AI as a Public Utility

This post is co-published with the Vanderbilt Policy Accelerator’s AI Governance Project.

** ** **

OpenAI’s Sam Altman recently said that he “see[s] a future where intelligence is a utility like electricity or water.” The idea that AI should be treated as a public utility has become increasingly common, with commentators in publications as diverse as Jacobin and Reuters suggesting the framework.

What would it mean to treat AI as a public utility? Along with some other scholars, I have spent the last decade working to understand public utility regulation in the United States and around the world, both historically and at present. This work has led to, among other things, the first new legal textbook in the field in a generation, Networks, Platforms, and Utilities: Law and Policy (2022), and a forthcoming Cambridge Handbook on the Law of Networks, Platforms, and Utilities (2026), in addition to many articles and books by myself and others.

In this post, I’m going to distill this body of research to offer a primer on how to think about public utility regulation and its tools – and how to apply them to AI. I’ll often refer to the field of law as NPU law and specific enterprises as NPUs, drawing on my and my colleagues’ work clarifying the field’s scope as “networks, platforms, and utilities.”

The Foundations of Public Utility Regulation

When people think of a “public utility,” they usually think of electricity, water, or sewer services. But the field of public utility law is much older and encompasses much more than these modern conveniences.

Core sectors that were historically governed by public utility principles include transportation (railroads, maritime shipping, buses and trucking, and airlines), energy (electricity, oil and gas pipelines), telecommunications (postal service, telegraph, telephone, broadcasting), and money and finance (banking). But there are and were many others, from horseshoeing blacksmiths and wharves to grain elevators and slaughterhouses to property and casualty insurance.

What did all these vastly different businesses have in common? The test for determining which firms were “in” and which were “out” was never mechanical; it was analogical, more akin to having a family resemblance than to satisfying a checklist. But there are common features:

• NPUs tend to be means to other ends – that is, they are usually inputs into a variety of other activities. As a result, they are often thought of as infrastructural enterprises and frequently provide essential services.
• NPUs tend to be services rather than goods. Because they are not tradable, these services confer a kind of social and economic power on the provider.
• NPUs usually have one or more of the following characteristics that limit the effectiveness of competition: natural monopoly or oligopoly, when the entire demand can be efficiently served by one or a few firms; network effects, where the value of the service increases with the number of users; functional monopolies, when a firm has an effective monopoly at the point of sale, for some set of users, or due to high switching costs; and horizontal interconnection, when firms in the same layer need to be connected to each other, such as postal services across jurisdictional boundaries.

NPU enterprises were historically governed by a special body of law not only because policymakers recognized that they possessed distinctive economic and social characteristics but also because policymakers wanted to ensure NPUs adhere to and advance a variety of public policy goals. These include:

• Industrial policy: building infrastructure at scale throughout a geography; ensuring continuity of service and service reliability; developing an integrated system; and preventing destructive competition.
• Fair commercial practices: preventing monopoly and oligopoly abuses, such as high prices and low outputs; poor service quality; and discrimination between similarly situated users, including favoritism and value extraction.
• Social goals: promoting commerce and industrial development; expanding or ensuring universal access to the service; and enabling (and sustaining) democracy while mitigating the rise of oligarchy.

People today often think of regulation as pitting businesses against consumers or as antithetical to development and innovation – but that is not the case with NPU law’s goals. Some of the goals are about preventing abuses of power that are business-to-business, rather than business-to-consumer, such as preventing monopoly abuses and value extraction when transporting goods. Some of the goals are about speeding development of the industry and scaling it, like building integrated and reliable electricity systems. Some of the goals are about ensuring dynamic innovation, including encouraging investment in businesses that use NPU services by making NPUs reliable at fair prices. And some of the goals are ecosystem goals: they seek to build social and economic structures that enable a range of societal values, such as universal communications services. These multiple goals and their mechanisms have often caused a lot of confusion: Some people describe the field as “antimonopoly” because it curbs abuses of power, while others claim the whole field of law is a function of “regulatory capture” because businesses can benefit from regulation.

The Tools of Public Utility Regulation

Because many different types of enterprises fit within these parameters and each sector has its own characteristics, “public utility regulation” was never a single blueprint applied indiscriminately across every industry with NPU characteristics. It also evolved over time as policymakers figured out how to address problems and configure the legal regime. Instead of seeing it as a unified package of rules, the better way to think about public utility law is as a set of tools that can also be configured and customized to the goals and problems of a specific sector.

The primary tools of NPU law fit into three broad categories.

