Redistributed

A handful of companies control frontier AI development. OpenAI, Google DeepMind, Anthropic, Meta — and increasingly few others — determine what foundation models exist, what they can do, and on what terms the rest of the world gets to use them. This concentration is not incidental. It follows from the economics of the technology: training frontier models requires billions in compute, proprietary datasets at scale, and engineering talent that a few employers have locked up.

The question is whether competition law can do anything about it. The answer, so far, is: not much — and not because regulators aren't trying.

The Concentration Problem

The numbers tell a stark story. Microsoft has invested over $13 billion in OpenAI and hosts its models exclusively on Azure. Amazon has committed up to $4 billion in Anthropic. Google both builds its own frontier models through DeepMind and runs the cloud infrastructure that many AI companies depend on. Meta trains and releases large models while leveraging its social media monopoly for data and distribution.

These are not conventional market positions. They represent vertical integration at a scale that antitrust law has rarely confronted: the same companies control the compute (cloud infrastructure), the models (foundation model development), and the applications (consumer and enterprise products). When Google embeds Gemini into Search, Gmail, Docs, and Cloud, it is leveraging an existing monopoly to establish dominance in a new market — the textbook antitrust concern. But the mechanism is more tightly integrated, and the market boundaries more ambiguous, than traditional competition analysis can easily handle.

The FTC and DOJ have noticed. The FTC under Chair Lina Khan opened inquiries into the Microsoft-OpenAI relationship, examining whether a $13 billion investment plus exclusive cloud hosting constitutes an acquisition in all but name. The DOJ's antitrust case against Google, which secured a landmark ruling in 2024 finding that Google maintains an illegal monopoly in search, has obvious implications for Google's AI strategy — particularly its bundling of AI features into products that already dominate their markets. The FTC has also scrutinized Amazon's investment in Anthropic under similar logic.

These efforts are genuine, and they represent regulators engaging seriously with novel market structures. But scrutiny is not remedy. And the deeper problem is structural: competition law was built for a different kind of market.

Why Traditional Antitrust Fails

Antitrust law in the United States, since Robert Bork's The Antitrust Paradox reshaped it in the 1980s, has been organized around a single principle: consumer harm, measured primarily through prices. A merger is anticompetitive if it will raise prices. A monopoly is illegal if it harms consumers through higher costs or reduced output.

Foundation models break this framework in several ways.

The price problem. Many AI products are offered to consumers for free. Google Search is free. ChatGPT has a free tier. Meta's Llama models are open-weight. If consumer harm means higher prices, and prices are zero, where is the harm? This is the same challenge that antitrust faced with social media platforms, and the same challenge it has largely failed to resolve.

The harm is real but indirect: reduced competition means fewer alternatives, less innovation from independent developers, and more power concentrated in companies that monetize through data extraction, advertising, and enterprise lock-in rather than direct consumer pricing. As Tim Wu argues in The Curse of Bigness, the consumer welfare standard systematically ignores harms to competition itself — harms to workers, suppliers, potential competitors, and the structure of markets.

The market definition problem. Antitrust cases begin by defining the relevant market. Is the market "AI models"? "AI APIs"? "Search"? "Productivity software"? "Cloud computing"? The answer determines everything. If the market is "AI models broadly," then the existence of open-source alternatives suggests competition is healthy. If the market is "frontier AI models capable of state-of-the-art reasoning," then three or four companies control virtually all of it.

Google's lawyers will argue that AI-enhanced Search competes with ChatGPT, Perplexity, and every other information retrieval tool — therefore the market is competitive. An antitrust enforcer might counter that Google's bundling of AI into a monopoly product (Search) forecloses competition in AI-powered search specifically. Neither framing is obviously wrong, which is the problem. Market definition has become a litigation tool rather than an analytical one.

Vertical integration obscures the harm. When the same company runs the cloud infrastructure, trains the model, and sells the application, anticompetitive behavior becomes invisible. Google does not need to refuse access to competitors — it simply gives itself preferential access to its own compute, data, and distribution channels. Amazon hosts Anthropic's models on AWS while competing with those same models through its own Alexa and Bedrock offerings. The conflicts of interest are structural, not episodic.

