Why Faster AI Diffusion Does Not Guarantee More Competition
This series examines the relationship between artificial-intelligence diffusion and competition: who adopts AI, who develops it, who can turn access into a durable advantage, and whether new firms can still challenge established ones. The immediate question is not whether AI will spread. It is whether the benefits of spreading AI will be distributed widely enough to keep markets contestable. Readers will gain a way to distinguish a technology becoming common from economic power becoming dispersed, and to judge what the early evidence actually supports.
The apparent contradiction is easy to state. A language model, a forecasting service, or an image-generation tool can be rented by a small business without building a research laboratory. That access may help a young firm produce credible work, test a product, or serve a niche that was previously too expensive to reach. Yet the same technology may reward firms that already have better data, skilled employees, computing access, managerial capacity, distribution, and the money to redesign their operations. A tool can become widely available while the capability to use it well remains concentrated.
That distinction matters because competition is not measured by adoption alone. A firm may use AI to lower costs and pass the savings to customers. It may use the same reduction in cost to improve quality, expand its share, and eventually make rivals less viable. A high market share may reflect a firm outperforming competitors, or it may coexist with durable barriers and high markups. The outcome depends on the mechanism and on the time horizon, not on the presence of AI by itself.
AI changes the conditions of competition, not just the speed of work
AI matters to market structure because it changes the relationship between inputs, capabilities, and scale. Non-generative systems such as anomaly detection, recommendation, and forecasting tools often require data, technical integration, and workers who can interpret their outputs. Generative AI can produce text, images, software, or other content across a much wider range of activities. Both can reduce the marginal cost of particular tasks, but neither removes the need to decide where a system fits, how its output is checked, and how work is reorganized around it.
There are therefore two opposing pathways.
The first is a contestability pathway. If an AI service supplies useful capability at a low entry cost, a smaller or younger firm can reach an acceptable level of quality without reproducing every asset of an incumbent. GenAI may help a business develop a product, tailor an offering, search for information, or communicate with customers more efficiently. Faster experimentation can expose underserved niches. In this pathway, AI increases the number of firms that can try to compete and makes established firms respond to new offers.
The second is an entrenchment pathway. Effective use may depend on complementary assets that are not rented with the model: proprietary data, high-quality workflows, specialized staff, managerial attention, customer relationships, and the ability to finance experimentation before returns arrive. Established firms often possess more of these assets. They may use AI to improve a product, differentiate it, gather more data, or lower costs while retaining the advantages that made them large in the first place. If data, talent, or critical computing resources are concentrated, the upstream advantage can reinforce downstream market power.
These pathways can operate at the same time. An AI system can lower the technical barrier to entry while raising the organizational barrier to effective adoption. A firm that has access to the same model as its competitor may still be unable to integrate it into pricing, product development, service delivery, or compliance. The relevant question is not simply “Who has AI?” but “Who can convert AI access into a repeatable capability, and who can prevent others from doing the same?”
The link between productivity and market power is also not one-directional. If AI lowers a firm's marginal cost, competition may push that firm to reduce prices or improve quality. If rivals cannot match the investment, the same cost advantage may increase market share. A firm that faces weak competitive pressure may retain more of the gain as a higher markup. Market concentration can likewise arise because a firm has outperformed rivals, or because fewer firms remain able to challenge it. The observed outcome is therefore a signal to investigate, not a verdict about the health of competition.
Several features of AI make that investigation especially important. Developing a customized system can involve high fixed or sunk costs even when using it later is inexpensive. Repeated use can improve predictions or workflows through learning and new data, creating an advantage that is difficult for a newcomer to reproduce. Versatile systems may also generate economies of scope: one firm's investment can support applications across many products or functions. At the provider level, a growing ecosystem of integrations can attract more users, while the cost of moving embedded data, processes, and applications to another provider creates switching costs. These mechanisms do not guarantee concentration, but they explain why an early lead can become more valuable as diffusion proceeds.
The distinction between upstream and downstream competition sets the scope for the evidence that follows. Upstream developers may compete over models, computing resources, data, and talent; downstream users and application developers compete in particular industries using those inputs. A downstream market can become more contestable even while an upstream input is concentrated, if the input is accessible on reasonable terms. Conversely, a seemingly open downstream tool may reinforce incumbents if only firms with the right skills and assets can use it productively. This series therefore treats the AI ecosystem as connected layers rather than assuming that competition in one layer settles the outcome in another.
