Cheap Goods for Everyone? The Impact of Market Power in Artificial Intelligence on Welfare and Inequality
A common fear is that transformative AI might concentrate power among a few companies and individuals. Despite its promise, TAI might then leave ordinary people worse off. In this essay, Susan Athey and Fiona Scott Morton explore this problem, noting that while TAI could deliver productivity gains to consumers through lower prices or better products, such outcomes aren’t guaranteed. They propose policy measures, focused on protecting market competition, to ensure TAI’s benefits are broadly shared.
Introduction
AI has generated tremendous excitement in the technology community and the press. Economists predict dramatic growth, investors pour in money, and governments position themselves for the future. Yet much of the policy debate overlooks a key issue: whether AI’s resource savings reach end consumers. Competition drives innovation, quality, and lower prices. Without it, firms retain cost savings as profit, charge high prices, and stifle downstream innovation, reducing application quality and variety. Market power also allows firms to “pick winners” or integrate into other value chain segments, limiting customer choice and diminishing competition in adjacent industries.
Workers who are displaced or face lower wages as a result of AI-fueled automation will be worse off unless they simultaneously experience lower-priced goods and services, that is, unless the cost savings from their displacement are passed on to them. And for a country as a whole, when AI is an imported factor of production, market power can mean the difference between increases or decreases in national income, while a dependence on a single foreign supplier creates ongoing risk of disruption. Since AI is a general-purpose technology used not just by end consumers but also (and especially) by businesses, market power in AI can affect industry structure, prices, output, and the organization of production in downstream industries, effects that in turn feed back into factor markets, particularly labor markets. Thus, understanding the sources and consequences of market power becomes critical for macroeconomic and trade policy regarding AI.
This paper contemplates the consequences of market power in AI in the scenario where AI is transformative, as distinct from conventional technological progress. Transformative AI does not alter business choices at the margin but instead drives a larger reorganization of economic activity and distribution of surplus across groups and countries. Our model considers reallocation of resources across sectors, impacts on price levels and sectoral income, and feedback effects of changes in income on demand for goods and services, which in turn determine demand for workers in each sector. To analyze the transformative AI scenario, we take a macro view and solve for a general equilibrium. This macro framing is unusual for an analysis of competition, where a more typical focus is on one market. The paper goes one step further by considering the impact of being open to trade in a world where many nations are adopting AI. The widespread availability of AI implies that a country is unlikely to sustain growth through exports of goods in the sector where AI displaces workers.
Why Should We Be Concerned About Market Power?
Why is market power a concern when many firms invest heavily in AI today? AI products are created in a multilayer value chain that includes components such as chips, computing power, proprietary data, language models, applications, distribution, devices, and systems integration. Each layer presents potential barriers to entry—including scale economies, sunk costs, capabilities, intellectual property, data access, switching costs, network effects, and control over key distribution channels or sources of supply. Even if most layers were competitive today, history shows bottlenecks in any layer can foster market power. Concentration in one layer creates the opportunity for firms to raise prices and capture more profit from the entire value chain. Furthermore, specific customer groups may have limited AI supplier options due to trade restrictions or specialized needs.
“Even if most layers were competitive today, history shows bottlenecks in any layer can foster market power. Concentration in one layer creates the opportunity for firms to raise prices and capture more profit from the entire value chain.”
Some layers of the AI value chain are global with multiple competitors, but others are shaped by national policies or constraints. For example, governments may restrict foreign chips for security reasons, reducing domestic competition, or they may subsidize local entrants to boost global competition. Export controls can limit access to critical inputs, sometimes leaving countries with only one supplier.1 Similar constraints apply to foundation models, which may not be offered in some jurisdictions due to regulatory costs. Local language needs may also restrict application availability. Systems integrators or defense contractors may hold market power, charging high markups supported by regulatory barriers.
