To unleash AI innovation, stop model providers from picking the winners

Artificial intelligence (AI) promises to find its way into every corner of our digital lives. The headlines focus on ever-bigger models, but the future will be determined by applications built on those models. Who builds those applications—the ones that will reshape medicine, education, software development, and mundane back-end enterprise services—has not yet been decided. Whether future applications will be determined by oligopolists or entrepreneurs depends on who controls the infrastructure those applications need.

Three companies—Anthropic, OpenAI, and Google—have established themselves as the dominant foundation model builders. Application developers build their tools atop those models. Total payments for such access reached over $37 billion in 2025, tripling in a single year, according to Menlo Ventures.

These AI models have become the essential infrastructure of the 21st century. In the startup incubator Y Combinator’s summer 2025 cohort, for instance, nine out of 10 companies were AI-native, meaning they were built from the ground up with AI. Nearly all were dependent on the three dominant companies for the models underpinning their offerings.

A startup building a coding agent or medical diagnostics tool writes code that sends requests to an AI model, and the model presents the results. The startup does not build its own model, which can cost hundreds of millions of dollars; instead, it buys access to a model. This may be efficient, but it exposes a profound competitive deficiency.

The problem is that the companies that control access to those foundation models are also developing their own applications. Anthropic sells access to Claude while marketing Claude Code and Claude Cowork, its own coding agent and enterprise tool. Thousands of third-party AI application developers rely on OpenAI, while it has begun a spree of acquiring AI applications developers across coding, personal finance, and health tech. Google has already embedded its Gemini models in search, Gmail, Maps, and YouTube and is also expanding its applications through acquisition.

When model providers build, buy, and invest in applications, they end up competing with the startups that depend on them. The fact that those companies build applications is not the problem, but it becomes a problem when they use their platform to limit or control the activities of others. Such a conflict is an invitation to abuse.

We have already seen it happen. Last year, “vibe coding”—using AI to write software without the need to understand coding itself—took Silicon Valley by storm. Startups were using the major models to develop AI coding agents in competition with the model builders themselves. End users were the beneficiaries of that good, old-fashioned American competition.

But, in mid-2025, Anthropic cut off Windsurf—another popular coding agent—from its ability to access Claude after reports surfaced that OpenAI might acquire the company. An Anthropic cofounder was blunt about why: it would be “odd,” he said, to sell Claude to a company entering a rival’s orbit. A dominant supplier exercised a kill switch against a downstream competitor.

More recently, SpaceX, which acquired xAI and offers its Grok foundation model for third-party developers, signaled that it will likely acquire Cursor, the leading independent AI coding company. While Cursor’s coding tool currently lets users pick among many underlying models, it and xAI will have an incentive to preference one another once they’re under one corporate structure.

The pattern is sadly familiar. Internet service providers blocked and throttled competing voice, file sharing, and video streaming services. Amazon deprioritized third-party sellers. Google demoted competitors in search results. Each time, control of essential infrastructure was used to tilt adjacent markets toward the oligopolists. Each time, competition-driving innovation was suffocated, and end users were made worse off.

Policymakers have dealt with these issues before—requiring railroads, then telegraph and telephone companies, to provide nondiscriminatory access. AI models are similarly essential infrastructure for the AI economy. The same principle should apply: Foundation model providers should be barred from unjust discrimination in pricing, speed, or quality of service. They should not be able to cut off a customer because it competes in other areas. They should not be allowed to throttle competitors while fast-tracking their own applications. And they should not use data generated by those customers or based on surveilling their customers to build copycat products.

Policy levers exist on both sides of the Atlantic. In the U.S., Congress and state legislatures can enact neutrality rules for AI foundation models, as we have each proposed. In Europe, the Digital Markets Act (DMA) already prohibits self-preferencing, tying, and discriminatory platform access—a framework that maps directly onto AI’s emerging abuses in principle, although the European Commission recently decided not to expand DMA coverage to generative AI services.

Are Gmail’s AI overviews the best that AI can do for email? Is Claude Code the final word in software development? We should want new companies to enter these spaces and compete with innovative new offerings. They cannot do so if the firms supplying their essential infrastructure are also their competitors capable of defining how they use the models and able to pull the plug at will.

It is a simple question of whether oligopolistic gatekeepers or competition is the best route to the future of inventive AI applications.