In this commentary, Professor Ibrahim Ozturk examines the political economy of artificial intelligence (AI), highlighting its productive potential alongside distributional inequalities, financial vulnerabilities, and environmental costs. Distinguishing infrastructure investment from AI’s diffusion across the economy, he explores how productivity gains affect employment, wages, prices, inflation, and market concentration. The analysis considers how data governance, computing infrastructure, and financing arrangements shape the distribution of benefits and risks. Professor Ozturk also examines how labor displacement and economic insecurity may create opportunities for authoritarian populist mobilization, while AI-enabled surveillance raises concerns about democratic accountability. He advocates policies that distribute technological gains more equitably, strengthen institutional oversight, and protect household security. The central benchmark is a decent, dependable standard of living throughout employment transitions and retirement.
By Ibrahim Ozturk
Artificial intelligence (AI) is rapidly reshaping the global economy, offering unprecedented opportunities for productivity growth, innovation, and economic transformation while raising profound questions about inequality, financial stability, environmental sustainability, and democratic governance. As investment in AI infrastructure accelerates, the diffusion of its applications across the real economy presents complex challenges concerning employment, wages, prices, inflation, and the distribution of economic gains. While AI can expand productive capacity and reduce costs, its benefits may become concentrated among powerful firms and capital owners, potentially transferring financial and environmental burdens to households.
These developments also raise concerns about market concentration, labor displacement, economic insecurity, and the political conditions that may facilitate authoritarian populist mobilization. Against this backdrop, this article examines the interconnected macroeconomic, distributional, financial, environmental, and institutional dimensions of AI, emphasizing the importance of effective governance, equitable access to technological benefits, and policies that protect economic security. The central challenge is ensuring that technological progress translates into shared prosperity, allowing households to maintain a decent and dependable standard of living throughout employment transitions and into retirement.
An AI system may help a worker complete a task in half the time. Who gains the other half? A firm may use the savings to lower prices, improve its product, reduce working hours, or cut its payroll. Its technology supplier may capture part of the gain through subscription fees, while shareholders benefit from higher margins. The same improvement in technical capability can therefore produce very different outcomes for workers, consumers, and investors. Those outcomes matter politically because they determine who can withstand disruption and who must depend on the discretion of employers, platforms, or public officials.
The distribution of these gains depends on two processes that unfold at different speeds. The investment cycle commits capital to data centers, chips, power, and networks before their revenues are known. The diffusion cycle depends on whether firms and public institutions can reorganize production around useful applications. Their interaction determines when costs are incurred, when productivity gains appear, and whether firms, households, or public institutions capture them.
Historical Lessons from Technological Transitions
The historical record of major technological transformations provides critical empirical and theoretical insights into the current AI trajectory. Joseph Schumpeter’s paradigm of “creative destruction” details how radical innovations continuously revolutionize the economic structure from within, destroying old industries and creating new ones. However, as Karl Polanyi famously demonstrated in his analysis of the Industrial Revolution, the expansion of the self-regulating market inflicts severe social dislocation during the transition. The pivotal political-economic question centers on the temporal duration of this transition: How long does the period of destruction last before new livelihoods emerge, and who possesses the financial resilience to endure it?
Britain’s Industrial Revolution offers a sobering historical precedent known as “Engels’ Pause.” As Allen documents, for nearly half a century after the initial adoption of industrial technologies, output per worker expanded significantly while real wages remained stagnant. Corporate profits claimed an escalating share of national income, accumulating to early industrial capitalists while working-class households suffered prolonged wage depression and deteriorating living conditions. Even if real wages eventually rise in the long run, the intergenerational timing of gains is central: workers displaced in one decade cannot feed their families on the theoretical equilibrium of a subsequent generation. Similarly, the rollout of electricity as a general-purpose technology (GPT) demonstrates that macro-level productivity acceleration requires extensive, multi-decade investments in complementary machinery, organizational workflows, and human capital. As Perez emphasizes, financial capital frequently accelerates the installation phase of new technological paradigms, driving speculative asset valuations long before real-economy productivity gains materialize.
The central historical challenge for AI is whether public institutions, labor markets, and non-tech sectors can cultivate human-centric applications as quickly as financial markets build out computational capacity. As Acemogluargues, technological trajectories are not preordained; innovation can be explicitly directed toward augmenting human capabilities rather than merely automating labor to cut immediate payroll costs.
Productivity, Prices and the Cost of Living
The task framework developed by Acemoglu and Restrepo distinguishes between automation that displaces labor and innovation that creates new tasks for people. An occupation often contains both kinds of opportunity. A nurse may spend less time on documentation and more time with patients. A firm could instead use the saved hours to raise staffing ratios. Which outcome follows depends on management decisions, demand for services, and workers’ and service users’ ability to influence deployment. The ILO’s occupational exposure estimates accordingly help identify where tasks may change. Actual displacement depends on adoption, cost, regulation, and work organization.