Industry Structure Rules. The first set of NPU tools is about the structure of the industry.

• Structural separations. In many NPU sectors, a firm was not allowed to own or operate any other business in the sector in which it operated – no vertical integration, no conglomerates. The concern is that if a firm owns the utility or infrastructural layer and business lines in other layers, it could favor its own integrated business lines over competitors or leverage its power and monopoly in one sector into others. Separations are an administrable way to prevent these conflicts of interest and ensure that infrastructure is available to all users on fair terms, while not overly concentrating economic power in large integrated firms.
• Entry restriction. Entry restriction involves the government pre-authorizing a firm to enter into the sector before it is permitted to operate, usually through a licensing requirement. There are three primary situations in which entry restriction is deployed: (1) preventing destructive competition, in which many firms expend high amounts of capital and can end up going bankrupt in fierce competition with each other; (2) ensuring affordable service provision even in high-cost, low-volume contexts – think about the uniform pricing of the postage stamp, which enables sending letters to or from small towns and rural areas that cost more to serve; or (3) ensuring safety and security, such as licensing for airlines, nuclear power, and many other areas. Critically, only the first two justifications require limiting entry and competition. We can call these strong-form entry restrictions. The third justification could permit many firms and competition, so long as they meet minimum safety and security requirements. These are weak-form entry restrictions.
• Ownership and control rules. These shape who governs the utility. Both public and private enterprises can operate according to NPU principles: the Post Office is a public enterprise; railroads in the United States are private. Ownership and control rules also go beyond the public/private debate. For example, many NPU sectors have been subject to foreign ownership and investment restrictions to ensure domestic control of infrastructural resources.

Access and Service Rules. A variety of rules exist to ensure that a wide range of users have reliable access to the service, often to guarantee that they can make investments in reliance on that access. If an electric company can shut off your electricity on a whim, people might be deterred from investing in new factories or businesses that rely on electricity.

• Equal access mandates. These rules require NPUs to grant equal access to all similarly situated users – without discrimination.
• Universal service requirements. Some NPUs are required to serve every person within a specific geographic area – for example, telephone service or electric service within a jurisdiction.
• Exit restrictions. Sometimes NPUs are barred from withdrawing service to communities: think about the consequences for a factory if the sole railroad going to the plant shuts down.
• Interconnection mandates or interoperability rules. These rules ensure that users can use multiple utilities that span across jurisdictions, such as different railroads connecting to each other with the same rail standards, or can switch between firms, like porting a telephone number from one provider to another.
• Quality of service requirements. These rules ensure that the NPU offers some minimal, reliable level of service.

Price and Profit Rules. Finally, NPU law has often regulated prices and profits, in part because NPUs are often natural or functional monopolies or oligopolies and in part to ensure affordable and fair access to a broad set of users.

• Non-discrimination or neutrality rules. These rules prevent the NPU from charging different prices to similarly situated users, in effect picking winners and losers, or playing favorites based on arbitrary or corrupt criteria. Posted, transparent pricing requirements are a classic example.
• Rate setting. This could include capping prices at a maximum, setting minimum prices, or setting the exact price. The last of these is common in the electricity context: a regulator calculates prices taking into account capital and operating costs, adds a rate of return, and then figures out a uniform price to charge users.
• Profit sharing rules. Given that these businesses are often public services that have been outsourced to the private sector or given government-granted monopolies, in some cases governments have required profit sharing to ensure that the public gets some of the upside from the profits of the NPU firm.

The International Dimension. Many NPU sectors are also subject to international collaboration, law, and regulation. International bodies have been longstanding: the International Telegraph (now Telecommunication) Union, the Universal Postal Union, the International Civil Aviation Organization, and the International Atomic Energy Agency are but a few of these bodies. While their histories are different, international organizations have focused on:

• Balancing Sovereignty and Collaboration. These bodies of law often include preserving areas for national regulation and public policy, while simultaneously enabling international cooperation and collaboration.
• Coordination Rules and Technical Standard Setting. These bodies often established substantive coordination rules and technical standards so NPUs could operate across national boundaries.
• Safety and Security. These bodies of law also frequently involve addressing safety and security, including through setting standards and establishing reporting or investigatory mechanisms for divergences from those standards.

The tools of NPU law can be configured differently for different sectors. Historically, there has not been a one-size-fits-all package for every sector. In the middle of the twentieth century, for example, AT&T was a national monopoly, regulated as a public utility. The airlines were regulated as public utilities, but as part of a system of regulated competition that enabled roughly a dozen major airlines to operate. Banking law dispersed operations to thousands of small banks around the country.

