The AI Value Capture Divide: Who Really Wins From AI

AI is concentrating value in a small set of companies that control infrastructure, data, distribution, and integration, while most others remain consumers of capability rather than owners of it.

In early 2025, surveys from McKinsey & Company began to show a pattern that has since become more pronounced. While a large majority of organizations reported adopting AI in some form, only a small share indicated that it had meaningfully affected their bottom line. McKinsey’s “State of AI” report found that although adoption had crossed 70–80% in many sectors, only a fraction of companies reported significant cost reductions or revenue gains tied directly to AI deployment.

The survey also found that organizations are beginning to take steps that drive bottom-line impact, including redesigning their workflows as they deploy gen AI and putting senior leaders in critical roles, such as overseeing AI governance. Companies with at least $500 million in annual revenue are changing more quickly than smaller organizations.

At the same time, the companies supplying the infrastructure behind AI have reported a very different trajectory. NVIDIA’s data center revenue more than doubled year-over-year in 2024, driven largely by demand for AI chips used in training and inference workloads. Cloud providers such as Microsoft and Amazon have committed tens of billions of dollars to AI infrastructure, including data centers and specialized compute capacity, as part of their long-term strategy.

The widespread adoption of AI at the application layer is occurring alongside a concentration of value at the infrastructure and platform layers. As more companies integrate AI into their workflows, the demand for compute, models, and cloud services increases, reinforcing the position of the firms that control these layers.

The divergence is therefore structural. Many organizations are becoming more capable through the use of AI tools, but they remain dependent on systems they do not control. The value generated by these tools does not remain evenly distributed across users. It flows toward the layers that enable and scale their use.

The question then is not whether AI creates value, but it is where that value accumulates as the system expands.

Why Capability Does Not Equal Advantage

The speed at which AI tools have spread across organizations has created a surface-level impression of democratization. Models are accessible through APIs, embedded into enterprise software, and increasingly bundled into productivity tools. From marketing teams generating content to developers using AI-assisted coding, the underlying capability is no longer scarce in the way it once was.

What is less visible is how quickly this access begins to neutralize itself. When a capability becomes widely available through shared infrastructure, it tends to reduce differentiation rather than increase it. This pattern has appeared in earlier technology cycles.

Cloud computing, for example, lowered the cost of infrastructure and expanded access to computing power, but it did not create sustained advantage for most firms that adopted it. Instead, it shifted advantage toward those who controlled the platforms through which that infrastructure was delivered.

AI is following a similar trajectory, but at a faster pace. Enterprise adoption data reflects this dynamic. A 2023–2024 survey by McKinsey & Company found that while over 70% of organizations reported using AI in at least one business function, only a much smaller share reported material financial impact, with fewer than a third seeing significant cost reductions or revenue increases tied directly to AI. The gap reflects how that capability is distributed.

At the operational level, the effects are measurable but limited. Studies on tools such as GitHub Copilot have shown productivity improvements of up to 55% in specific coding tasks under controlled conditions. These gains are real, but they are also replicable. Any team using the same tools can achieve similar improvements, which means that efficiency increases across the board without necessarily shifting competitive position.

This creates what might be described as a compression effect. As AI tools reduce the cost and time required to perform certain tasks, they narrow the performance gap between organizations. Tasks that once differentiated between firms start to become standardized, and the baseline for acceptable performance rises.

The structure of AI delivery reinforces this effect. Most organizations are not building models from scratch. They are accessing capabilities through platforms operated by a small number of providers. Whether a company integrates services from Microsoft, builds APIs from OpenAI, or deploys models through Google, the underlying systems are shared across a wide set of users. This shared dependency limits the extent to which any single organization can differentiate purely through access to AI.

The financial data points in the same direction. While enterprise users report incremental gains, the most significant revenue growth has been captured by infrastructure providers. NVIDIA reported data center revenue growth exceeding 200% year-over-year during the peak of the AI expansion, driven by demand for GPUs used in training and inference. It reflects where the economic leverage sits within the system.

There is also a second-order effect that is less discussed. As AI tools become integrated into workflows, they begin to standardize how work is performed. Outputs become more consistent in tone, structure, and approach, particularly in areas such as content generation, coding patterns, and analytical summaries. This standardization can improve efficiency, but it can also reduce variation, making it harder for organizations to differentiate based on execution alone.

The result is a separation between access and advantage. AI can make organizations more capable, but it does not automatically change their position in the system. Advantage depends on factors that are not evenly distributed, control over infrastructure, access to proprietary data, integration into core systems, and the ability to shape how capabilities are delivered at scale.

The gap between widespread AI adoption and limited enterprise value points to a simple conclusion: access to AI is no longer a competitive advantage. As technology becomes widely available, advantage shifts to how deeply AI is connected to proprietary data, workflows, decision-making, and execution. Companies that make these connections can turn AI into an operating capability. Those that do not may adopt the same tools but are unlikely to create the same value.