Lina Khan, in her influential 2017 law review article "Amazon's Antitrust Paradox," argued that platform companies use vertical integration to cross-subsidize competitive efforts, disadvantage rivals who depend on their platforms, and extract value from both sides of a market. The same logic applies, with greater force, to AI. The companies providing the compute are also the companies training the models are also the companies selling the products. The entire stack is captive.

The Kill Zone

The venture capital industry has a term for it: the "kill zone." This is the area around a dominant platform's core products where startups cannot survive. Build a feature that a major AI lab could replicate? Your funding dries up, because investors know that Google or Microsoft can simply copy it and bundle it into their existing products at no marginal cost.

The kill zone is not theoretical. When OpenAI launched its Code Interpreter (now Advanced Data Analysis), it immediately threatened dozens of data analysis startups. When Google integrated AI summarization into Search, it undermined the business model of every AI-powered search startup. When Microsoft embedded Copilot into Office 365, it made independent AI writing assistants redundant for most enterprise customers.

This is the mechanism through which concentration suppresses innovation. It is not that the dominant companies innovate less — they may innovate more, by some measures. It is that they determine the direction of innovation, and that direction serves their existing business models. Independent AI development — models optimized for different values, different use cases, different communities — is structurally disadvantaged.

Sai Vipra and Anton Korinek, in their Brookings Institution analysis of market concentration in foundation models, document how this dynamic operates specifically in AI. They find that concentration raises barriers to entry (compute costs alone exclude most potential competitors), suppresses innovation that threatens incumbent business models, and gives a small number of companies disproportionate influence over the trajectory of AI development. When four or five firms control the technology, they also control what the technology becomes.

Historical Parallels

We have been here before.

AT&T controlled American telecommunications as a vertically integrated monopoly for most of the twentieth century. It owned the network, manufactured the equipment, and provided the service. The 1984 breakup — structural separation of local and long-distance service — unleashed decades of innovation. The internet as we know it emerged from the competitive long-distance market and the open network interfaces that the breakup required.

Microsoft's antitrust case in the late 1990s addressed bundling: using Windows dominance to foreclose competition in web browsers. The case did not result in a breakup, but the consent decree and the scrutiny itself created space for Firefox, then Chrome, then the modern web. The Department of Justice's theory — that a monopolist cannot use dominance in one market to eliminate competition in adjacent markets — is directly applicable to Google's AI integration.

The current Big Tech cases continue this trajectory. The DOJ's Google search monopoly ruling in 2024 found that Google's exclusive default search agreements (paying Apple $20 billion annually) constituted illegal monopoly maintenance. The remedy phase, still underway, could reshape how AI-powered search reaches consumers.

But each historical case took years — often decades — to resolve. AT&T's case began in 1974 and settled in 1982. Microsoft's case ran from 1998 to 2001. The Google search case started in 2020. AI markets are consolidating now.

What Redistribution Requires

The redistribution stakes are direct. Concentrated AI markets mean that the enormous value AI generates — through productivity gains, automation, new capabilities — flows to a small number of shareholders at a small number of companies. As of early 2026, the "Magnificent Seven" tech companies account for a staggering share of US stock market capitalization, and their AI investments are a primary driver.

Competition would distribute value more widely. More competitors means lower API prices for developers, more diverse models serving different communities, more negotiating power for workers and data providers, and more innovation from companies that are not constrained by protecting an advertising or cloud monopoly.

This is not abstract. When there are three providers of frontier AI APIs, they set prices. When there are thirty, the market sets prices. When one company controls the dominant AI-powered search engine, it decides what information reaches users. When ten companies compete, users have choices. Competition is a redistribution mechanism — perhaps the most powerful one that markets provide.

Proposed Remedies

Several approaches could address AI market concentration, each with different implications.