The Two Big Questions

In thinking about how to regulate AI as a public utility, there are two big questions:

(1) What business line or lines are we talking about when we say “artificial intelligence”? To heavily simplify the AI tech stack, we could be talking about (a) semiconductor chips that are used within (b) cloud computing data centers which train and process inference for (c) foundation models, which are used to build (d) applications.

(2) Given a particular business line, what tools from the public utility toolkit should apply?

Let’s start with cloud computing.

Cloud. Cloud computing has high capital costs, it isn’t easily replicable, it’s hard to exit due to a variety of lock-in effects, and it’s a service that is infrastructural – it is a means to many other ends. In a sense, compute is the electricity of the technological age. As a result, it’s no surprise that cloud isn’t very competitive – three big providers, Amazon Web Services, Microsoft Azure, and Google Cloud, dominate the field. Given this infrastructural role, one also might be particularly worried about these cloud providers leveraging their power up and down the stack.

My colleague Asad Ramzanali has written extensively about how one might regulate the cloud using public utility principles, as have others. One can easily imagine applying structural separations, non-discriminatory access and pricing rules, and potentially interoperability requirements and uniform standards at this layer. What would this look like? It would mean that cloud businesses couldn’t own other businesses – models, applications, anything else. In short, AWS, Azure, and Google Cloud would be spun off as their own businesses. They’d be required to serve similarly situated users with non-discriminatory prices. And there could be mandatory uniform standards so that users could easily port their work across cloud providers – think of that as akin to standardized railway gauges.

Strong-form entry restriction probably doesn’t make sense. There isn’t a need for universal service in low-volume, high-cost times or places, nor is this a case of destructive competition. The best case for entry restriction would be a weak-form entry-restriction rule, one that does not limit the number of providers but does require them to meet minimum safety and security standards, perhaps for cybersecurity, or perhaps some know-your-customer rules if there are security concerns about user identity. Foreign ownership restrictions might also make sense, given the importance of both data and compute power and the dangers of them being in the hands of adversaries.

Models. Building frontier models requires an enormous amount of data and very costly amounts of compute. This makes them relatively difficult to develop, and the sector is unlikely to be terribly competitive (I think of 4-6 players as not that competitive; compare your local restaurant market – it is much more so in most big cities!). At the same time, there are people who think that open-weight models will be able to provide competition, so regulation (and particularly utility style-regulation) is not necessary. And there are still others who worry that open-weight models might be more dangerous to society. 

This is one of the more challenging areas to analyze for a few reasons. First, there is the issue of what constitutes “open.” Openness in AI isn’t a binary, but rather a “gradient” and the degree of openness and terms and conditions involved might matter a great deal. Indeed, in some cases, “open” systems have still led to concentration. Second, at the moment, there appears to be a difference between frontier models and non-frontier models; open-weight models may be able to operate effectively but not at the frontier. Third, and relatedly, there is the question of what happens if open-weight models become good enough and cheap enough that they are commoditized and highly effective for most consumer and commercial applications (some believe this is already the case). In that scenario, there might be significant competition at the model layer, even if the frontiers of model development are dominated by a limited number of players. Fourth, the question of open/closed may not be the most relevant one depending on one’s policy goals. To the extent one is interested in interoperability, equal access, international standards and cooperation, and safety and security, one could want rules to apply to both open- and closed-weight models.

With those caveats in mind, there are still some reasons to think that some NPU tools could be helpful at the model layer.

First, structural separations would prevent model providers from taking over the applications market or integrating with and leveraging their power in the cloud, as is already the case for Google and Microsoft. Here, the idea would be that model providers could not be vertically integrated, owned by, or own others in the sector – cloud, applications, etc. Neutrality rules could help ensure non-discriminatory access, but they will be more difficult to administer when implemented without a clear separation rule. Non-discriminatory pricing rules and equal-access mandates would likely make sense in any scenario, since their downsides are few if any. This latter point might seem like no big deal, but the Windsurf case illustrates the value of it. There, Anthropic cut off Windsurf’s access to certain Anthropic model APIs for the stated reason that OpenAI had reportedly agreed to buy Windsurf (Anthropic had also just launched its own competitor to Windsurf, Claude Code).

