Does AI Make Life Cheaper? The Productivity-Prosperity Gap

by

Introduction: The Promise of Automation

Something unusual is happening in the economy. The tools available to a typical knowledge worker in 2026 — AI assistants capable of drafting documents, synthesizing research, writing and debugging code, translating across languages, and analyzing data in seconds — represent a genuine leap in productive capacity. Tasks that once consumed hours now take minutes. Work that once required specialized expertise can be delegated, at least partially, to software available at a fraction of what equivalent human labor would have cost. By almost any measure of technical efficiency, this is an extraordinary moment.

And yet the experience of many ordinary households in the United States looks quite different. Renting an apartment in a major metropolitan area consumes a larger share of income than it did a decade ago. Healthcare premiums and medical costs continue to rise faster than wages. Published tuition at public four-year universities has risen sharply since the turn of the century, far outpacing general inflation, even if net prices after aid have grown more slowly. Groceries and utilities press harder on household budgets than the efficiency narrative alone would suggest.

This gap — between measurably rising productivity and the persistent difficulty of affording ordinary life — is what this article sets out to examine.

Defining the terms

The word productivity has a precise meaning in economics. Labor productivity refers to output per hour worked: how much an economy, a firm, or a worker produces in a given unit of time. When this rises, the same inputs generate more output. That is, in principle, the foundation of rising living standards. More efficient production should mean lower costs, lower prices, higher wages, or some combination of all three.

Prosperity is a different concept. It refers to whether people are actually better off — whether real incomes cover real expenses, whether households can meet their needs without excessive financial stress, whether the gains from economic growth reach ordinary people rather than concentrating narrowly.

The critical question this article asks is not whether productivity is rising. It is. The question is whether that productivity growth is being transmitted into genuine prosperity for typical American households — or whether the gains are being captured elsewhere, while essential costs continue to rise.

A gap that began long before AI

The divergence between productivity and typical worker outcomes is not a new problem. According to data from the U.S. Bureau of Labor Statistics, labor productivity in the nonfarm business sector has risen substantially since the early 1970s — roughly doubling over the following five decades. Real hourly compensation over the same broad period rose much more slowly, lagging productivity by a considerable margin 1, 2.

This separation was not always the pattern. BLS data indicate that from roughly 1947 to 1973, labor productivity and real hourly compensation grew at broadly similar rates — around 2.8 and 2.6 percent annually. After 1973 the relationship weakened, and in the first decade of the twenty-first century productivity grew at roughly 2.5 percent annually while compensation grew at around 1.1 percent 3, 4. The gap between efficiency and ordinary compensation appears to be a structural feature of the modern economy, not a temporary anomaly — and it predates AI by decades.

Alongside the compensation gap, labor’s share of nonfarm business output moved lower. For much of the postwar period it was comparatively stable, around the low 60s percent. Since the early 2000s that share has trended down, reaching near 58 percent by 2024 5. The shift is broad-based across advanced economies and appears linked to technology, global integration, and changing labor market institutions — though different studies assign different weights to each 6, 7, 8.

This matters for how we read the AI moment. Generative AI arrives inside an economy that already has an established tendency to route the gains from technological progress away from median workers. Whether AI changes that dynamic or deepens it is the central distributional question this article examines.

The cost-of-living side

The productivity-pay gap is only part of the story. Since 2000, the overall U.S. price level has risen by roughly 75 to 80 percent, while typical earnings appear to have improved only modestly in real terms 9, 10. The categories that dominate household budgets — housing, healthcare, and education — have risen considerably faster than headline inflation. Shelter costs appear to have roughly doubled between 2000 and 2024, outpacing overall inflation by a meaningful margin 9, 11, 12. Medical care prices and employer-sponsored family health insurance premiums have also risen far faster than general inflation 9, 13. College tuition at public four-year universities climbed from roughly $3,500–$4,000 in 2000 to around $10,000–$11,000 by the mid-2020s in published prices 14, 15.

The result is a household economy in which some things have become dramatically cheaper — software, streaming, digital communications, and increasingly AI tools themselves — while the largest fixed costs of a stable middle-class life have become significantly more expensive.

The argument ahead

Some private-sector analyses project that generative AI could add trillions of dollars in annual economic value and raise labor productivity growth by meaningful fractions of a percentage point per year over the next decade 16, 17. These projections are worth taking seriously. But they describe scenarios and potential outcomes, not realized economy-wide data. More cautious analysis, notably Acemoglu (2024), suggests aggregate TFP gains may be considerably smaller once diffusion frictions and realistic task-automation shares are accounted for 18.

