The Oligarchy Nobody Planned For

The artificial intelligence market has consolidated with breathtaking speed. By the end of 2025, three companies—Anthropic, OpenAI, and Google—controlled nearly 90 percent of the $37 billion enterprise AI market, according to analysis by Menlo Ventures. Anthropic commands 40 percent, OpenAI holds 27 percent, and Google maintains 21 percent. For an industry that barely existed five years ago, this concentration rivals the most ossified sectors of American capitalism.

What makes this oligopoly particularly remarkable is not its existence, but the velocity with which it formed. The capital requirements alone have become prohibitive. Microsoft, Meta, Alphabet, and Amazon will collectively spend $610 billion on AI infrastructure in 2026—triple what they invested just two years ago. Only a handful of companies possess the financial firepower to enter this market. Even Apple, with its nearly $3 trillion valuation and unmatched ecosystem, chose partnership over competition, opting to work with Google rather than spend hundreds of billions developing its own foundation model.

Yet this consolidation, achieved through what appeared to be natural market forces, now contains the seeds of its own disruption. The three dominant players face a choice that will define the next era of technology: maintain their position as neutral infrastructure providers, or leverage their control to compete directly with their customers. Most have chosen the latter, creating a structural crisis that threatens the entire edifice.

When Your Landlord Becomes Your Competitor

Consider the predicament facing thousands of software companies built on top of these foundation models. Financial technology firms, legal services providers, and enterprise software vendors have spent the past two years constructing sophisticated applications using APIs from OpenAI, Anthropic, or Google. They believed they were building on stable ground. They were wrong.

Anthropic has ventured aggressively into application markets, directly competing with companies like Thomson Reuters that have historically dominated legal and financial services. The pattern is repeating across the industry. When a foundation model company—which profits from licensing its base technology—simultaneously offers finished applications in the same domain, it creates an impossible situation for downstream developers. These application companies rely on model access for their viability, yet they face the constant threat that their suppliers might degrade or eliminate that access entirely.

The threat is not merely theoretical. In an industry where switching costs are high and technical lock-in is substantial, the leverage is asymmetric to the point of absurdity. A foundation model company can observe which applications prove most profitable, then enter that market directly. If an application developer attempts to migrate to a competitor's model, they face retraining their systems, reoptimizing for different API structures, and potentially fragmenting their product. Meanwhile, the original model provider retains the ability to throttle access or adjust pricing in ways that make the application economically unviable.

This dynamic has not escaped notice among regulators and antitrust scholars. The structural similarity to vertical monopolies of previous eras—railroad companies that also owned the goods shipped on their rails, or telecommunications companies that manufactured their own handsets—is unmistakable. Yet the AI market presents novel challenges. The product is information; the barrier to entry is computational capital; and the winners have embedded themselves across entire industries with remarkable speed.

The Spending Spree That Signals Desperation

The staggering capital expenditures flowing into AI infrastructure reveal something curious about the market's current leaders. Despite their dominance, they are spending as if they remain vulnerable.

Meta has committed up to $72 billion annually to AI infrastructure, sums that dwarf its entire historical investment in research and development. Microsoft, despite its strategic partnership with OpenAI, continues to invest heavily in its own AI capabilities. Google, which invented the transformer architecture underlying all modern large language models, faces the peculiar humiliation of being displaced from its own invention—Anthropic and OpenAI have captured greater market share despite Google's substantial head start and technical pedigree.

This spending suggests an underlying anxiety. The consolidation may appear stable from the vantage point of 2026, but each company recognizes the fragility of its position. OpenAI, the market leader just two years ago, has been overtaken by Anthropic, largely through superior execution on safety and alignment issues. Anthropic, despite its 40 percent market share, invests as if challengers lurk perpetually on the horizon. Google, burdened by legacy systems and organizational complexity, spends aggressively to maintain relevance in a market it should dominate.

The deeper anxiety concerns moats. In software, sustainable competitive advantage typically derives from network effects, switching costs, or proprietary data. AI model providers possess none of these in reliable form. Network effects are diffuse and weak—a legal technology company using Claude benefits little from knowing that financial services firms also use Claude. Switching costs exist but are surmountable with engineering effort. Proprietary data, theoretically the strongest moat, proves elusive in a domain where training data comes from web scraping, synthetic generation, and partnerships that can be replicated by competitors.

The Regulation Conundrum Nobody Wants to Solve

Policymakers have begun examining the AI market with the same skepticism they once directed toward tech platforms, and the oligopolistic structure is impossible to ignore. Yet regulating this market presents genuinely novel challenges that traditional antitrust frameworks struggle to address.

The European Union and various national governments have proposed rules governing AI transparency, safety, and bias. The United States has engaged in extensive commentary but limited action. What remains largely absent from regulatory discussion is the structural question: should foundation model providers be permitted to compete in application markets simultaneously? Should they face obligations to provide equal access to all downstream developers?

The answer is not obvious. Preventing vertical integration could stifle innovation—foundation model companies might argue they need direct market feedback to improve their products. Mandating equal access could impose massive compliance burdens and paradoxically entrench incumbents by raising barriers to entry. There is no precedent for regulating a market this nascent, this capital-intensive, and this central to future economic value creation.

What is clear is that the current path leads toward one of two outcomes, neither satisfactory. Either the three incumbent providers maintain their oligopoly through a combination of scale advantages and vertical integration, creating a bottleneck through which all AI-enabled commerce must flow; or continued competition for dominance triggers a regulatory response sufficiently heavy-handed that it dampens innovation entirely. The middle path—a competitive market with reasonable safeguards—remains elusive.

The Endgame Nobody Sees Clearly

The most striking feature of the current moment is the absence of visible strategy from any of the major players. Microsoft maintains its partnership with OpenAI while building independent capabilities, a position that provides optionality but commits to nothing. Google simultaneously presents itself as a responsible regulatory actor while competing ferociously in foundational capabilities. Anthropic grows rapidly but remains dependent on capital from its investors, constrained by a business model that might not generate profits at the scale required to justify its valuation.

Meta's strategy—spending $72 billion annually to position itself as an AI company rather than a social media company—represents perhaps the clearest bet, but its outcome remains uncertain. The company is attempting to transform its identity while demonstrating that its open-source model philosophy can compete against proprietary alternatives. Whether this succeeds depends on factors largely outside Meta's control: the evolution of open-source model capabilities, the decisions of enterprise customers to adopt open models versus proprietary ones, and the regulatory environment that ultimately emerges.

Apple's decision to partner rather than compete deserves closer examination. It suggests that even the most capable technology companies recognize the futility of entering a market so capital-intensive and so crowded. Yet Apple's strategy also means accepting dependence on a competitor for capabilities that will define user experience. In a domain where differentiation increasingly derives from AI quality, outsourcing to a rival is not a position of strength.

The Uncomfortable Truth

The AI market's current structure—dominated by three companies that control nearly 90 percent of enterprise spending—will not persist unchanged. The question is what replaces it.

One possibility is continued consolidation, with one or two survivors dominating the landscape so completely that regulators intervene. Another is fragmentation, driven by the emergence of superior technology, lower-cost alternatives, or successful open-source models that reduce the advantage of proprietary systems. A third is regulatory intervention that forcibly restructures the market, breaking up the leaders or imposing obligations that alter competitive dynamics.

What seems least likely is stability. The spending levels are unsustainable, the tensions between infrastructure provision and application competition are unresolvable within current structures, and the regulatory pressure will intensify as AI capabilities expand and the stakes of AI dominance become impossible to ignore.

The architects of today's AI oligopoly created something powerful but ultimately unstable. They may have already lost the ability to shape what comes next.