As with paradigm-shifting technologies before it, artificial intelligence raises myriad policy implications for decision makers and analysts alike.
The federal policy debate around AI often focuses on it as an application: think “AI for drones,” or “AI for manufacturing.” The political debate is focused on AI’s insatiable demand for energy: think “nukes for AI” or NIMBYism and data center buildouts. The debate in Congress surrounds the role that states play in AI regulation and whether they can or should be preempted by a federal national standard.
But for antitrust purposes, the AI debate is less about how the technology is applied, or powered, or even regulated. The antitrust debate is rightly about competition within the AI technology stack itself. When it comes to antitrust, AI is for the most part just another technology in need of quiet, nerdy analysis on a case-by-case basis. This fact may prove disappointing to those in search of a unitary, administration-wide “AI Competition Policy,” but that is not how antitrust enforcement works.
Ultimately, the goal of antitrust law in AI should not be to pick winners and losers or to slow technological progress. Rather, the goal is to preserve the conditions under which that progress remains possible. Antitrust law can ensure that the marketplace remains sharp, competitive, and innovative rather than becoming a place controlled by entrenched gatekeepers.
U.S. antitrust statutes are common-law framework statutes. It is the job of the courts to color between their lines by creating judicial precedent. As much as an AI Competition Policy may appeal to some, it runs counter to the common-law enforcement model. While policy can be discerned from the case law, to state a policy up front puts the cart before the horse.
Taking a case-by-case approach starts with analyzing the relevant markets involved. The AI stack is built on layers of infrastructure, data, engineering talent, software, and distribution channels. Antitrust markets within this stack include data itself, foundation models, developer tools, cloud computing, and semiconductors. Each layer can be considered a separate market in need of bespoke analysis and precedent—at least for now. Antitrust law could one day apply to one dominant AI platform controlling the full stack absent competition, a “One Ring to Rule Them All” AI scenario, so to speak. But we are not there yet.
Even though the AI market is broadly competitive today, there are antitrust hawks in the Neo-Brandeisian school who would apply the precautionary principle to AI in order to avoid the mistakes of the past. These hawks, such as Yale professor Fiona Scott Morton, argue that an ounce of prevention is worth a pound of cure. They favor early intervention and argue that AI markets exhibit many of the same structural features that prompted historic antitrust scrutiny in railroads, operating systems, and Big Tech platforms. These features include high fixed costs, network effects, economies of scale, vertical integration, and at times high switching costs for consumers.
Antitrust hawks note that Web 1.0 and Web 2.0 were left unscathed by antitrust scrutiny for too long, resulting in several markets tilting toward monopoly. This state of affairs remained largely unchallenged until 2020, when the first Trump administration’s Department of Justice sued Google and Facebook, alleging the companies had monopolized online search and social media, respectively. Six years on, these cases are still working their way through the courts and are on appeal in the District of Columbia Circuit. According to the antitrust hawks, if Web 1.0 and Web 2.0 are the standard, a hands-off approach to antitrust in AI markets could be too little too late.
In contrast, antitrust doves in the George Mason tradition strenuously oppose early antitrust intervention in AI markets. The doves argue that enforcers should avoid punishing innovation and should instead let the AI ball rip for now. AI accelerationists in the Trump administration further argue that anything that slows down AI adoption, whether it be regulation or antitrust enforcement, must be eschewed on national security grounds. In a world where the U.S. is in a neck-and-neck race for global AI dominance, the U.S. government should stand aside.
The logical conclusion to this line of thinking seems to be that even if AI markets tilt toward monopoly, the government still should not intervene, as if a state-sanctioned private monopoly is needed to compete with those of China. Yet this presents a false dichotomy. The U.S. free market approach is a proven winner that has withstood the test of time. Impartial antitrust enforcement is a key and essential component of that proven model. We do not need to become like China in order to beat China.
The challenge, therefore, is to strive for balanced and targeted, meticulous antitrust enforcement. Competition authorities can and should distinguish between conduct that reflects legitimate innovation and conduct that unlawfully entrenches dominance and market power. High market shares alone are not illegal under U.S. law. Big is not Bad, but neither is Big Beautiful. What matters is exclusionary conduct that harms the competitive process itself.
What might this conduct look like?
The AI stack could theoretically trigger nearly every classic antitrust issue into a single ecosystem, including control of infrastructure, network effects, vertical integration, platform gatekeeping, and rapid technological change. Fortunately, our common-law antitrust statutes are flexible enough to address many of these issues using existing precedent, though enforcement will require careful economic analysis and a nuanced understanding of how AI markets work. Once again, antitrust intervention should be grounded in facts and evidence, not magical thinking.
How should antitrust enforcers act today to analyze the AI stack? The AI stack begins with compute. Training frontier AI models requires enormous amounts of computational power, particularly advanced graphics processing units and specialized accelerators. A small number of firms already dominate these markets, which creates several potential antitrust issues.
First, a dominant chip manufacturer could engage in exclusionary conduct by tying hardware to proprietary software ecosystems or entering exclusive arrangements that foreclose rival model developers. This seems to be the issue under scrutiny at the Federal Trade Commission as it reportedly investigates Arm Holdings’s attempt to monopolize parts of the semiconductor market by leveraging control over essential central processing unit blueprints while competing downstream in data center chips.
The history of antitrust enforcement in technology markets provides instructive precedents for the AI stack. In the Web 1.0 operating-system era, antitrust enforcers worried that a platform owner could leverage its dominance in one layer into adjacent markets. Similar concerns may arise today if a dominant cloud provider privileges its own AI products through pricing advantages, preferential access to compute, bundling, or technical integration.