Evidence of complementarity deserves attention. In a study of customer support, AI assistance raised productivity by about 15 percent on average, with larger benefits for less experienced and lower-skilled agents. Access to advice previously available mainly through experience can help workers learn and perform more complex tasks. At the same time, a firm can use the same technology to change staffing levels or monitor individual performance. A productivity increase observed in one workplace tells us what the tool achieved in that setting. The division of the gain among wages, profits, and consumers is a separate empirical question. Estimates of aggregate productivity are sensitive to the share of tasks affected, the savings actually realized, and the cost of implementation.
The aggregate effect depends on the economic weight of the tasks affected, the savings realized after implementation costs and the scale of adoption. At the firm level, faster completion of a task raises output only when the task constrains production or the resources released can be put to productive use. A hospital that reduces administrative work may be able to see more patients if staff time is the binding constraint. If beds or specialist capacity are scarce, the gain may instead appear in shorter paperwork queues or better service quality. Investment in software, training, and organizational redesign also absorbs resources during the transition.
Control over specialized chips, cloud capacity, models, and distribution channels can influence where profits accrue. Large firms can capture substantial gains when other firms depend on their infrastructure or face high switching costs. Historical evidence on superstar firms links shifts in market share toward firms with lower labor shares to the aggregate decline in labor’s share. In AI markets, the relevant competition questions concern access to inputs and the terms on which firms can switch suppliers or reach customers. Control over training data and intellectual property adds another barrier: contested access to copyrighted works and uncertain licensing terms can favor incumbents with legal and computing scale. High market valuations indicate expectations about future earnings. Evidence of durable rents requires examining realized returns, entry barriers, and the scope for competition.
A decline in unit costs can lower prices when customers have credible alternatives and firms compete for their business. The same gain can widen margins or be captured upstream by providers of models and cloud infrastructure. The effect on the cost of living depends on the prices of goods and services that households actually buy, their budget shares and their disposable incomes. Cheaper digital services may raise consumer welfare while housing, food or utility bills continue to strain household budgets. For that reason, the relevant outcome is real purchasing power after essential expenses.
The inflation effect depends on timing. During the buildout, investment can raise demand for construction, equipment and electricity before new capacity comes online. As adoption spreads, lower unit costs and greater productive capacity can ease price pressures. Stronger investment or household spending can offset part of that effect. A one-time fall in the price of a service lowers its contribution to measured inflation during the adjustment period. A sustained change in aggregate inflation depends on how supply, demand and monetary policy respond.
Data Centers, Environmental Costs and Financial Risk
Data centers make the physical demands of AI visible. New facilities require power connections, cooling, and land. Their local effects depend on the energy mix, water availability, and utility tariff design. A utility may commit to network upgrades before a facility begins operating. If the project is delayed or uses less capacity than contracted, the allocation of the resulting costs depends on the tariff and the developer’s commitments. Local authorities should distinguish temporary construction jobs from continuing employment and compare expected tax receipts with abatements and the cost of public infrastructure. The International Energy Agency’s (IEA) analysis underlines the importance of location and grid capacity when assessing data center electricity demand.
Political resistance reaches beyond immediate neighbors. In a May 2026 survey reported by Axios, 49 percent of registered voters supported a temporary moratorium on new data center construction. Among respondents opposed to data centers, 8 percent said they knew of a facility near their home. That figure measures reported awareness of proximity. In July, Navigator Research found that 65 percent of respondents in its survey of competitive congressional districts agreed with a statement accusing AI executives of using their wealth to influence politics for their own benefit. Taken together, these results are consistent with broader distrust of how the investment boom is governed. Local referendums and campaigns in the November midterms gave that distrust a political outlet.
The scale of financing makes the distribution of risk equally important. Van Nieuwerburgh estimates that a scenario in which the United States builds 188 GW of AI capacity by 2032 would require nearly $9 trillion in investment, equivalent to about 3.2 percent of annual GDP over 2025–2032. The estimate depends on the assumed capacity and costs. Project loans, leases, joint ventures, and private credit spread funding across several parties, and their terms determine where losses fall. An investor who has yet to commit can defer spending. A leveraged project under construction can face continued interest charges while a grid connection or permit delays revenue.
Accounting depreciation spreads an asset’s recorded cost over its useful life. Investment recovery depends on cash flow, while debt maturity determines when creditors must be paid or refinanced. These schedules can diverge. A building and its power connection may serve multiple generations of equipment. Accelerators can lose economic value more quickly as newer chips become available. The project must cover operating costs and equipment replacement as well as debt service. A project that relies on a small number of tenants is also exposed to their capacity to honor commitments. If borrowing costs rise as utilization disappoints, the financing problem can become acute well before the site loses productive value.
