The United States power grid is increasingly strained by the surging electricity demand driven by the AI boom. Efforts to modernize the power infrastructure are unlikely to keep pace with the rising demand in the coming years. Barak and Eli Orbach explore why competition in AI markets may create an electricity demand shock, examine the associated social costs, and offer several policy recommendations.
The artificial intelligence boom began in late 2022 with OpenAI’s release of ChatGPT. Since then, competition among developers of large AI models has been the primary driver of this frenzy. Companies are racing to establish themselves in emerging AI markets. They do so by developing increasingly advanced models that surpass rivals in performance, scalability, and efficiency.
However, entering and succeeding in the AI race is fraught with challenges. Companies must invest heavily in infrastructure and secure a reliable supply of critical inputs, including data, specialized chips, talent, and electricity. The needed capital expenditures create significant barriers to entry, and investors are increasingly questioning the feasibility of expected returns.
Among the challenges, the power constraint—namely, the limits on available electrical supply—is considered the most pressing. From 2000 to 2023, U.S. annual electricity consumption fluctuated between 3.734 and 4.157 terawatt-hours (TWh) (see Figure 1). Boston Consulting Group and Goldman Sachs project that it will exceed 5.0 TWh by 2030 (see Figure 2). As electricity demand rises, supply growth and efficiency gains are expected to fall behind.
Over the past few months, business and financial news organizations have published numerous reports detailing how soaring electricity demand is straining the power grid, delaying the retirement of fossil-fuel power plants, and limiting further advancements in AI technologies. The Wall Street Journal has explained this dynamic:
As long as the competition between makers of AI continues to spur companies to use these ever more capable, ever more power-hungry models, there’s no end in sight to how much more electricity the global AI industry will demand. The only question, then, is at what rate its consumption of power will increase. (emphasis added).
Tech executives have also emphasized the significance of the power constraint. For example, at the 2024 World Economic Forum, OpenAI CEO Sam Altman stated that the future of AI would require significantly more power than anticipated, making an “energy breakthrough” essential. Similarly, Amazon CEO Andy Jassy remarked that “there’s just not enough energy right now” to meet the demand from companies developing large AI models.
Policy debates surrounding the AI boom address important concerns and risks but often overlook its impact on the power grid. Lawmakers, regulatory agencies, and scholars must improve their understanding of the distinctive characteristics of AI markets driving the forming demand shock. A stringent industrial policy must be developed and urgently implemented.
The AI Power Grab
The beating heart of the AI boom is the foundation model (FM). FMs, such as OpenAI’s GPT and Google’s Gemini, are complex AI systems pre-trained on vast datasets, enabling them to perform diverse tasks across many domains. Their properties are driving the AI race. FMs are general-purpose technologies—transformative innovations that disrupt industries, enable a wide array of new technologies, and lead to significant productivity gains. Examples of other general-purpose technologies include the steam engine, electricity, and the internet. FM markets are expected to be oligopolistic, at least in the foreseeable future. It is essential to understand why.
Firstly, like digital ecosystems, FMs benefit from extreme economies of scale and scope, as well as network effects. Secondly, the development and training of FMs require gargantuan capital investments, meaning that firms without the deep pockets and infrastructure of big tech will be limited in their capacity to innovate and compete. FM startups typically rely on partnerships with these tech giants, secured through equity sales. Thirdly, a shortage of parts and inputs for AI data centers relative to new, heightened demand constrains the number of firms capable of developing FMs, a bottleneck that will continue to impede markets in the coming years. Fourthly, first-mover advantages in FM markets create significant barriers to entry. Early entrants leverage their infrastructure and experience to develop more advanced FMs, making it difficult for later entrants to catch up, as illustrated by Apple’s challenges to catch up with successful early entrants. Finally, firms in upstream and downstream markets have strong incentives to partner with established companies rather than newcomers.
These conditions have set the stage for the AI race, where firms compete through capital investments and vertical arrangements to secure their positions in emerging AI markets.
The Social Costs of the AI Race
The commercialization of transformative innovations often leads to costly creative destruction. While the losses incurred by those affected are significant, our focus here is on other categories of social costs arising from the increasing demand for electricity. Given the uncertainty surrounding the magnitude and characteristics of the electricity demand shock, current estimates of its effects remain speculative. Nevertheless, some anticipated impacts can be characterized.
1. Electricity Prices. Electricity is a general-purpose input essential to all sectors and households. Therefore, the most immediate effect of the demand surge will be increases in electricity costs, which will be felt mainly by those connected to power grids serving clusters of AI data centers. Over time, this could create inflationary pressures. Further, in the U.S., changes in energy prices also tend to have political implications.
2. Shortages. The rising electricity demand has already impacted the ability to build new AI data centers or expand existing ones. For example, most AI data centers are concentrated in Data Center Alley in Loudoun County, Virginia, where the current power infrastructure can no longer support additional construction. These shortages will continue to worsen until expansions of power generation and transmission capacities are attained.