A map of the AI ecosystem
The evidence is easier to interpret if two distinctions are kept separate. The first distinguishes AI users from AI developers. Users purchase or apply AI in their own operations; developers build AI systems or commercial solutions. A company that develops a tool only for internal use belongs principally on the user side of the competition question, because the relevant issue is how the tool changes its core business. A company supplying AI to other firms raises a different question about control over technology, inputs, or customer dependence.
The second distinction is between generative and non-generative AI. Combining the two dimensions produces four useful categories.
- Generative AI users apply versatile systems to tasks such as drafting, coding, customer service, analysis, or product development. Their competitive effect depends heavily on human supervision, complementary skills, and organizational change. The technology may help an entrant reach customers, or help an incumbent extend its lead.
- Generative AI developers include firms building foundation models as well as firms creating smaller or more specialized systems. Frontier development involves large fixed costs, data, computing power, and scarce talent. These conditions can support intense competition among leading developers while still making entry difficult and encouraging vertical integration.
- Non-generative AI users apply predictive or analytical systems to specific operational problems. Cloud services can reduce the fixed infrastructure needed to adopt them, but data quality, process readiness, and technical expertise remain important constraints.
- Non-generative AI developers range from small specialized providers to large platforms offering standardized and bundled services. Tailored solutions can be accessible, while dependence on a provider, switching costs, and bundled infrastructure can reduce competitive pressure over time.
This map is not a claim that every firm fits neatly into one box. It is a safeguard against moving an observation from one part of the ecosystem to another without checking the mechanism. Evidence about a foundation-model developer cannot be used as direct evidence about a retail firm using a chatbot. Evidence about a firm's AI patent cannot automatically show that its customers face higher prices. The actor, technology, and position in the value chain matter.
The organizational layer matters just as much. Firm-level adoption records whether an enterprise reports formal AI use. Functional deployment asks where inside the enterprise the technology is used. Worker-task measures ask whether employees use AI in particular activities. These layers can diverge: AI may run as background infrastructure without daily worker interaction, or employees may experiment with external tools before a firm formally adopts them. A headline adoption rate therefore says little about how deeply AI has entered production or whether it can generate a durable advantage.
How to read the early evidence
The studies considered in this series use different windows onto the same question. The Census Bureau's business survey follows firm adoption, functions, worker tasks, organizational investment, and reported outcomes in the United States. The OECD combines firm surveys from France and Portugal with occupational exposure measures, patents, financial data, market-concentration measures, and start-up records. These sources complement one another, but their numbers are not interchangeable.
The early picture is consequently mixed rather than uniform. The Census evidence shows that firm adoption remains concentrated in larger and knowledge-intensive businesses, while worker-task use can appear without formal firm adoption. It also shows that current task effects are more often augmentative than substitutive, even as organizational investment is more closely associated with employment declines than worker-task breadth is. This suggests that a firm can gain performance from AI before the technology produces a large net change in headcount.
The OECD's evidence on earlier, mainly non-generative firm use in France and Portugal does not show a systematic increase in markup growth among adopters. At the same time, GenAI exposure appears easier to exploit for firms with stronger human capital and complementary assets, and concentration in AI-related patenting is positively correlated with concentration in selected product markets. The AI start-up ecosystem is dynamic and well funded, yet AI start-ups—especially those focused on GenAI—are frequently acquired by large incumbents. None of these observations, taken alone, settles the future of competition.
The correct reading rule is therefore four-part. First identify the actor: user, developer, entrant, incumbent, or platform. Then identify the layer: firm, function, task, innovation, or market. Next identify the relevant outcome: productivity, sales, market share, markup, patent concentration, entry, or acquisition. Finally ask what the study design can establish. A correlation may reveal a mechanism worth monitoring without proving that AI caused a market to become concentrated.
That discipline keeps the main question in view. Fast diffusion can still produce concentrated markets when access spreads faster than the complementary capabilities needed to use it, when critical inputs and learning accumulate among a small group, or when successful challengers are absorbed before they become independent rivals. It can also expand competition when lower costs, better tools, and open entry outweigh those forces. The next step is to look inside firms and measure how adoption actually travels across functions and tasks; without that view, the competitive meaning of “AI diffusion” remains too coarse to judge.