Janeway separates the productive legacy of speculative investment from the financial consequences of its collapse. Railways, electricity networks, and fiber networks could continue to serve users after their original investors suffered losses. AI infrastructure may have a similar afterlife, although that possibility varies across power connections, buildings, and rapidly aging chips. The applications that justify investment on a social basis can take longer to spread than the contracts that finance the investment allow. The risk to the wider economy depends on who has financed the buildout and how leveraged those institutions are. Jordà, Schularick, and Taylor (2015) and Janeway (2026) have shown that historical asset bubbles fueled by credit tended to be followed by deeper recessions. For AI, that finding motivates scrutiny of banks’ direct and indirect exposures, including through private credit. It leaves the scale of those exposures to be established empirically.
A useful test is to compare the operator’s expected return with the project’s wider benefits and costs. The latter include productivity in sectors using AI, net employment and tax benefits in host communities, the resources drawn from local power and water systems, and the opportunity cost of public subsidies. A profitable project can impose costs on households if tariffs transfer network expenses to them. An unprofitable project can leave useful infrastructure behind. The terms of approval and financing determine how much of either outcome is shared with the public.
Electricity and water also link environmental constraints to financing. A delayed grid connection can leave a completed facility unable to earn revenue while interest continues to accrue. If the utility has already expanded capacity and the project is canceled, tariffs and contracts determine whether the developer, the utility or other ratepayers bear the cost. Water availability must be assessed at the basin level, and emissions depend on the generation mix and the power actually added to serve the load.
Income Distribution, Wealth and Economic Security
Four distributional questions follow. The first concerns wages among workers. The second concerns labor’s share of income relative to capital. The third concerns the ownership of assets that generate capital income. The fourth concerns where gains accrue geographically. An AI tool can narrow skill differences among employees while shifting income towards capital owners. The wealth effects can persist if higher capital returns are reinvested and ownership remains concentrated.
International distribution raises another issue. Countries reliant on labor-intensive exports can face weaker demand for some tasks even as firms elsewhere capture automation gains. Access to affordable computing, skills, and local infrastructure can help those countries develop their own uses. Their capacity to tax internationally mobile income also influences whether productivity gains finance public services where workers bear transition costs. US–China semiconductor controls and efforts to build sovereign AI capacity raise the cost of access for third countries, making compute dependence a new channel of geopolitical and developmental inequality. Redistribution within the economies that host the largest technology firms can leave these cross-border effects largely untouched.
Labor income, wealth and security can move in different directions. Labor’s share can decline even as some real wages rise, while higher average wages can coexist with fewer routes into stable employment. Job loss, volatile hours and insecure retirement affect people’s ability to plan and withstand shocks. A distributional assessment should therefore track real disposable income after essential expenses, exposure to income losses and pension adequacy alongside wage dispersion, labor’s share and wealth concentration.
Economic Insecurity, Surveillance and Authoritarian Populism
Workers’ ability to refuse unfavorable terms depends partly on their access to other jobs, savings, and public support. A technological shock can weaken those alternatives before it reduces headline employment. Workers may accept worse conditions because replacement seems more plausible. Younger entrants can face fewer opportunities to gain experience if employers automate the tasks through which they previously learned. The distribution of these risks matters for political trust, particularly where the costs fall on communities that already have few alternatives.
Evidence from previous waves of automation identifies one pathway. Anelli, Colantone, and Stanig find that exposure to industrial robots increased support for the radical right in their study of 13 Western European countries. The study concerns a particular kind of automation and a particular political context. It offers a mechanism to test in the AI transition. Parties can frame economic insecurity as a loss of status or control and promise protection through exclusion or personal loyalty. The strength of that appeal depends on local institutions, political organization, and the narratives that parties and media advance. The backlash against data centers likewise can lead to ordinary democratic demands for transparency and cost sharing. Opposition becomes relevant to authoritarian populism when political actors mobilize it against pluralism or use selective protection to reward supporters.
A second pathway concerns the tools available to governments. Beraja et al. document how procurement of facial recognition technology in China helped strengthen political control and supported further development of the technology. Public demand for surveillance can therefore help create a market for capabilities that reinforce the incumbent’s power. This evidence concerns a specific regime and technology. It illustrates why oversight of procurement, access to personal data, and avenues of legal challenge matter as AI spreads into public administration.
The two pathways can interact when citizens with fewer economic alternatives also face more intrusive monitoring. If access to work or public benefits is mediated by opaque scoring systems, challenging a decision may become costly for people who rely on it. Where the government controls access to benefits or tolerates selective enforcement, it can use economic insecurity to reward loyalty. Opaque models can also amplify bias in credit, hiring and public-benefit decisions, while synthetic media can lower the cost of electoral deception and weaken trust in public evidence. Rights to explanation, appeal, independent review and provenance disclosures matter because they make both automated decisions and manipulated information contestable.