Even in global markets, the history of digital markets has many examples of firms expanding market power by integrating into adjacent value chain layers and excluding competitors—behavior that may create a scenario in which new entrants need to introduce products at multiple levels to succeed. For example, mobile phone providers might restrict AI application providers’ access to app stores or key functionality, forcing them to enter the market through their own operating system or device.2 Similar challenges have played out with browsers3 and exclusive chip supply agreements.4 Anticipating such dynamics, firms may enter into deals or vertically integrate to guard against or exercise market power.5 The past is instructive: Tech giants like Apple and Google have long acquired adjacent firms and integrated across hardware, software, and deep supply chains, often using exclusive agreements.
“The past is instructive: Tech giants like Apple and Google have long acquired adjacent firms and integrated across hardware, software, and deep supply chains, often using exclusive agreements.”
Market power in downstream industries also threatens economic welfare. When dominant firms adopt labor-displacing technology, they often retain the cost savings rather than passing them on to consumers. This can mean falling wages and difficult transitions for workers, while consumer prices remain high. Even if profits stay in the country, they may not circulate widely through the economy.
In this article, we argue that even a small risk of market power warrants making competition a core consideration in AI regulation and policy. For example, AI safety rules may protect consumers, but if they reduce competition, they can cause new consumer harms and weaken long-run market incentives for safety. Ensuring regulation does not undermine competition is important across industries, as promoted by both the current and previous US presidential administrations, but it is especially critical for a general-purpose, transformative technology such as AI.6
Modeling the Macroeconomic Impact of Labor-Displacing AI
“Since AI affects industries globally, productivity gains may not lead to increased exports or sector growth.”
The macroeconomics literature has argued that productivity shocks usually help economies by enabling more output with fewer inputs, often fueling growth through sectoral expansion for domestic consumption or for exports. However, these shocks can also increase inequality or raise prices in other sectors by pulling in factors of production to fuel the growth. We argue AI will not primarily act this way. Since AI affects industries globally, productivity gains may not lead to increased exports or sector growth. Countries relying on cheap labor for exports may lose that edge, making AI-driven export expansion unlikely. Further, when AI displaces labor, it pushes workers into sectors with lower marginal returns rather than pulling them from elsewhere. On the capital side, AI may have more traditional effects, e.g., raising demand for computing and thus inducing shortages and higher costs.
More broadly, much of the macro literature on AI focuses on its impact on the capital-labor ratio and the inequality issues that might arise. Nordhaus presents a model in which AI increases capital investment and growth, emphasizing that returns accrue primarily to capital: “Capital eventually gets virtually all the cake, but the crumbs left for labor—which are really small pieces of the increasingly huge mountains of cake—are still growing at a phenomenal rate.”7 However, we argue in our companion paper, “Artificial Intelligence, Competition, and Welfare,” that this logic breaks down when consumer prices do not fall alongside costs due to market power.8
AI as an imported factor of production with market power
Traditional macro models often assume fixed industry structure and markups, even in the face of technical change, and do not contemplate monopoly or oligopoly suppliers of critical inputs. These features are vital for understanding the impact of AI. We argue that it is important to further consider the impact of AI on international trade, as some countries may lose their export market advantages due to globally available AI. In addition to any AI-driven worker displacement, market power in the AI supply chain may result in national income leaking abroad due to high technology prices.
“In addition to any AI-driven worker displacement, market power in the AI supply chain may result in national income leaking abroad due to high technology prices.”
And, although a vast economics literature has grappled with market power in digital markets, almost all of that literature takes a partial equilibrium perspective. We argue that general-purpose technologies have general equilibrium effects that are first-order, and we show that these can feed back into the incentives of the firms with market power.
To capture the case of AI, particularly for a small country that may import AI technology, we build a model of an open economy that imports AI as a factor of production, one that displaces other factors, in particular labor.9 The economy is otherwise closed, so that each sector faces diminishing returns to production (constrained by the marginal utility of home consumers). This ensures that AI productivity shocks do not allow a country to lean into exporting the AI-powered goods, consistent with the hypothesis that goods that use AI as a primary factor of production can be produced in many places, and countries that previously relied on cheap labor to drive exports may need to reallocate into different sectors. Of course, some countries may have specific resources, data, or other specialized assets that do allow them to increase exports as a result of AI, but we focus on the alternative.
