The business debate about artificial intelligence often treats model performance as the main measure of competition. A new OECD study, released on 30 July 2026, suggests that the more consequential contest is broader. It includes who controls innovation, computing infrastructure, distribution, customer relationships and the routes through which start-ups reach scale.
The paper draws on multiple firm-level datasets and makes an important distinction between companies that develop AI and those that use it. It finds no significant association between adoption of non-generative AI and increased market power. Exposure to generative AI may create opportunities for smaller firms, even as companies with stronger existing capabilities retain advantages.
That is not a simple concentration story. The OECD also finds that concentration in AI innovation is correlated with greater sales concentration, while AI-related patenting is associated with faster growth in mark-ups, particularly in information and communications technology. The start-up ecosystem remains active and attracts substantial venture capital, but young companies are frequently acquired by large incumbents.
The emerging picture is therefore dynamic but uneven: entry and experimentation continue, while value can still accumulate around established positions.
Competition operates across a stack
An AI product depends on more than a model. It may require specialised chips, cloud capacity, data, technical talent, application software and access to users. A company can face vigorous competition at one layer while remaining dependent on a small number of providers at another.
This helps explain why partnerships and acquisitions matter. They can give a start-up the capital, compute and distribution needed to compete. They can also tighten links between critical inputs and customer-facing services, making it harder for an independent rival to reach the same market. The UK Competition and Markets Authority has previously identified three related risks in foundation-model markets: restricted access to essential inputs, incumbents using existing positions to distort choice, and partnerships reinforcing power across the value chain.
These outcomes are not inevitable. But they mean that counting model launches or funded start-ups is an incomplete measure of market health. The durability of competition depends on whether firms can obtain inputs on workable terms, reach customers and remain viable without being absorbed.
Switching is becoming a strategic capability
For buyers, the practical question is not whether several AI vendors exist. It is whether changing vendors is technically and economically realistic. Proprietary interfaces, tightly coupled cloud services, accumulated prompts and evaluation data, custom workflows and staff training can create substantial switching costs.
Procurement teams should therefore test portability before deployment. Contracts can address access to logs and generated assets, export formats, service termination, model substitution and the ownership of fine-tuning or evaluation work. Technical teams can separate business rules from a provider-specific interface and maintain benchmark sets that allow alternatives to be assessed consistently.
A multi-provider architecture is not always necessary or economical. Redundancy has costs, and a single integrated service may be the best choice for a bounded workload. The objective is not to avoid commitment; it is to understand which commitments are reversible and which create structural dependence.
Patents are a signal, not the whole scoreboard
The OECD’s link between AI patenting and mark-up growth deserves careful interpretation. Correlation does not prove that patents caused higher prices or weaker competition. Patents can also reflect genuine innovation and the ability to turn research into valuable products.
Still, the finding gives boards and policymakers a reason to watch the combination of intellectual property, data advantages, infrastructure control and distribution. Any one element may be contestable. Their combination can be harder to challenge, especially when scale improves the product and attracts still more users.
Competition monitoring must also distinguish AI developers from adopters. When a retailer, manufacturer or professional-services firm uses AI to reduce costs or improve quality, the effect may intensify competition in its own market. When AI itself is the product, control of the technology and its inputs can produce a different pattern. Broad claims that AI will either democratise every sector or entrench every incumbent obscure this distinction.
The next advantage is institutional
Companies cannot determine the structure of global AI markets, but they can improve their bargaining position. That means mapping dependencies across compute, models, data and distribution; giving procurement teams enough technical expertise to evaluate lock-in; and treating interoperability as an operating requirement rather than a compliance phrase.
For governments, the priority is continued observation grounded in firm-level evidence. Merger review, cloud-market oversight and rules for dominant digital platforms may all affect the AI landscape, but interventions should target demonstrated bottlenecks rather than assume that every large partnership is harmful. Poorly calibrated restrictions could deny smaller firms the investment and reach they need.
The OECD evidence is an early measurement of a market still changing quickly. Its central lesson is nevertheless durable: innovation and concentration can advance at the same time. The winners will not be determined only by who builds the most capable model. They will also be shaped by who controls the surrounding system—and whether customers and challengers retain credible choices.
Featured photograph: Helpameout via Wikimedia Commons, licensed under CC BY-SA 3.0.