Policies for Sharing Gains and Protecting Independence
The structure of the AI economy creates strong pressures toward faster wealth concentration and a more unequal distribution of income. A central question for democratic political economy is whether these pressures can coexist with a decent and dependable standard of living for the broad majority. That standard has concrete implications: households should be able to meet ordinary expenses without persistent anxiety, withstand a spell of unemployment or a temporary loss of earnings without financial distress, and look ahead to retirement with reasonable confidence. The policy test is whether people can count on that level of security throughout their lives as technological change alters the jobs and income sources on which they depend.
This distinction matters for understanding the appeal of authoritarian populism. When people lose confidence that established institutions can offer a credible path to a secure life, promises of protection from leaders who claim to speak for those left behind can gain force. A dependable income floor, backed by enforceable social rights and adequate pensions, would help restore the material basis for confidence in democratic institutions. How widely that security is shared should be a central criterion for judging the distributional consequences of the AI transition.
The immediate task is to align investment decisions with the costs they create. Utilities and regulators can require large users to cover attributable network upgrades and establish credible payment commitments if demand falls short. Local authorities can publish expected fiscal costs and benefits, assess water use in the affected basin, and tie incentives to measurable commitments. Financial supervisors need visibility into the routes by which leveraged projects can transmit losses. Separately, AI agents linked to payments, national accounts or critical infrastructure introduce operational and cyber risks: unauthorized instructions, data leakage and correlated model failures can transmit shocks even without excessive leverage. These measures clarify who bears the risk before the public is asked to absorb it.
Tax policy can then address who receives the gains. The case for reviewing the treatment of capital income and economic rents becomes stronger if automation shifts the tax base away from wages while raising demand for social protection. The design should distinguish returns needed to finance investment from rents sustained by market power or privileged access to scarce resources. The IMF staff analysis favors strengthening capital income taxation over a general tax on generative AI use. Models that find a role for taxing automation show that the result depends on transition conditions and workers’ ability to move into new occupations. International coordination can help keep mobile profits within reach of taxation, while leaving room for governments to choose how to support local adjustment.
The use of revenue is as important as its collection. Portable benefits and income support can allow workers to change jobs without losing basic security. Education systems must also prepare entrants for AI-complementary judgment, digital literacy, and transferable skills; reskilling is more valuable when it connects to actual demand, and workers can afford to take it up. Collective bargaining can give employees a say in how productivity gains translate into pay, working time, or the quality of services. Broad-based capital ownership, including through independently governed public or worker funds, could extend access to investment returns. Such arrangements require transparent rules for governance and distribution.
Public institutions also influence which applications are developed. Open procurement and access to shared computing resources can expand the pool of firms and researchers able to develop applications. Governments can use procurement to specify observable goals such as reducing administrative burdens, improving service quality, or giving professionals more time with patients and students. Risk-tiered audits, human oversight, data provenance and enforceable rights to challenge automated decisions are especially important when the public cannot choose another provider. Copyright licensing and creator remuneration should be addressed without allowing rights aggregation to entrench dominant platforms. The test is whether these arrangements increase the capacity of people and institutions to make their own decisions as they adopt the technology.
In conclusion, the AI transition will be shaped by the gap between the pace of infrastructure investment and the rate at which useful applications spread. Investment can strain power systems and balance sheets before productivity gains are realized. As those gains arrive, competition, ownership, and bargaining institutions will determine how much reaches households through prices, wages, and public services. The democratic consequences depend partly on whether people can absorb disruption without becoming dependent on discretionary protection. A credible response must link investment risk to the distribution of returns, safeguard critical systems and epistemic integrity, and preserve households’ ability to plan their lives. Without institutions that govern ownership, data, and public accountability, productivity alone cannot guarantee shared prosperity or democratic resilience.
Governing AI’s Creative Destruction: Shared Prosperity, Financial Risks, and Democratic Resilience
In this commentary, Professor Ibrahim Ozturk examines the political economy of artificial intelligence (AI), highlighting its productive potential alongside distributional inequalities, financial vulnerabilities, and environmental costs. Distinguishing infrastructure investment from AI’s diffusion across the economy, he explores how productivity gains affect employment, wages, prices, inflation, and market concentration. The analysis considers how data governance, computing infrastructure, and financing arrangements shape the distribution of benefits and risks. Professor Ozturk also examines how labor displacement and economic insecurity may create opportunities for authoritarian populist mobilization, while AI-enabled surveillance raises concerns about democratic accountability. He advocates policies that distribute technological gains more equitably, strengthen institutional oversight, and protect household security. The central benchmark is a decent, dependable standard of living throughout employment transitions and retirement.
