, attention often gravitates toward capability, what models can do, how quickly they are improving, and where they may lead. That focus is understandable because the visible progress has been substantial. At the same time, a parallel shift has been unfolding in the structure through which that progress is being made, one that is less visible but no less consequential.

The system that supports modern AI, spanning compute, models, and distribution, has developed under conditions that favor scale, integration, and sustained access to resources. Organizations that operate across multiple layers of this stack are able to influence not only the direction of the technology, but also the terms under which it is accessed and deployed.

Governments, through policy and strategic positioning, are shaping the environment in which these systems evolve, linking technological development with broader economic and geopolitical considerations.

Within this landscape, figures such as Sam Altman, Demis Hassabis, Satya Nadella, and Jensen Huang are often seen articulating where artificial intelligence is headed. Their influence extends beyond the systems they help build. It reflects their position within a structure that is becoming more interconnected as it scales, where decisions made at one layer can shape outcomes across others.

The distribution of capability continues to expand, and new applications are emerging across domains. At the same time, the conditions under which that capability is developed remain uneven, shaped by infrastructure, capital, and access to platforms that operate at global scale. This combination creates a system in which progress and concentration advance together, each reinforcing the other in ways that are not always immediately visible.

The trajectory of artificial intelligence will continue to be defined by advances in capability, but it will also be shaped by the structure through which those advances are produced and distributed. Understanding that structure clarifies how influence is exercised, how access is determined, and how the benefits of these systems are likely to be allocated.

The future of AI will be measured not only by what the technology can do, but by whoever has the ability to build it, control it, and decide how it is used.