We build on insights from the trade literature, and particularly the literature on offshoring.10 However, we also highlight differences from the standard setup. The trade literature has generally shown that the opening of an economy to a new factor of production is good for the productivity of that economy, and because it expands the set of choices, the economy is better off and can be reorganized more efficiently,11 boosting the marginal productivity of all factors. But trade can create challenges: As production reorganizes across sectors and factors of production, some types of workers and factors of production are in lower demand, leading to redistribution. In particular, workers doing non-tradeable tasks win, while those making goods that are now offshored lose when the new factor is imported. But when there is sufficient displacement from the tradable sector, workers in the non-tradable sector can also be harmed.
“When harm from trade is possible, a standard policy response is a tariff on the imported factor, which helps the workers who are displaced by the imports. We argue that the intuition that tariffs help displaced workers may not carry over to the case of AI.”
When harm from trade is possible, a standard policy response is a tariff on the imported factor, which helps the workers who are displaced by the imports. We argue that the intuition that tariffs help displaced workers may not carry over to the case of AI. First, worker groups who remain employed by AI-powered sectors may lose if tariffs reduce output and thus employment. Second, if AI has led to a full reorganization of production to automate away jobs, small or moderate increases in the price of AI may not induce firms to bring back the types of workers who were automated. Third, an increase in the price of AI may feed back into consumer prices, which makes workers worse off.
This is just one example where some of the traditional insights may not apply. Opening to trade in a new factor usually involves a limited share of the economy, relatively small changes in factor prices, and local, rather than global, changes in production technologies. Standard results also depend on the assumption that the technology is sold at cost. If instead the new factor has one owner with market power, and the new factor can entirely substitute for labor across a large share of the economy, these models no longer deliver the right intuition.
The impact of market power in AI
In our companion paper,12 we focus on the impact of AI on two sectors of labor, which we refer to as skilled and unskilled, although those labels may not take on conventional meanings here. “Unskilled” is defined not as a level of education or income, but rather by whether the worker can be replaced with AI, while a skilled worker is less replaceable. Starting with the introduction of competitively provided AI, the direct impact of AI on a worker type depends on whether AI is a substitute or a complement for labor in that sector. A worker who is more of a substitute will be displaced when AI is available. In our general equilibrium model, it matters both whether the workers can be productive in other sectors and also whether they are substitutes or complements for other types of workers in other sectors. If they are substitutes for workers in other sectors, their displacement drives down the wages of all workers in other sectors. Our insights, derived from a general equilibrium model, complement the theoretical and empirical work of Autor and Thompson,13 who focus on the question of whether, after AI automates some tasks, the requirements for the job remain differentiated (i.e., require expertise).
“In our general equilibrium model, it matters both whether the workers can be productive in other sectors and also whether they are substitutes or complements for other types of workers in other sectors.”
Under competition, AI may lead to the distribution and reallocation challenges described above. However, it is more likely that the value creation (enabling economies to make more with less) leads to an overall increase in a country’s welfare. Even in the case of competitive provision, we note that there may be externalities created by firm adoption decisions. A firm may be just indifferent between hiring workers and purchasing AI, but from the country’s perspective, hiring workers recirculates profits in the domestic economy, while paying a foreign provider trades off increased consumption (through efficiency) against decreased national income.
We then compare the competitive benchmark to one with market power in AI, where prices may be increased just to the point where firms are indifferent about adopting. We note that if AI has been adopted, wages may be lower than without AI, and we account for this when analyzing the incentives for a firm to “deviate” by not adopting AI. We also assume that the AI provider considers the effect of its high prices on national income (and indirectly demand for its product) when setting its price.
“A firm may be just indifferent between hiring workers and purchasing AI, but from the country’s perspective, hiring workers recirculates profits in the domestic economy, while paying a foreign provider trades off increased consumption (through efficiency) against decreased national income.”
We show that an increase in AI prices often hurts all workers in an economy, especially when displaced workers are substitutes for other workers across sectors. Further, when prices are set so that firms are just indifferent about adopting AI, under reasonable conditions, a country may be worse off as a whole after the introduction of AI, and tariffs imposed after adoption can make things worse. In addition, displaced worker groups may be hit especially hard by the double whammy of reduced wages and high consumer prices. We further analyze the role of two-part tariffs, showing that if a country attempts to regulate input prices of AI, it will be important to regulate both fixed and marginal fees, as a profit-maximizing monopolist provider of AI will use both.