By Ibrahim Ozturk
Artificial intelligence (AI) is rapidly reshaping the global economy, offering unprecedented opportunities for productivity growth, innovation, and economic transformation while raising profound questions about inequality, financial stability, environmental sustainability, and democratic governance. As investment in AI infrastructure accelerates, the diffusion of its applications across the real economy presents complex challenges concerning employment, wages, prices, inflation, and the distribution of economic gains. While AI can expand productive capacity and reduce costs, its benefits may become concentrated among powerful firms and capital owners, potentially transferring financial and environmental burdens to households.
These developments also raise concerns about market concentration, labor displacement, economic insecurity, and the political conditions that may facilitate authoritarian populist mobilization. Against this backdrop, this article examines the interconnected macroeconomic, distributional, financial, environmental, and institutional dimensions of AI, emphasizing the importance of effective governance, equitable access to technological benefits, and policies that protect economic security. The central challenge is ensuring that technological progress translates into shared prosperity, allowing households to maintain a decent and dependable standard of living throughout employment transitions and into retirement.
An AI system may help a worker complete a task in half the time. Who gains the other half? A firm may use the savings to lower prices, improve its product, reduce working hours, or cut its payroll. Its technology supplier may capture part of the gain through subscription fees, while shareholders benefit from higher margins. The same improvement in technical capability can therefore produce very different outcomes for workers, consumers, and investors. Those outcomes matter politically because they determine who can withstand disruption and who must depend on the discretion of employers, platforms, or public officials.
The distribution of these gains depends on two processes that unfold at different speeds. The investment cycle commits capital to data centers, chips, power, and networks before their revenues are known. The diffusion cycle depends on whether firms and public institutions can reorganize production around useful applications. Their interaction determines when costs are incurred, when productivity gains appear, and whether firms, households, or public institutions capture them.
Historical Lessons from Technological Transitions
The historical record of major technological transformations provides critical empirical and theoretical insights into the current AI trajectory. Joseph Schumpeter’s paradigm of “creative destruction” details how radical innovations continuously revolutionize the economic structure from within, destroying old industries and creating new ones. However, as Karl Polanyi famously demonstrated in his analysis of the Industrial Revolution, the expansion of the self-regulating market inflicts severe social dislocation during the transition. The pivotal political-economic question centers on the temporal duration of this transition: How long does the period of destruction last before new livelihoods emerge, and who possesses the financial resilience to endure it?
Britain’s Industrial Revolution offers a sobering historical precedent known as “Engels’ Pause.” As Allen documents, for nearly half a century after the initial adoption of industrial technologies, output per worker expanded significantly while real wages remained stagnant. Corporate profits claimed an escalating share of national income, accumulating to early industrial capitalists while working-class households suffered prolonged wage depression and deteriorating living conditions. Even if real wages eventually rise in the long run, the intergenerational timing of gains is central: workers displaced in one decade cannot feed their families on the theoretical equilibrium of a subsequent generation. Similarly, the rollout of electricity as a general-purpose technology (GPT) demonstrates that macro-level productivity acceleration requires extensive, multi-decade investments in complementary machinery, organizational workflows, and human capital. As Perez emphasizes, financial capital frequently accelerates the installation phase of new technological paradigms, driving speculative asset valuations long before real-economy productivity gains materialize.
The central historical challenge for AI is whether public institutions, labor markets, and non-tech sectors can cultivate human-centric applications as quickly as financial markets build out computational capacity. As Acemogluargues, technological trajectories are not preordained; innovation can be explicitly directed toward augmenting human capabilities rather than merely automating labor to cut immediate payroll costs.
Productivity, Prices and the Cost of Living
The task framework developed by Acemoglu and Restrepo distinguishes between automation that displaces labor and innovation that creates new tasks for people. An occupation often contains both kinds of opportunity. A nurse may spend less time on documentation and more time with patients. A firm could instead use the saved hours to raise staffing ratios. Which outcome follows depends on management decisions, demand for services, and workers’ and service users’ ability to influence deployment. The ILO’s occupational exposure estimates accordingly help identify where tasks may change. Actual displacement depends on adoption, cost, regulation, and work organization.
Evidence of complementarity deserves attention. In a study of customer support, AI assistance raised productivity by about 15 percent on average, with larger benefits for less experienced and lower-skilled agents. Access to advice previously available mainly through experience can help workers learn and perform more complex tasks. At the same time, a firm can use the same technology to change staffing levels or monitor individual performance. A productivity increase observed in one workplace tells us what the tool achieved in that setting. The division of the gain among wages, profits, and consumers is a separate empirical question. Estimates of aggregate productivity are sensitive to the share of tasks affected, the savings actually realized, and the cost of implementation.
