Signal
An AI model can get cheaper while the advantage around it gets harder to copy.
A new OECD working paper finds early evidence of both forces. Smaller firms may gain access to tools that once demanded more capital and technical infrastructure. Firms with stronger people, data, workflows, distribution, and financing may still be better placed to turn the same access into durable gains.
The barrier moved. It did not disappear.
What changed
On July 30, the OECD released Competition in the Age of AI: Initial Evidence from Microdata. This paper separates AI users from AI developers and generative AI from earlier, task-specific systems. That matters because each group faces a different market.
For firms using non-generative AI in France and Portugal from 2011 through 2022, the study found no systematic increase in its measures of market power. AI adopters were generally larger and more productive, but their markups were not substantially higher than those of non-users.
But the generative-AI evidence is earlier and thinner. Using the occupations inside Portuguese firms as a proxy for potential exposure, the paper finds that small and young firms may be well positioned to use generative AI. It also finds that companies with stronger human capital and complementary assets may be better able to capture the benefits.
Meanwhile, the view changes again when the paper looks at AI development. Concentration in AI patenting was positively correlated with sales concentration in the studied industries. Firms with AI-related patents also showed an additional markup premium mainly in ICT sectors. That premium disappeared in the specification using firm fixed effects, which makes the result useful for monitoring and too weak for a causal verdict.
That same tension appears in the start-up layer. AI companies continue to enter, and venture funding has flowed heavily toward generative-AI firms. Incumbents also acquire AI start-ups frequently. Those deals can fund entry, give founders an exit, and move technology into wider use. They can also remove a future competitor before it has time to grow.
Why it matters
Model access is the easiest part of this market to see. A price drops. A new open model arrives. An API gives a tiny company capabilities that once required a large technical team.
But access and productive use are separate events.
A business still has to find a useful problem, reorganize work, train people, connect data, check the output, reach customers, and absorb the cost of mistakes. A larger incumbent may already own those surrounding capabilities. A smaller company may move faster or discover a market the incumbent missed. The model price tells us little about which one will keep the gain.
That shifts the practical competition question. After access broadens, where does advantage accumulate?
Look at the capability stack around the model: skilled workers, proprietary data, customer relationships, integrations, switching costs, capital, and the ability to buy a promising company. These assets may become more important as the model itself becomes easier to obtain.
This is also why AI users and AI developers should stay in separate columns. A retailer applying an AI tool faces a different set of costs and advantages from a company building models, cloud infrastructure, or AI software. A broad claim that “AI increases competition” or “AI creates monopoly” flattens several markets into one sentence.
The OECD paper gives us a more useful operating frame. Cheap tools can widen the starting gate while the race still favors firms with stronger support systems.
What this does not prove
The main adoption data cover non-generative AI in two countries through 2022. They do not describe the current generative-AI market.
The Portugal analysis measures occupational exposure as a proxy for potential adoption. It does not show which firms used generative AI well or what happened afterward. The patent and concentration findings are correlations with more than one plausible direction. The start-up sample covers venture-backed or patenting firms rather than every AI company.
The authors call the evidence mixed and preliminary. Their proxies are imperfect, recent generative-AI activity is only partly visible, and some competition effects may take years to appear. This card is an orientation tool, not business, legal, competition-policy, financial, or investment advice.
What to watch
- Observed adoption: firm-level data showing who uses generative AI, for which tasks, and with what result.
- Capability gaps: whether gains cluster around firm size, skilled staff, existing data, or the ability to redesign work.
- Acquisitions: what happens to products, teams, customers, and competing roadmaps after an incumbent buys an AI start-up.
- Switching costs: whether a company can move its workflow to another provider without rebuilding the surrounding system.
- Sector differences: whether the ICT patent-and-markup pattern appears elsewhere as adoption matures.
The next useful measure is not how many firms can reach the model. It is who can turn that access into an advantage, who can keep it, and who gets absorbed along the way.