What Policies Address These Challenges?
Outcomes where consumers are left worse off are not inevitable. There are many policy instruments that can be used to protect competition and the benefits it provides, and other policy instruments can benefit workers directly while increasing the productivity of the economy.
There may be opportunities to create public policies that achieve multiple objectives at once: (1) increase entrepreneurship and the contribution of a country to the AI value chain; (2) improve the quality of critical services that contribute to well-being, long-term human capital, and growth of a country; and (3) address the specific challenges faced by displaced workers. In particular, governments can invest in initiatives (potentially powered by AI) that help workers transition into service sectors, including education, health care, elder care, childcare, or policing. In these initiatives, technology may be a complement to transitioning workers, increasing their productivity by reducing the required expertise to be productive. In many countries, growth is constrained by an insufficient number of educated professionals in certain service sectors. If AI can help workers new to a job or a field provide high-quality service, then it becomes cheaper for the government to provide these services. Governments can implement such policies through, e.g., innovation competitions, procurement, or building capability. If governments procure AI-driven services locally, they may be able to jump-start the local knowledge base, leading to more entrepreneurship in adjacent areas.
“Technology may be a complement to transitioning workers, increasing their productivity by reducing the required expertise to be productive.”
In the area of competition policy, governments can scrutinize the AI industry and competition between the different players at different levels of the value chain, looking for places where firms with market power attempt to thwart competition in the same market or in adjacent markets.14 A nation could consider creating a digital regulatory agency charged with following developments in AI and conducting studies on topics such as safety, national security, effects on labor, and any other important issue, with a mandate to consider the impact of regulations on competition. Such a regulatory agency could be tasked to review mergers, alliances, investments, and contracts between parties in the AI stack.
One policy idea that has some precedent in digital markets is to regulate interoperability between layers of the stack.15 Dominant players at each level may not choose to support interoperability because their market power is stronger without it.16 But smaller firms and entrants will gain from being able to connect to the neighboring layers of the stack in a nondiscriminatory way. When each layer of the stack has competitors and entry is possible, monopolization of any layer will be more difficult.
“A nation could consider creating a digital regulatory agency charged with following developments in AI and conducting studies on topics such as safety, national security, effects on labor, and any other important issue, with a mandate to consider the impact of regulations on competition.”
AI-based coding tools could also be used to port software products from one cloud or chipset to another, or applications from one foundation model to another, thereby reducing switching costs.17 Lower switching costs typically intensify competition and lower markups, to the benefit of consumers. Although the private sector will likely be incentivized to create such “portability” tools, if there are gaps, public policy might create explicit incentives or funding for the creation of such tools.
Governments can choose pro-competitive industrial policy in the area of AI. At least as they exist today, open-source models can be downloaded by users and used indefinitely; they can also be modified on local (or cloud-based) infrastructure through a process known as “fine-tuning” that can keep a model up to date with world events, research discoveries, and language. Since using the models only requires the user to pay for computational costs, the availability of “good-enough” open-source models puts an upper bound on the price that proprietary foundation model providers can charge for more basic functionality—although this depends on how the relative quality evolves over time. Governments can promote innovation in “open” foundation models or sponsor research that enables large-scale testing of open-source models for safety or other public policy goals.
Permanent loss of domestic production capabilities is another risk. AI may displace domestic economic activity that is costly to rebuild because it involves learning-by-doing or organizational capital. There may be agglomeration effects in the supply chain that will be difficult to reproduce once lost. In that case, dependence on a supplier of AI will increase. That supplier could be foreign, or it could be an individual with goals that do not align with those of the elected government. A country’s ability to bargain over and regulate AI prices may deteriorate if its outside option degrades because it can no longer produce goods and services alone. Governments may need to invest in maintaining local capabilities and consider how to maintain multiple sources of AI that can be paired with local capabilities, in order to maintain longer-run resilience against the exercise of market power.