The aggregate effect depends on the economic weight of the tasks affected, the savings realized after implementation costs and the scale of adoption. At the firm level, faster completion of a task raises output only when the task constrains production or the resources released can be put to productive use. A hospital that reduces administrative work may be able to see more patients if staff time is the binding constraint. If beds or specialist capacity are scarce, the gain may instead appear in shorter paperwork queues or better service quality. Investment in software, training, and organizational redesign also absorbs resources during the transition.
Control over specialized chips, cloud capacity, models, and distribution channels can influence where profits accrue. Large firms can capture substantial gains when other firms depend on their infrastructure or face high switching costs. Historical evidence on superstar firms links shifts in market share toward firms with lower labor shares to the aggregate decline in labor’s share. In AI markets, the relevant competition questions concern access to inputs and the terms on which firms can switch suppliers or reach customers. Control over training data and intellectual property adds another barrier: contested access to copyrighted works and uncertain licensing terms can favor incumbents with legal and computing scale. High market valuations indicate expectations about future earnings. Evidence of durable rents requires examining realized returns, entry barriers, and the scope for competition.
A decline in unit costs can lower prices when customers have credible alternatives and firms compete for their business. The same gain can widen margins or be captured upstream by providers of models and cloud infrastructure. The effect on the cost of living depends on the prices of goods and services that households actually buy, their budget shares and their disposable incomes. Cheaper digital services may raise consumer welfare while housing, food or utility bills continue to strain household budgets. For that reason, the relevant outcome is real purchasing power after essential expenses.
The inflation effect depends on timing. During the buildout, investment can raise demand for construction, equipment and electricity before new capacity comes online. As adoption spreads, lower unit costs and greater productive capacity can ease price pressures. Stronger investment or household spending can offset part of that effect. A one-time fall in the price of a service lowers its contribution to measured inflation during the adjustment period. A sustained change in aggregate inflation depends on how supply, demand and monetary policy respond.
Data Centers, Environmental Costs and Financial Risk
Data centers make the physical demands of AI visible. New facilities require power connections, cooling, and land. Their local effects depend on the energy mix, water availability, and utility tariff design. A utility may commit to network upgrades before a facility begins operating. If the project is delayed or uses less capacity than contracted, the allocation of the resulting costs depends on the tariff and the developer’s commitments. Local authorities should distinguish temporary construction jobs from continuing employment and compare expected tax receipts with abatements and the cost of public infrastructure. The International Energy Agency’s (IEA) analysis underlines the importance of location and grid capacity when assessing data center electricity demand.
Political resistance reaches beyond immediate neighbors. In a May 2026 survey reported by Axios, 49 percent of registered voters supported a temporary moratorium on new data center construction. Among respondents opposed to data centers, 8 percent said they knew of a facility near their home. That figure measures reported awareness of proximity. In July, Navigator Research found that 65 percent of respondents in its survey of competitive congressional districts agreed with a statement accusing AI executives of using their wealth to influence politics for their own benefit. Taken together, these results are consistent with broader distrust of how the investment boom is governed. Local referendums and campaigns in the November midterms gave that distrust a political outlet.
The scale of financing makes the distribution of risk equally important. Van Nieuwerburgh estimates that a scenario in which the United States builds 188 GW of AI capacity by 2032 would require nearly $9 trillion in investment, equivalent to about 3.2 percent of annual GDP over 2025–2032. The estimate depends on the assumed capacity and costs. Project loans, leases, joint ventures, and private credit spread funding across several parties, and their terms determine where losses fall. An investor who has yet to commit can defer spending. A leveraged project under construction can face continued interest charges while a grid connection or permit delays revenue.
Accounting depreciation spreads an asset’s recorded cost over its useful life. Investment recovery depends on cash flow, while debt maturity determines when creditors must be paid or refinanced. These schedules can diverge. A building and its power connection may serve multiple generations of equipment. Accelerators can lose economic value more quickly as newer chips become available. The project must cover operating costs and equipment replacement as well as debt service. A project that relies on a small number of tenants is also exposed to their capacity to honor commitments. If borrowing costs rise as utilization disappoints, the financing problem can become acute well before the site loses productive value.
Janeway separates the productive legacy of speculative investment from the financial consequences of its collapse. Railways, electricity networks, and fiber networks could continue to serve users after their original investors suffered losses. AI infrastructure may have a similar afterlife, although that possibility varies across power connections, buildings, and rapidly aging chips. The applications that justify investment on a social basis can take longer to spread than the contracts that finance the investment allow. The risk to the wider economy depends on who has financed the buildout and how leveraged those institutions are. Jordà, Schularick, and Taylor (2015) and Janeway (2026) have shown that historical asset bubbles fueled by credit tended to be followed by deeper recessions. For AI, that finding motivates scrutiny of banks’ direct and indirect exposures, including through private credit. It leaves the scale of those exposures to be established empirically.