To the extent one is worried about the safety and security concerns with frontier models themselves (as opposed to applications that run on the models) or thinks that model-providers are the best cost-avoider on safety and security, then entry restrictions might also be justifiable. Here it is worth distinguishing, again, between strong-form and weak-form entry restrictions. Entry restrictions for AI models need not be designed to limit entry and competition if the goal is safety and security rather than universal service or preventing destructive competition. Firms that meet safety and security goals could be permitted entry, even if there are many of them. This does not mean that some amount of safety and security rules would allow every entrant; it might prevent some who will not or cannot meet the safety and security requirements and those requirements could be stronger or weaker, which would have implications for entry. But the purpose of such weak-form entry restriction rules is not to prevent new-entry and competition.

It is also worth distinguishing between entry restriction and pre-approval for deploying models. Entry restriction is about the enterprise’s ability to operate at all, that is, to have a charter and the legal right to operate. That can be conditioned on meeting various safety or security standards. Pre-approval for deploying models is distinct from entry restriction. That’s more like the FDA needing to approve every particular drug a pharmaceutical company makes before it can sell it to the general public. This operates independently of whether those pharma companies are legally allowed to exist.

One might also argue that there are safety and security reasons for strong-form entry restriction. That is, public policy should limit entry and competition on purpose to further safety goals. This argument is likely to be one from administrability: limited competition is needed because it would be too hard to approve many entrants and then monitor their compliance with safety and security rules. It’ll be easier, practically speaking, to do it just for a few.

The best economic argument for strong-form entry restrictions is that the model layer is experiencing destructive competition, meaning that model companies are in a price war, cutting prices to below cost, in an attempt to survive in a competitive winner-take-all race. This leads to a lot of wasted capital and a higher risk of a boom-and-bust cycle or an AI bubble bursting, with the possible consequence being a crash, recession, or depression – and all the externalities and society downsides that causes. One of the aims of strong-form entry restrictions is to prevent these consequences and create investment stability. Importantly, because they are designed to restrict competition, strong-form entry restrictions need to be coupled with rate regulation.

Chips. This is a difficult case for NPU law. At first blush, chips are a bad fit for NPU regulation. Chips are goods, not services, and despite there being limited competition at this layer of the AI stack, they do not have a lot of the features of NPUs, historically speaking (network effects, desire for geographical service, etc.). The best case for applying NPU tools might be one based on functional monopoly. Leading-edge chip manufacturing (TSMC’s business) is a functional monopoly because the capital, ASML’s extreme ultraviolet (EUV) lithography machines, is so rare and expensive (ASML has a monopoly) and the manufacturing skill required is highly specialized and limited.

NVIDIA’s business is a hard case too because it sells chips (goods), even though its value-add is in chip design. Recall that generally speaking, goods are not NPUs and that NPU toolkit is usually not applied to them. In any case, chip design is highly specialized with very few people and organizations that can operate at the frontiers. This too might make it a functional monopoly, even if not a natural one. The functional monopoly might justify structural separations and neutrality rules (preventing cross ownership/financing and picking winners and losers), rate regulation to prevent monopoly rents, and interoperability rules and standard-setting to commoditize the sector. But it alone would not likely justify entry restriction. As with many layers of the AI tech stack, this assessment only scratches the surface.

Applications. As a category, the application layer seems least suited to public utility regulation. But the key is “as a category.” There is a high variance in types of applications and the functions that they serve. There will be some AI applications that are NPU-like, performing infrastructural services and meeting many of the basic NPU criteria. But there will also be many AI applications that do not meet the criteria and that can exist in highly competitive sectors. This makes assessing the category of “applications” difficult, if not impossible. A better approach will be to assess categories of applications or specific applications to determine whether NPU tools should apply.

Rule of Law and the American Public Utility Tradition

Pulling it all together, the American tradition of public utility regulation has a lot to offer debates over artificial intelligence. Its core approach is to think through particular business lines and assess their specific dynamics. Historically, the tools of NPU law have been configured to address the nature and problems of each layer, with policymakers sometimes adopting regulated monopoly and other times regulated competition as the overarching market structure. This system of regulation has often created stability, even as it has enabled the development of industry, and it balances public and private interests.

Perhaps one of the most important benefits of public utility regulation is that it offers a rule-of-law-based way to address a wide range of economic, social, and national security problems and issues. At a time when people on both left and right, in industry and outside of it, are concerned about expansive presidential power and its use in ways that are not necessarily transparent, systematic, or even-handed, we need a way to govern AI that is grounded in the rule of law. There are other strategies to regulate this industry, of course: no regulation, self-regulation, nationalization. But public utility law offers a coherent and effective way to address a wide range of problems – and in the process, to ensure we have a country based on laws, not the whims of a few powerful people.