This was the lesson from the DOJ’s first foray into high-tech antitrust when it sued Microsoft in the late 1990s. In that case, Microsoft was found to have leveraged its operating system monopoly into the then-nascent online internet browser market. The DOJ won the case having successfully demonstrated the ways in which Microsoft used its clout in operating systems to preference its own browser and foreclose competition. A similar scenario could come to pass with cloud, should the evidence support it.
Data was not a critical input in the browser wars, but it is surely a critical input in the AI stack. Large language models depend on massive datasets for training and refinement, and data can be a bottleneck standing in the way of scale. Today, a heated debate focuses on copyright and fair use when it comes to training data, but there are also competition dimensions to the issue.
The hard-fought battle over copyright’s fair use doctrine and whether it applies to data used to train AI models foreshadows the antitrust question. Antitrust law traditionally does not require firms to share assets with competitors. There is not a widespread or inbuilt duty to deal under U.S. antitrust law. However, concerns can arise when dominant firms use exclusivity agreements, acquisitions, or technical restrictions to prevent rivals from obtaining necessary data inputs to train models.
This is not the first time this issue has arisen, and recent case law sheds light on it. Proprietary search data and its use to foreclose competition was a core issue in the DOJ Google search antitrust case, and the same principle applies to AI. In the DOJ search case, the judge found that Google had unlawfully monopolized the market. Instead of breaking the company up, the judge recognized the vital role data plays in enabling competition, and crafted remedies aimed at freeing up search index and user interaction data for use by certain competitors. The same logic holds for the role data plays in competitive AI markets.
Another antitrust issue is whether incumbent AI platforms are using mergers and investment playbooks to neutralize competition from smaller rivals. Mergers, exclusive licensing arrangements, preferential cloud commitments, and governance rights may all receive attention from the antitrust agencies. Even when an investment falls short of a formal acquisition, antitrust agencies may investigate whether it effectively gives an incumbent influence over a potential competitor.
Section 7 of the Clayton Antitrust Act is especially important in this context because it prohibits mergers and acquisitions whose effect “may be substantially to lessen competition.” Historically, antitrust agencies have focused on horizontal mergers between direct competitors. In AI, however, concerns may also involve nascent competitors and even vertical investments involving firms upstream or downstream from one another.
A dominant AI platform could, for example, acquire an innovative AI startup before it matures into a true competitive threat. Agencies increasingly worry about so-called “killer acquisitions” or serial acquisitions that prevent disruptive competition from emerging. Antitrust hawks point to mergers in the Web 2.0 era to argue that, for example, Facebook’s acquisition of nascent rivals WhatsApp and Instagram altered the face of competition in social media. The hawks argue that antitrust enforcers were not sufficiently vigilant at that time and cleared these acquisitions without meaningful scrutiny, allowing the social media market to tilt toward monopoly. Antitrust enforcement caught up with Facebook in 2020, but the hawks believe this was too little too late.
Antitrust law can also reach labor markets within the AI stack. This is separate from the important and broader debate around the impact of AI and automation on American workers in general. There, the tentpoles rest somewhere between the recent papal encyclical on AI and Universal Basic Income.
Echoing Rerum Novarum, Pope Leo XIV warns that AI should not undermine workers’ dignity or concentrate wealth and power in the hands of a few firms. Technological progress should benefit society broadly, according to Leo. Antitrust concerns as they relate to labor pivot primarily on competition for elite AI researchers, which is intense. Antitrust agencies have increasingly scrutinized labor market restraints such as no-poach agreements and wage suppression in labor markets generally. Acquihires and exclusive employment arrangements involving key technical talent may receive attention if they substantially reduce competition for specialized labor. The FTC is thought to be taking a close look at this issue in particular.
Finally, there is the important role that the government plays in keeping markets competitive when it is the only customer in those markets. Every year, the Pentagon spends billions of taxpayer dollars on AI for national security purposes. Ensuring these markets stay competitive benefits not only the public purse in the form of lower prices to the taxpayer, it also benefits our national security to have more than one firm innovating in AI.
Innovation competition will be key to achieving AI national security goals, and nothing drives innovation better and faster than competition. Here, the Pentagon and other agencies engaged in AI procurement can play an important role educating antitrust enforcers on the benefits and drawbacks of individual mergers from a competition standpoint. Such customer testimony is key to the ability of antitrust enforcers to challenge a merger in court. Having the Pentagon on the side of free markets and competition is key to any antitrust court challenge to a national security impacted merger.
This approach is to be preferred over the “do nothing” scenario Sen. Josh Hawley (R-MO) described in his speech at the American Compass gala in June. “Let me be blunt,” he argued. “Left to itself, artificial intelligence will not choose moral liberty. It will choose the passions. It will choose appetite. It will gather power into the fewest hands the world has ever seen and call the result progress. That is the rule of the strong over the weak. That is the dissolution of our moral covenant. It is the oldest temptation in the history of humanity, now just dressed up in silicon.”
These words echo the sentiments of another Republican Senator, Ohio’s John Sherman, who told Congress in support of his eponymous 1890 antitrust legislation, “if we will not endure a king as a political power, we should not endure a king over the production, transportation, and sale of any of the necessaries of life.”
History does not repeat itself, but it often rhymes.
Tough choices. I don't know where we end up but any policy that is centered in the judiciary is guaranteed to be dysfunctional. The MA case was a 20 year fizzle.