More importantly, even if significant productivity gains materialize, the experience of the past several decades gives reason to ask whether those gains will flow broadly to workers and households, or concentrate at the upper end of the income and wealth distribution. This article develops that question through three main lines of argument: the historical divergence between productivity and pay; the structural cost pressures that productivity growth elsewhere does not automatically relieve; and the market and institutional conditions that determine who captures what efficiency generates.

The central claim is not that AI is harmful or that technological progress is illusory. It is more specific: productivity is not the same as prosperity, and a faster economy is not automatically a more affordable one.

From Computers to AI: A Brief History of Efficiency

The promise that technology will make work more efficient is not new. What changes from era to era is the type of work being automated and — crucially — who captures the resulting gains.

The postwar decades demonstrated that broad-based diffusion is possible. From 1947 to 1973, labor productivity and real hourly compensation grew at broadly similar rates; a more productive economy did translate, with reasonable reliability, into higher pay for workers. After 1973 that alignment weakened. In the decade from 2000 to 2009, productivity grew at roughly 2.5 percent annually while real hourly compensation grew at around 1.1 percent 3, 4.

The labor-share data tell the same story in a different register: labor’s claim on output was comparatively stable through the postwar era, then drifted lower from roughly the early 2000s onward 5. The causes are multiple — technology, globalization, institutional change — and no single explanation accounts for the full shift 6, 7, 8. The point is that the structural disconnect between productivity and broadly shared worker outcomes predates AI and is rooted in how gains move through institutional channels, not just in how technology performs.

Artificial intelligence is now being presented as the latest iteration of this pattern — a general-purpose technology with the potential to raise productivity across a wide range of tasks. Whether it proves categorically different from earlier technology waves or follows a similar arc of delayed and uneven diffusion remains genuinely open. What the historical record suggests is that productivity gains, however real, do not automatically deliver broad prosperity. The mechanisms connecting efficiency to worker outcomes — labor bargaining power, product market competition, wage-setting institutions, and public policy — matter as much as the technology itself.

The Basic Economic Expectation: Productivity Should Lower Costs

There is a straightforward economic logic behind the expectation that technological progress should make life more affordable. In standard economic reasoning, productivity growth sets off a chain of effects: when firms can produce more output with the same inputs, unit costs fall; in a competitive market, lower costs put pressure on prices; firms that pass savings to consumers attract more business. Over time, efficiency gains tend to flow through to lower prices, higher real wages, or both.

This mechanism is not theoretical fiction. Manufacturing offers an intuitive illustration: when production processes become more efficient and competitive pressure is strong, unit costs can fall and prices may follow. Digital goods suggest a more recent version of the same logic, where very low marginal costs and broad replication create conditions in which the textbook model can operate. These are not universal outcomes, but they show that the expected chain of effects is not purely hypothetical.

The optimism surrounding AI reflects a similar intuition. Some major private-sector analyses project very large potential gains from generative AI 16, 17. These projections reflect the underlying logic: more output per unit of input, multiplied across a large economy, should produce significant aggregate gains.

But the textbook model is incomplete. The chain from productivity to consumer benefit depends on conditions that do not always hold in practice. Productivity gains must actually materialize at scale — for AI specifically, this remains uncertain 18. Even when productivity rises, the pass-through to wages and prices depends on market structure and labor bargaining: U.S. productivity has risen substantially since the early 1970s, but real compensation has lagged considerably 1, 2, and OECD and NBER research finds the productivity-pay link to be meaningful but incomplete 19, 20. And the sectors where efficiency gains are easiest to achieve are not necessarily where consumers face the greatest cost pressure.

Productivity is necessary, not sufficient, for prosperity. The mechanisms that translate efficiency into broadly shared gains — competitive product markets, strong labor bargaining, appropriate public investment — shape the outcome as much as the technology itself.

The Cost-of-Living Paradox

Despite decades of productivity growth, the cost of key essentials has risen faster than overall inflation and typical wages for many U.S. households.

Housing

Housing is the single largest expenditure category for most households. BLS shelter CPI data indicate that housing costs rose roughly 100 to 120 percent from 2000 to 2024, meaningfully faster than overall inflation 9, 11, 12. Harvard Joint Center for Housing Studies reports describe renter cost burdens and price-to-income ratios as near or above historic highs by the early 2020s 11, 12.

Housing is not a textbook Baumol sector. Its cost dynamics are driven by land scarcity, restrictive zoning, construction financing conditions, and the financialization of residential property — not slow service-sector productivity growth 12, 21. AI tools in architecture or permitting may reduce friction at the margin, but they do not change zoning maps, reduce land values in constrained markets, or resolve the political economy of housing supply restriction.