Not all global providers of AI may make investments in customizing and distributing to small countries due to fixed costs, country-specific regulations, or trade policies. Weaker local competition opens a role for industrial policy that incentivizes providers to overcome the fixed investments needed for local customization, thereby enabling downstream industries to build on the technology. In some cases, local entrepreneurs may be able to undertake the “last mile” of customization, and there may be positive (or negative) externalities that can be internalized through industrial policy.
Tax policy and distributional policies can be used to address income inequality. Because capital may have a rate of return that is much higher than in the past, and the economy may shift to be more capitalintensive, taxation of capital will be a crucial issue for public policy.18 Revenue from taxing labor-replacing technology could be used to pay for the policies needed to help workers adjust to AI or to provide economy-wide public goods. Political economy may be influenced by competition policy indirectly, since corporate power may play a role in scenarios where a small number of firms provide technology that much of the economy relies upon.
Our companion paper19 lays out a policy-relevant research agenda on these and other topics in more detail. In our view, tackling these research questions is critical for economists seeking to guide policymakers to enhance the benefits and mitigate the harms of transformative, and potentially labor-displacing, artificial intelligence.
“In our view, tackling these research questions is critical for economists seeking to guide policymakers to enhance the benefits and mitigate the harms of transformative, and potentially labordisplacing, artificial intelligence.”
1. Gracelin Baskaran, “China’s New Rare Earth and Magnet Restrictions Threaten U.S. Defense Supply Chains,” Center for Strategic and International Studies, October 9, 2025, https://www.csis.org/analysis/chinas-new-rare-earth-and-magnet-restrictions-threaten-us-defense-supply-chains; Barath Harisas and Andreas Schumacher, “Where the Chips Fall: U.S. Export Controls Under the Biden Administration from 2022 to 2024,” Center for Strategic and International Studies, December 12, 2024, https://www.csis.org/analysis/where-chips-fall-us-export-controls-under-biden-administration-2022-2024.
2. There have been many cases where regulators or courts found mobile operating system providers in violation of competition law or accepted settlements related to competition concerns, particularly revolving around behavior where phone providers used their control over their installed base of users to exclude application providers from hardware access or from distribution through their app stores. For example, see European Commission, “Commission Accepts Commitments by Apple Opening Access to ‘Tap and Go’ Technology on iPhones,” press release, July 11, 2024, https://ec.europa.eu/commission/presscorner/api/files/document/print/fin/ip_24_3706/IP_24_3706_EN.pdf; Stephen Nellis and Foo Yun Chee, “Apple Changes App Store Rules in EU to Comply with Antitrust Order,” Reuters, June 26, 2025, https://www.reuters.com/legal/litigation/apple-changes-app-store-rules-eu-comply-with-antitrust-order-2025-06-26/. In the Google Search antitrust case, a federal judge ruled that Google had maintained its monopoly in search through exclusive distribution deals with competing mobile operating system provider Apple; Unites States of America v. Google LLC, 1:20-cv-03010 (D.D.C.), Document 1033, https://cdn.arstechnica.net/wp-content/uploads/2024/08/US-v-Google-Opinion-8-5-2024.pdf.
3. Long-term exclusive business deals between browsers and related products such as search engines and toolbars have been a feature of the browser market since at least the early 2000s. In April 2025, Perplexity.ai executives testified at the remedies trial that they had faced barriers to gaining distribution for their AI product due to exclusionary conduct by Google; Karina Montoya, “Analyzing Week One of Google Search’s Antitrust Remedies Trial,” TechPolicy.Press, April 29, 2025, https://www.techpolicy.press/analyzing-week-one-of-google-searchs-antitrustremedies-trial/. Later in 2025, Google integrated its proprietary Gemini AI into the Chrome browser two weeks after the judge in the Google Search case allowed it. Shortly thereafter, Perplexity released a competing browser that integrated Perplexity’s AI service.
4. Foo Yun Chee, “EU Watchdog Probes Nvidia Hardware Bundling as It Scrutinises Run:ai Deal,” Reuters, December 4, 2024, https://www.reuters.com/technology/eu-watchdog-probes-nvidia-hardware-bundling-it-scrutinises-runai-deal-2024-12-04/.