A useful test is to compare the operator’s expected return with the project’s wider benefits and costs. The latter include productivity in sectors using AI, net employment and tax benefits in host communities, the resources drawn from local power and water systems, and the opportunity cost of public subsidies. A profitable project can impose costs on households if tariffs transfer network expenses to them. An unprofitable project can leave useful infrastructure behind. The terms of approval and financing determine how much of either outcome is shared with the public.
Electricity and water also link environmental constraints to financing. A delayed grid connection can leave a completed facility unable to earn revenue while interest continues to accrue. If the utility has already expanded capacity and the project is canceled, tariffs and contracts determine whether the developer, the utility or other ratepayers bear the cost. Water availability must be assessed at the basin level, and emissions depend on the generation mix and the power actually added to serve the load.
Income Distribution, Wealth and Economic Security
Four distributional questions follow. The first concerns wages among workers. The second concerns labor’s share of income relative to capital. The third concerns the ownership of assets that generate capital income. The fourth concerns where gains accrue geographically. An AI tool can narrow skill differences among employees while shifting income towards capital owners. The wealth effects can persist if higher capital returns are reinvested and ownership remains concentrated.
International distribution raises another issue. Countries reliant on labor-intensive exports can face weaker demand for some tasks even as firms elsewhere capture automation gains. Access to affordable computing, skills, and local infrastructure can help those countries develop their own uses. Their capacity to tax internationally mobile income also influences whether productivity gains finance public services where workers bear transition costs. US–China semiconductor controls and efforts to build sovereign AI capacity raise the cost of access for third countries, making compute dependence a new channel of geopolitical and developmental inequality. Redistribution within the economies that host the largest technology firms can leave these cross-border effects largely untouched.
Labor income, wealth and security can move in different directions. Labor’s share can decline even as some real wages rise, while higher average wages can coexist with fewer routes into stable employment. Job loss, volatile hours and insecure retirement affect people’s ability to plan and withstand shocks. A distributional assessment should therefore track real disposable income after essential expenses, exposure to income losses and pension adequacy alongside wage dispersion, labor’s share and wealth concentration.
Economic Insecurity, Surveillance and Authoritarian Populism
Workers’ ability to refuse unfavorable terms depends partly on their access to other jobs, savings, and public support. A technological shock can weaken those alternatives before it reduces headline employment. Workers may accept worse conditions because replacement seems more plausible. Younger entrants can face fewer opportunities to gain experience if employers automate the tasks through which they previously learned. The distribution of these risks matters for political trust, particularly where the costs fall on communities that already have few alternatives.
Evidence from previous waves of automation identifies one pathway. Anelli, Colantone, and Stanig find that exposure to industrial robots increased support for the radical right in their study of 13 Western European countries. The study concerns a particular kind of automation and a particular political context. It offers a mechanism to test in the AI transition. Parties can frame economic insecurity as a loss of status or control and promise protection through exclusion or personal loyalty. The strength of that appeal depends on local institutions, political organization, and the narratives that parties and media advance. The backlash against data centers likewise can lead to ordinary democratic demands for transparency and cost sharing. Opposition becomes relevant to authoritarian populism when political actors mobilize it against pluralism or use selective protection to reward supporters.
A second pathway concerns the tools available to governments. Beraja et al. document how procurement of facial recognition technology in China helped strengthen political control and supported further development of the technology. Public demand for surveillance can therefore help create a market for capabilities that reinforce the incumbent’s power. This evidence concerns a specific regime and technology. It illustrates why oversight of procurement, access to personal data, and avenues of legal challenge matter as AI spreads into public administration.
The two pathways can interact when citizens with fewer economic alternatives also face more intrusive monitoring. If access to work or public benefits is mediated by opaque scoring systems, challenging a decision may become costly for people who rely on it. Where the government controls access to benefits or tolerates selective enforcement, it can use economic insecurity to reward loyalty. Opaque models can also amplify bias in credit, hiring and public-benefit decisions, while synthetic media can lower the cost of electoral deception and weaken trust in public evidence. Rights to explanation, appeal, independent review and provenance disclosures matter because they make both automated decisions and manipulated information contestable.
Policies for Sharing Gains and Protecting Independence
The structure of the AI economy creates strong pressures toward faster wealth concentration and a more unequal distribution of income. A central question for democratic political economy is whether these pressures can coexist with a decent and dependable standard of living for the broad majority. That standard has concrete implications: households should be able to meet ordinary expenses without persistent anxiety, withstand a spell of unemployment or a temporary loss of earnings without financial distress, and look ahead to retirement with reasonable confidence. The policy test is whether people can count on that level of security throughout their lives as technological change alters the jobs and income sources on which they depend.
