The AI Duopoly Is Not a Market Failure. It Is the Market

If two companies control nearly nine-tenths of a market, standard antitrust intuition says something has gone wrong. The market is failing. Concentration has throttled competition. Regulators should act.

That reading is tidy. It is also wrong—or at least incomplete—for the artificial intelligence sector as it exists in mid-2026.

According to data reported by The Information and circulated across startup investment channels on 19 May 2026, Anthropic and OpenAI together account for 89% of revenue across the 34 most mature companies selling AI applications or models. Meanwhile, Polymarket odds data from the same date show a 64% implied probability that Anthropic holds the top AI model position by the end of June. The two figures are different metrics—revenue share versus frontier capability—but they point in the same direction: two companies have become the gravitational centre of an entire industry.

That is not a glitch. It is the architecture.

The winner-takes-most logic

AI development does not behave like consumer software. The latter rewards distribution and network effects; the former rewards inference scale, proprietary training data, and the capital intensity required to push frontier capability forward. Those inputs are not evenly distributed. They flow toward organisations that have already demonstrated they can ship at frontier scale, which means they flow toward the same two organisations that have already done so.

This is not new. Cloud infrastructure consolidated around AWS, Azure, and GCP. Search consolidated around Google. Social media consolidated around Meta and, for a time, Twitter. In each case, the winner's advantage compounded: more users generated more training signal, which produced better products, which attracted more users. The AI sector is running the same loop at a faster cycle time.

The 89% revenue figure is not evidence that Anthropic and OpenAI gamed a system. It is evidence that they built the systems others are building on top of. When your product is the substrate, revenue concentration follows almost by definition.

Why this time might still be different

The counter-argument has merit. AI is not social media. The technology is dual-use, infrastructure-critical, and increasingly embedded in national security planning. Concentration at the model layer creates single points of failure that consumer internet consolidation did not.

If Anthropic or OpenAI suffers a catastrophic failure—an alignment breakdown, a security breach at the model权重 level, a regulatory shutdown—the downstream effects cascade through every company in that 89%. Startups that built their entire product on one provider's API have no alternative ready. The resilience argument for competition is stronger here than it was for, say, the search market.

There is also the question of national competition. China has DeepSeek, Zhipu AI, and Baidu's Ernie. The European AI ecosystem, despite significant investment, has not produced a frontier-scale competitor. The United States' current dominance is partly a function of first-mover advantage and partly a function of the regulatory environment allowing rapid capital aggregation. If that dominance becomes structural and permanent, it reshapes the geopolitical AI map in ways that go beyond market economics.

The Polymarket odds—64% favouring Anthropic specifically—reflect market confidence in a single firm's trajectory, not confidence in the ecosystem's resilience. Those are not the same bet.

The governance gap

What is striking about the current moment is how little of the concentration debate is happening in public policy terms versus investment terms. The market is pricing Anthropic's lead. Venture investors are routing capital toward the duopoly's orbit. The regulatory conversation, where it exists at all, remains fixated on whether these companies are engaging in traditional antitrust violations—which, under current definitions, they often are not.

The harder question is whether the antitrust framework, designed for a world of widget manufacturers and regional market power, is the right tool for an industry where the "widgets" are also the infrastructure. Regulating Anthropic and OpenAI as utilities would require a conceptual leap that neither US nor European competition law has yet made. Regulating them as platforms creates its own complications, since they are simultaneously buyers (of compute, talent, data) and sellers.

The 89% revenue share makes the structural question unavoidable. At some level of concentration, the market-correcting mechanisms that antitrust assumes—new entrants, price competition, innovation pressure—stop working. The data from 19 May 2026 suggests the AI industry may already be past that threshold, or dangerously close to it.

What the duopoly means, concretely

For enterprise buyers, the current landscape means dependency on two vendors with divergent safety philosophies, governance structures, and commercial incentives. Anthropic's Constitutional AI approach and OpenAI's more commercial posture are not interchangeable. Choosing one is a strategic decision, not merely a procurement decision.

For startups, the picture is bifurcated. Companies building on top of the frontier models occupy a category The Information's data confirms: they are generating the remaining 11% of sector revenue. That is not a small market—AI-adjacent spending is large enough that 11% represents billions annually—but it is a market that exists in the shadow of the two providers it depends on. Pricing changes, API access restrictions, or capability gaps at the model layer land immediately on every downstream builder.

For regulators, the structural question has arrived before the doctrinal tools to answer it. The EU AI Act addresses risk classification and compliance requirements. It does not address what to do when two companies constitute the industry's operating system. That gap will need to be closed, one way or another.

The concentration is real. The market is not failing in the narrow sense—revenue is being generated, products are being shipped, capability is advancing. What the 89% figure reveals is that the AI industry has built its foundation on an extraordinarily narrow base. Whether that is a feature or a vulnerability depends entirely on what you believe the next eighteen months will demand of it.

This article was structured around market-structure and revenue-concentration data rather than the wire framing of individual company announcements.

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

• https://t.me/producthunt/11334
• https://t.me/venture/29438