Artificial intelligence (AI) promises to permeate every corner of our digital lives. While the headlines focus on increasingly large models, the future will be determined by the applications built on those models. It is not yet decided who will build the applications that will reshape healthcare, education, software development, and everyday back-end enterprise services. Whether future applications are decided by oligopolists or entrepreneurs depends on who controls the infrastructure the applications require.
Anthropic, OpenAI, and Google have established themselves as leading foundational model builders. Application developers build tools based on these models. According to Menlo Ventures, total payments for such access will reach more than $37 billion in 2025, a threefold increase in one year.
These AI models have become essential infrastructure of the 21st century. For example, in startup incubator Y Combinator’s summer 2025 cohort, 9 out of 10 companies were AI natives, meaning they were built from the ground up using AI. Almost all companies relied on three powerful companies for the model that underpinned their products.
Startups building coding agents and medical diagnostic tools write code that sends requests to AI models, which then display results. The startup won’t build its own model, which can cost hundreds of millions of dollars. Instead, purchase access to the model. While this may be efficient, it exposes a serious competitive flaw.
The problem is that the companies that control access to these underlying models are also developing their own applications. Anthropic sells access to Claude while marketing its own coding agency and enterprise tools, Claude Code and Claude Cowork. While thousands of third-party AI application developers rely on OpenAI, OpenAI is starting to attract large numbers of AI application developers in coding, personal finance, and health technology. Google has already built the Gemini model into Search, Gmail, Maps, and YouTube, and is expanding its applications through acquisitions.
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 a problem, but it becomes a problem when they use their platforms to limit or control the activities of others. Such conflicts are an invitation to abuse.
We’ve already seen it happen. Last year, “vibe coding,” which uses AI to create software without the need to understand the coding itself, took Silicon Valley by storm. Startups were using leading models to develop AI coding agents in competition with the model builders themselves. End users benefited from good old American competition.
However, in mid-2025, Anthropic cut off access to another popular coding agency, Windsurf’s Claude, after reports surfaced that OpenAI might acquire the company. Anthropic’s co-founder spoke candidly about why. Selling Claude to a company moving into a rival’s orbit is “bizarre,” he said. A powerful supplier has exercised a kill switch on a downstream competitor.
SpaceX, which recently acquired xAI and is offering its Grok-based model to third-party developers, has indicated that it is likely to acquire Cursor, a leading independent AI coding company. Currently, Cursor’s coding tools allow users to choose between many underlying models, but once Cursor and xAI are brought under one corporate structure, there will be an incentive to prioritize each other.
This pattern is sadly familiar. Internet service providers blocked and restricted competing audio, file sharing, and video streaming services. Amazon has lowered the priority of third-party sellers. Google demoted competitors in search results. In each case, control of critical infrastructure was used to tilt adjacent markets toward the oligarchs. Each time, competitive innovation was stifled and end-users were left worse off.
Policymakers have grappled with these issues before, requiring railroads and then telegraph and telephone companies to provide nondiscriminatory access. AI models are likewise essential infrastructure for the AI economy. The same principles should apply. This means that foundation model providers should be prohibited from unfairly discriminating in pricing, speed, or quality of service. Don’t cut off customers just because you’re competing in other areas. You shouldn’t be allowed to throttle your competitors while accelerating your own applications. And data generated by those customers, or based on their monitoring, should not be used to build copycat products.
Policy instruments exist on both sides of the Atlantic. In the United States, Congress and state legislatures can enact neutrality rules for AI-based models, as we have each proposed. In Europe, the Digital Markets Act (DMA) already prohibits access to self-prioritizing, pegging and discriminatory platforms, a framework that in principle directly addresses new abuses of AI, but the European Commission recently decided not to extend the scope of the DMA to generative AI services.
Is Gmail’s AI Overview the best thing 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 products. You can’t do that if the company providing your critical infrastructure is also a competitor who can define how your models are used and plug in at will.
It’s a simple question: is oligopolistic gatekeeping or competition the best path to the future of inventive AI applications?