5. Recently, Microsoft expanded its AI offerings inside the Office365 productivity suite by including access to models beyond those provided by OpenAI, specifically Claude from Anthropic. In mid-2025, OpenAI announced deals with Broadcom, NVIDIA, and AMD in close succession, each of which will provide OpenAI with chips. It may be that this strategy increases competition because it helps sustain multiple competitors in AI chips, while OpenAI may have been motivated by a concern about relying on a single chip provider for its success.
6. For a discussion of the competition executive order and policy in the Biden administration, see, e.g., Bill Baer, “Competition Policy in the Biden Administration: Translating Campaign Poetry into Governing Prose,” 2024 Milton Handler Lecture at the New York Bar Association, Brookings Institution, May 28, 2024, https://www.brookings.edu/articles/competition-policy-in-the-biden-administration-translating-campaign-poetry-into-governing-prose/. The Trump administration announced a task force looking into anticompetitive regulations in March 2025: U.S. Department of Justice, Office of Public Affairs, “Justice Department Launches Anticompetitive Regulations Task Force,” press release, March 27, 2025, https://www.justice.gov/opa/pr/justice-department-launches-anticompetitive-regulations-task-force.
7. William D. Nordhaus, “Are We Approaching an Economic Singularity? Information Technology and the Future of Economic Growth,” American Economic Journal: Macroeconomics 13, no. 1 (January 2021): 299–332, https://doi.org/10.1257/mac.20170105.
8. Susan Athey and Fiona Scott Morton, “Artificial Intelligence, Competition, and Welfare,” Working Paper No. 34444 (National Bureau of Economic Research, 2025), https://doi.org/10.3386/w34444.
9. Athey and Scott Morton, “Artificial Intelligence, Competition, and Welfare.”
10. The offshoring literature has a large set of additional insights beyond what we discuss here that also may apply to the introduction of AI, many of which could be tailored to the problem of AI in future work.
11. Gene M. Grossman and Esteban Rossi-Hansberg, “Trading Tasks: A Simple Theory of Offshoring,” American Economic Review 98, no. 5 (2008): 1978–1997, https://doi.org/10.1257/aer.98.5.1978.
12. Athey and Scott Morton, “Artificial Intelligence, Competition, and Welfare.”
13. David Autor and Neil Thompson, “Expertise,” Working Paper No. 33941 (National Bureau of Economic Research, June 2025), https://www.nber.org/papers/w33941.
14. See endnotes 2–4 for examples.
15. The Digital Markets Act in the EU takes this approach to existing digital platforms; Fiona M. Scott Morton, Gregory S. Crawford, Jacques Crémer, David Dinielli, Amelia Fletcher, Paul Heidhues, and Monika Schnitzer, “Equitable Interoperability: The ‘Supertool’ of Digital Platform Governance,” Yale Journal on Regulation 40, no. 3 (2025): 1013–1055, https://www.yalejreg.com/print/equitable-interoperability-the-supertool-of-digital-platform-governance/.
16. See, e.g., Leevi Saarie, “The Rise and Fall of Nvidia’s Geopolitical Strategy,” TechPolicy Press, May 6, 2024, https://www.techpolicy.press/the-rise-and-fall-of-nvidias-geopolitical-strategy; the theory of this type of behavior is analyzed in Susan Athey and Fiona Scott Morton, “Platform Annexation,” Antitrust Law Journal 86, no. 3 (2023): 839–882, https://doi.org/10.2139/ssrn.3786434.
17. Although today there appear to be minimal switching costs from moving between LLM providers, over time this may change. And generally, “prompts” that work well for one model may not work for another, necessitating thorough testing of applications when they switch models.
18. E.g., William D. Nordhaus, “Are We Approaching an Economic Singularity?”; Daron Acemoglu and Pascual Restrepo, “Does the U.S. Tax Code Favor Automation?,” Brookings Papers on Economic Activity, 51, no. 1 (Spring 2020), 231–300, https://www.brookings.edu/wp-content/uploads/2020/12/Acemoglu-FINAL-WEB.pdf.
19. Athey and Scott Morton, “Artificial Intelligence, Competition, and Welfare.”