Where Value Actually Accumulates — Infrastructure, Models, Platforms

The distribution of value in AI does not follow usage. It follows control over the layers that enable usage. At the base of the system sits compute. Training and running modern AI models requires specialized hardware at scale, and this layer is dominated by a small number of suppliers.

The position of NVIDIA illustrates this clearly. Its data center business, driven largely by demand for AI chips, grew at a pace rarely seen in the semiconductor industry, with revenue more than tripling year-over-year during the peak of the AI expansion. This growth is not tied to a single application. It reflects the underlying demand for compute across the entire ecosystem.

Above this layer sits cloud infrastructure, where companies such as Amazon, Microsoft, and Google provide access to the compute required to train and deploy models. These firms are the gatekeepers of scale. Their investments in data centers, networking, and energy infrastructure position them as the primary interface through which most organizations access AI capabilities.

The financial commitments involved are substantial. Microsoft announced plans to invest tens of billions of dollars in AI-enabled data centers, reflecting the long-term importance of infrastructure in this market. These investments are not speculative. They are tied directly to usage, as every API call, model inference, and deployment runs through infrastructure that generates recurring revenue.

The model layer introduces another point of concentration. Developing frontier models requires access to compute, data, and specialized talent at a scale that is difficult for most organizations to replicate. Companies such as OpenAI and Anthropic operate at this layer, but their ability to do so is closely linked to partnerships with infrastructure providers. This creates a system in which control over models is intertwined with control over the resources required to build and run them.

What emerges from this structure is a form of vertical alignment. Infrastructure providers supply compute, host models, and distribute capabilities through APIs and platforms. Model developers rely on these layers, while application-level companies build on top of them. Value flows upward through this stack, accumulating at the points where control is most concentrated.

This pattern is consistent with earlier technology cycles, but it is amplified in AI by the scale of resources required. In cloud computing, infrastructure providers captured significant value, but the barriers to entry were lower. In AI, the combination of compute intensity, data requirements, and ongoing operational costs raises those barriers, making it more difficult for new entrants to compete at the same level.

There is also a reinforcing dynamic at play. As more companies adopt AI, demand for compute and infrastructure increases, which in turn strengthens the position of those who provide it. The growth of the ecosystem feeds back into the layers that control it, creating a cycle in which usage drives concentration rather than dispersion.

Understanding this distribution is essential to understanding how AI reshapes competitive advantage. The ability to use AI is widespread, but the ability to capture the value it generates is not.

The Distribution Layer — Why Platforms Capture the Edge

Control over infrastructure explains where value begins to accumulate. Control over distribution explains how it compounds, often quietly, through the channels that already sit between capability and the end user.

One way to see this is to look at how AI features are being deployed. Rather than launching as standalone products, many of the most widely used capabilities are being folded into existing software ecosystems. Microsoft has integrated Copilot across its productivity suite, including Word, Excel, and Teams, effectively placing AI inside tools that already have hundreds of millions of users. Analysts estimate that Microsoft 365 has more than 345 million paid seats globally, which means any incremental AI feature is distributed at a scale that most standalone products cannot approach.

A similar pattern is visible in search. Google has embedded generative AI into its core search product and workspace tools, layering AI functionality onto services that process billions of queries per day. This matters because distribution has already been solved. AI does not need to acquire users independently as it is inserted into existing behavior.

This changes the economics of competition. When AI is bundled into platforms, it is not priced as a standalone capability. It becomes part of a broader subscription or service offering. Microsoft’s Copilot, for example, has been priced as an add-on to enterprise plans, allowing the company to monetize AI through an existing revenue base rather than building a new one from scratch. This bundling strategy makes it difficult for independent providers to compete on capability alone, because the cost of switching is tied to the entire ecosystem, not just the AI feature.

There is also a measurable advantage in user reach and engagement. OpenAI’s ChatGPT reached over 100 million users within two months of launch, which was itself an indication of demand for standalone AI interfaces. Yet much of the subsequent growth in usage has come through integrations with larger platforms, including enterprise software and cloud services, where AI becomes part of a broader workflow rather than a separate destination.

The distribution advantage extends beyond reach into data. Platforms that sit closest to users capture interaction data at scale, including how queries are structured, how outputs are used, and where systems succeed or fail. This data feeds back into model improvement. The more a system is used within a platform, the more information that the platform gets to refine its behavior. Over time, this creates a compounding effect in which distribution improves capability, and improved capability reinforces distribution.

Industry analysis suggests that this feedback loop is already influencing competitive positioning. A report from Sequoia Capital described the AI market as one where infrastructure and platform layers are likely to capture the majority of economic value, with application-layer companies facing pressure on margins due to competition and dependency on underlying providers. The report highlights how distribution and access to users are critical in determining where value ultimately settles.