This distinction matters for understanding the appeal of authoritarian populism. When people lose confidence that established institutions can offer a credible path to a secure life, promises of protection from leaders who claim to speak for those left behind can gain force. A dependable income floor, backed by enforceable social rights and adequate pensions, would help restore the material basis for confidence in democratic institutions. How widely that security is shared should be a central criterion for judging the distributional consequences of the AI transition.
The immediate task is to align investment decisions with the costs they create. Utilities and regulators can require large users to cover attributable network upgrades and establish credible payment commitments if demand falls short. Local authorities can publish expected fiscal costs and benefits, assess water use in the affected basin, and tie incentives to measurable commitments. Financial supervisors need visibility into the routes by which leveraged projects can transmit losses. Separately, AI agents linked to payments, national accounts or critical infrastructure introduce operational and cyber risks: unauthorized instructions, data leakage and correlated model failures can transmit shocks even without excessive leverage. These measures clarify who bears the risk before the public is asked to absorb it.
Tax policy can then address who receives the gains. The case for reviewing the treatment of capital income and economic rents becomes stronger if automation shifts the tax base away from wages while raising demand for social protection. The design should distinguish returns needed to finance investment from rents sustained by market power or privileged access to scarce resources. The IMF staff analysis favors strengthening capital income taxation over a general tax on generative AI use. Models that find a role for taxing automation show that the result depends on transition conditions and workers’ ability to move into new occupations. International coordination can help keep mobile profits within reach of taxation, while leaving room for governments to choose how to support local adjustment.
The use of revenue is as important as its collection. Portable benefits and income support can allow workers to change jobs without losing basic security. Education systems must also prepare entrants for AI-complementary judgment, digital literacy, and transferable skills; reskilling is more valuable when it connects to actual demand, and workers can afford to take it up. Collective bargaining can give employees a say in how productivity gains translate into pay, working time, or the quality of services. Broad-based capital ownership, including through independently governed public or worker funds, could extend access to investment returns. Such arrangements require transparent rules for governance and distribution.
Public institutions also influence which applications are developed. Open procurement and access to shared computing resources can expand the pool of firms and researchers able to develop applications. Governments can use procurement to specify observable goals such as reducing administrative burdens, improving service quality, or giving professionals more time with patients and students. Risk-tiered audits, human oversight, data provenance and enforceable rights to challenge automated decisions are especially important when the public cannot choose another provider. Copyright licensing and creator remuneration should be addressed without allowing rights aggregation to entrench dominant platforms. The test is whether these arrangements increase the capacity of people and institutions to make their own decisions as they adopt the technology.
In conclusion, the AI transition will be shaped by the gap between the pace of infrastructure investment and the rate at which useful applications spread. Investment can strain power systems and balance sheets before productivity gains are realized. As those gains arrive, competition, ownership, and bargaining institutions will determine how much reaches households through prices, wages, and public services. The democratic consequences depend partly on whether people can absorb disruption without becoming dependent on discretionary protection. A credible response must link investment risk to the distribution of returns, safeguard critical systems and epistemic integrity, and preserve households’ ability to plan their lives. Without institutions that govern ownership, data, and public accountability, productivity alone cannot guarantee shared prosperity or democratic resilience.
Latest News
Fabiola Ermer: Illiberal Political Ideas Become Socially Embedded Through Less Visible Mechanisms
Governing AI’s Creative Destruction: Shared Prosperity, Financial Risks, and Democratic Resilience
The Hood on Rhodes and the Politics of Moral Purification
Fabiola Ermer: Illiberal Political Ideas Become Socially Embedded Through Less Visible Mechanisms
Prof. Frances E. Lee: I Anticipate Very Hardball Politics in the US Over the Next Two Years
Fabiola Ermer: Illiberal Political Ideas Become Socially Embedded Through Less Visible Mechanisms
Governing AI’s Creative Destruction: Shared Prosperity, Financial Risks, and Democratic Resilience
The Hood on Rhodes and the Politics of Moral Purification
Category
Fabiola Ermer: Illiberal Political Ideas Become Socially Embedded Through Less Visible Mechanisms
Governing AI’s Creative Destruction: Shared Prosperity, Financial Risks, and Democratic Resilience
The Hood on Rhodes and the Politics of Moral Purification
Fabiola Ermer: Illiberal Political Ideas Become Socially Embedded Through Less Visible Mechanisms
Prof. Frances E. Lee: I Anticipate Very Hardball Politics in the US Over the Next Two Years
Prof. Spasojević: Without Lustration and Accountability, Serbia’s Former Rulers Will Return