Structural separation would prohibit companies from simultaneously controlling compute infrastructure, foundation models, and consumer applications. This is the AT&T model: if you run the cloud, you cannot also train and sell the models that run on it. The advantage is clarity. The disadvantage is that vertical integration creates genuine efficiencies, and separation might slow some forms of innovation.

Interoperability mandates would require dominant AI platforms to support open interfaces, allowing third-party models and applications to integrate with their ecosystems. The EU's Digital Markets Act takes this approach for designated "gatekeeper" platforms, and its extension to AI services would impose obligations on companies like Google and Microsoft to open their AI ecosystems to competitors.

Open standards and public alternatives would reduce dependence on private infrastructure. Public compute facilities — like the US National AI Research Resource proposed in bipartisan legislation — would lower barriers to entry. Open model standards would prevent lock-in. Public foundation models, developed outside the commercial incentive structure, would provide alternatives to corporate AI.

Aggressive merger enforcement would scrutinize the investment structures that currently evade traditional acquisition review. If a $13 billion investment plus exclusive hosting is not an acquisition, what is? The FTC's examination of the Microsoft-OpenAI relationship tests whether existing merger law can reach these arrangements, or whether new legislation is needed.

The Regulatory Divide

The approach to AI competition differs sharply across jurisdictions. The EU's Digital Markets Act, which designates large platforms as "gatekeepers" and imposes ex ante obligations — interoperability, data portability, prohibition of self-preferencing — offers a structural model. If AI services are designated under the DMA, dominant providers would face obligations before harm occurs, rather than after years of litigation.

In the United States, the Khan FTC pursued a more aggressive enforcement posture than any recent predecessor, challenging mergers and investigating AI investments. But enforcement depends on political will, and the antitrust philosophy of any given administration determines whether these tools are used. Industry-friendly alternatives emphasize innovation over competition, arguing that heavy-handed enforcement will drive AI development overseas and that market forces will eventually produce competition.

The historical record suggests otherwise. Markets with high barriers to entry, strong network effects, and entrenched platform power do not self-correct toward competition. They consolidate until intervention forces them open. AI markets exhibit all of these features in concentrated form.

What Should Be Done

The honest assessment is that existing antitrust tools are inadequate for AI markets. Consumer harm measured through prices misses the problem. Market definition is ambiguous. Cases take too long. Vertical integration makes harm invisible to frameworks designed for horizontal competition.

What is needed is structural: updated competition law that recognizes concentration as harmful even absent price increases, that can act before markets have fully consolidated, and that addresses vertical integration in platform ecosystems specifically. The EU's DMA approach — ex ante rules for designated gatekeepers — is closer to what AI markets require than the US model of ex post enforcement through multi-year litigation.

For readers, this means the choices being made now about AI market structure will determine who benefits from AI for decades. Competition policy is redistribution policy. The question is not whether AI will create enormous value — it will. The question is whether that value will be captured by a handful of companies that happened to have the compute and the capital first, or whether competition will distribute it more widely.

That question is being decided now. And the answer depends less on technology than on whether democratic institutions can move as fast as markets consolidate.

Sources

Khan, Lina. "Amazon's Antitrust Paradox." Yale Law Journal 126, no. 3 (2017). Yale Law Journal
Vipra, Sai & Korinek, Anton. "Market Concentration Implications of Foundation Models." Brookings Institution, 2023. Brookings
Wu, Tim. The Curse of Bigness: Antitrust in the New Gilded Age. Columbia Global Reports, 2018. Columbia
European Commission. "Digital Markets Act." Regulation (EU) 2022/1925. EUR-Lex
U.S. Department of Justice. "Justice Department Sues Google for Monopolizing Digital Advertising Technologies." 2023. DOJ
Federal Trade Commission. "FTC Chair Khan and Officials Discuss Investments and Partnerships in AI." 2024. FTC
Bork, Robert. The Antitrust Paradox: A Policy at War with Itself. Basic Books, 1978. Publisher

General

AI and Antitrust: Why Competition Law Struggles with Foundation Models