This dynamic is not new, but it is intensified in AI. In previous technology cycles, distribution platforms determined which applications reached users. In AI, they also shape how intelligence itself is delivered. The interface becomes the product, and the model becomes part of the platform rather than a separate layer.

The result is a system in which access to users is as important as access to compute. Companies that control distribution can embed AI into workflows, capture usage, and monetize it through existing channels. Those that do not must compete within ecosystems where pricing, visibility, and user access are defined by others. This is where the edge consolidates.

Infrastructure determines who can build at scale. Distribution determines who remains closest to the user. Together, they define how value moves through the system and where it accumulates.

The Enterprise Trap — Why Most Companies Stay Consumers

For most organizations, the adoption of AI does not change their position in the value chain. It improves how work is done, but it does not alter where value is captured. The structure of the system makes this difficult to change.

One constraint is economic. Building AI capabilities at the infrastructure or model level requires capital, compute, and specialized talent at a scale that is out of reach for most firms. The cost of training advanced models has moved into the tens or hundreds of millions of dollars, depending on size and complexity, with ongoing expenses for inference and maintenance. Even companies with substantial resources find it more efficient to access these capabilities through external providers rather than replicate them internally.

This creates a dependency that is structural rather than temporary. Organizations integrate AI through APIs, cloud services, and platform features that are controlled by others. Each interaction, whether generating content, running analysis, or supporting a workflow, relies on infrastructure that sits outside the firm. The cost of this interaction may appear small at the unit level, but it scales with usage, turning operational activity into recurring expenditure tied to external systems.

There is also a technical dimension to this dependency. AI systems are not isolated tools that can be easily swapped or replaced. They are integrated into workflows, connected to data sources, and embedded in decision-making processes. Over time, this integration creates switching costs. Changing providers or rebuilding systems internally requires adjustments to processes, retraining teams, and potential disruption to ongoing operations.

Enterprise adoption patterns reflect this dynamic. Surveys by Deloitte have shown that while a majority of organizations are experimenting with AI, a much smaller proportion have moved into scaled, production-level deployment with measurable impact, in part due to challenges related to integration, governance, and cost management. The gap between experimentation and sustained advantage remains significant.

The nature of competition reinforces the trap. When multiple firms within an industry adopt similar AI tools, the gains in efficiency tend to be shared. A retailer using AI for demand forecasting may improve accuracy, but competitors using comparable systems can achieve similar results. The baseline improves, but relative positioning remains largely unchanged. In this environment, AI becomes a cost of staying competitive rather than a source of differentiation.

There is a parallel here with earlier technology cycles. Enterprise resource planning systems, cloud computing, and digital marketing tools all expanded capability across organizations, but they did not fundamentally redistribute advantage. Instead, they shifted value toward the providers of those systems. AI extends this pattern, but with greater intensity due to the scale of resources involved and the centrality of compute.

Some organizations attempt to move beyond this position by developing proprietary data assets or building specialized applications. These efforts can create localized advantages, particularly in domains where data is unique, or workflows are highly specific. However, they do not eliminate dependence on the underlying layers of infrastructure and models. Even differentiated applications often run on shared systems, which continue to capture a portion of the value generated.

The result is structural asymmetry. Most companies become more capable through the use of AI, but they remain consumers of the system rather than owners of it. Their gains are real, but they are bound by the terms set by those who control the layers beneath them.

This is an enterprise trap. Adoption is widespread; capability improves, and efficiency increases, but the position within the value chain remains largely unchanged. The benefits of AI are distributed, but the economic upside is concentrated.

Conclusion: Why AI Spreads but Value Concentrates

Artificial intelligence is often described through access. More firms can now use capabilities that were once confined to a small set of players. AI is available through APIs, embedded into software, and increasingly built into everyday workflows. At the level of usage, it looks like a broad diffusion of power.

A closer look shows a different structure. Usage is spreading, but control is concentrating. AI runs through a layered system in which compute, models, and distribution carry very different economics. Compute depends on capital, energy, hardware, and infrastructure at scale.

Models depend on sustained investment, data, and iteration. Distribution depends on platforms that already control user access, pricing power, and workflow entry points. These layers reinforce one another, so control in one strengthens position in the others.

That is why value does not disperse evenly as adoption rises. It accumulates most strongly at the layers through which everyone else’s usage flows. Each new deployment, each added workflow, and each increase in demand deepens the advantage of firms sitting closest to infrastructure and distribution. Scale at the edge keeps feeding leverage at the core.

For most companies, AI changes performance as it can improve efficiency, extend capability, and raise the speed and consistency of work. But those gains usually sit inside systems they do not control, on terms they do not define, and through dependencies they cannot easily escape.

So, the real question is no longer whether a company uses AI. It is where that company sits within the system that produces, powers, and distributes it. Firms that control infrastructure, own distribution, or build around proprietary data are positioned to turn AI into durable value. Most others will gain from it, but within limits set elsewhere. AI is spreading widely, but the control over where its value settles is not.