3. Misallocation of Resources. The emergence of lucrative markets often leads to excessive investment in new entries, creating technological bubbles. The FM market is no exception. Entrants apparently believe they can succeed; otherwise, they would not invest in the first place. However, some will inevitably be proven wrong. While their losses are their own, their participation in the AI race drives up the prices of inputs and complementary goods for everyone, creating externalities.
4. Economic Resilience. Developers of FMs have thus far prioritized rapid growth at the expense of safety and resilience. Specifically, decisions about building AI data centers sometimes overlook their impact on the grid’s resilience. For instance, over 70% of the world’s internet traffic passes through Data Center Alley. This level of geographic concentration poses a systemic risk to the economy.
5. Sustainability. The booming demand for electricity has forestalled plans to retire fossil-fuel power plants and forced tech giants to withdraw from their publicized commitments to slash their greenhouse gas emissions. Other companies are expected to face similar constraints. In other words, the energy demand shock has already undermined climate change initiatives.
Policy Blind Spots
While the term “demand shock” is rarely used to describe the current situation, it is widely recognized that the U.S. grid cannot support the soaring demand for electricity. Despite this, the federal government seems focused on AI’s potential to enhance power generation and transmission, often downplaying the associated challenges. For example, President Biden’s AI Executive Order guides the federal government on how to “harness[ ] AI for good and realiz[e] its myriad benefits requires mitigat[e] its substantial risks,” and the 2024 Presidential Economic Report presents an “economic framework for understanding artificial intelligence.” Both documents overlook power constraints and, therefore, illustrate policy blind spots.
Policy statements from the Department of Justice and Federal Trade Commission also fail to address the power constraints or acknowledge the unique characteristics of the AI race. For example, in July 2024, the DOJ, FTC, European Commission, and United Kingdom’s Competition & Markets Authority issued a Joint Statement On Competition in Generative AI Foundation Models and AI Products. This statement emphasizes that established tech giants may “entrench or extend” their power by leveraging their advantages in AI markets or suppressing market development. Furthermore, the agencies’ joint statement posits that “firms may attempt to restrict key inputs for the development of AI technologies,” such as “[s]pecialized chips, substantial compute, data at scale, and specialist technical expertise.” The agencies’ concerns regarding market concentration are sound but incorrectly assumes that there are alternatives to oligopoly. As explained above, there are not. They also do not refer to electricity, which is the principal constraint.
Recommended Policy Framework
We present two principal policy aspects that, if considered, could significantly enhance national preparedness for the AI electricity demand shock.
1. Time-Horizon Considerations. The AI race is a form of competition that is expected to generate high social costs in the short term, while some of its important benefits—including the modernization of the power system and economic growth—are expected to be attained in the long run. This means that public policies must balance time-horizon considerations. In the short run, it may be necessary to implement industrial policies—intrusive measures to mitigate market failures to promote growth, prosperity, and other national goals. For example, consider the following constraints:
A. Energy Consumption Constraints. Government-imposed constraints on inputs critical to innovation and growth should be considered only in extreme circumstances and applied cautiously. Concerns about potential grid failures might justify such constraints, possibly in the form of consumption taxes or quotas for AI companies.
B. Geographic Concentration Restrictions. Data center hubs like Data Center Alley offer considerable advantages by attracting infrastructure companies, specialized talent, and other service providers. However, they also pose systemic risks; extreme weather events, natural disasters, or significant grid failures affecting these hubs could disrupt the entire economy. Therefore, the federal government must develop national policies addressing risks associated with geographic concentrations of data centers, such as Data Center Alley.
2. Antitrust Limits. The DOJ and FTC’s statement on competition in AI markets raises concerns that the agencies’ approach to the AI race is neither practical nor well-informed. Antitrust law does not authorize courts or agencies “to decide whether a policy favoring competition is in the public interest. . . . Subject to exceptions defined by statute, that policy decision has been made by the Congress.” Instead, antitrust analysis is meant to “form a judgment about the competitive significance” of market transactions and practices. However, this established legal standard does not necessarily support the conclusion that competition among startups is superior to competition among tech giants.
Conclusion
Technological arms races, such as the AI race, are a recurring phenomenon. Examples include competition among railroads in the 19th century, the dot-com bubble, and cloud computing. One of the intriguing characteristics of the AI electricity demand shock is its temporal nature. While competition to secure a reliable power supply is expected to expand the electricity supply in the long run, it will exacerbate the power constraint problem in the short term. With a grid already straining to meet existing demand, developing informed national strategies to support the grid and enable an effective transition to the AI age is imperative.
Authors’ Note: For a bibliography of articles, reports, and papers on AI, competition, and the electricity demand shock, see here.
Authors’ Disclosures: the authors report no conflicts of interest. You can read our disclosure policy here.
Articles represent the opinions of their writers, not necessarily those of the University of Chicago, the Booth School of Business, or its faculty.