Healthcare

BLS medical care CPI rose roughly 115 to 130 percent from 2000 to mid-2024, well above overall inflation 9, 22. Employer-sponsored family health insurance premiums have more than doubled over comparable long-run windows, outpacing wages over the full span 13, 23.

Part of the explanation fits a Baumol-style framework: physician consultations, nursing care, and surgical procedures require sustained human attention that is difficult to automate without compromising quality. But empirical research suggests Baumol-type mechanisms explain only 15 to 40 percent of U.S. healthcare cost growth — with institutional factors, pricing power, and administrative overhead accounting for much of the remainder 24, 25, 26. AI may reduce some administrative burden, but as long as the labor-intensive core of clinical care remains, and as long as efficiency gains are absorbed by margins rather than passed to patients, relief to household healthcare budgets is likely to remain partial 18, 26.

Education

Published tuition and fees at public four-year universities rose from roughly $3,500–$4,000 in 2000–01 to around $10,000–$11,000 by the mid-2020s — an increase far exceeding general inflation 14, 15. Net prices after aid have grown more slowly, but the gap between published and net prices does not remove the affordability challenge for students and families who do not qualify for sufficient aid. Like healthcare, higher education has significant Baumol-type characteristics: teaching and mentorship are labor-intensive activities that are difficult to replicate at quality through automation alone 27, 21.

Food, energy, and the overall picture

Food-at-home CPI rose roughly 60 to 80 percent from 2000 to 2024; energy roughly doubled over the same period with high year-to-year volatility 9, 22, 28, 10, 29. These categories are more responsive to technological change than healthcare or housing, but supply chains, climate risk, and geopolitical pressures mean that efficiency in production does not translate smoothly into stable consumer prices.

Taken together, the data present a clear picture: the sectors that claim the largest shares of household budgets have seen real costs rise substantially. The structural features that drive this — land scarcity, labor intensity in core service delivery, regulatory complexity, supply constraints — were not created by a failure to adopt automation and are not obviously resolved by the arrival of better AI tools.

Why Digital Goods Become Cheap

Before examining why essential costs persist, it is worth understanding why some parts of the economy genuinely have become cheaper. That contrast is essential for understanding what AI can and cannot do for household affordability.

The economics of digital goods rest on an asymmetry between fixed and marginal costs. Creating the first version of a piece of software or an analysis requires substantial investment; distributing it to additional users often costs very little by comparison. Research in digital economics identifies several costs that digital technology tends to reduce: search, replication, transportation, tracking, and verification 30. OECD analysis on measuring digital transformation describes how this allows average costs to fall rapidly as scale increases 31. When marginal cost approaches near zero, the average cost per user falls with scale — a platform reaching tens of millions of users can price far below what a comparable human-intensive service would require.

This cost structure is visible, though imperfectly, in official U.S. price measurement. BLS and BEA treat information technology hardware and software as areas of rapid quality change, quality-adjusting prices to reflect capability improvements — which is why real ICT prices tend to fall significantly faster than economy-wide inflation 32, 33, 34.

AI may extend the digital scale model to a wider range of information tasks — drafting, summarizing, translating, generating code — that previously required skilled human time. In that sense, AI applies the same economics that made earlier digital goods cheap to a new and broader category of information work 30, 31, 35. But this extension is conditional: AI services depend on substantial ongoing infrastructure costs (computing hardware, electricity, cloud capacity, compliance), which means they do not approach zero marginal cost in the way a replicated software file might. Market structure also matters — the fact that a service can be delivered cheaply does not guarantee it will be priced cheaply.

The digital cost story is genuine and worth acknowledging. What it does not do is address the cost pressure that dominates most household budgets. Shelter, healthcare, and education remain large, structurally expensive, and resistant to the logic that makes replicating a digital file inexpensive.

Why Essential Goods and Services Remain Expensive

Essential goods and services face structural headwinds that limit automation-led price declines. The most influential theoretical lens for understanding cost pressure in labor-intensive services is William Baumol’s 1967 framework 27. In a market economy, wages tend to equalize across sectors. If productivity grows rapidly in some sectors but slowly in labor-intensive services, the wages needed to retain workers in those services rise along with the rest of the economy — even though their output per hour has not grown proportionally. Unit costs and relative prices in slow-productivity sectors rise over time not because they are failing, but because efficiency elsewhere is pulling wages up 27, 21.

This mechanism fits healthcare, education, and care work reasonably well at a structural level. But it does not explain everything. Empirical research places Baumol-type mechanisms at roughly 15 to 40 percent of full healthcare cost growth in some cross-country estimates, with institutional factors, pricing power, and administrative overhead explaining much of the remainder 24, 25, 26.

Housing, as discussed earlier, should be treated separately. Its cost dynamics are driven by land, regulation, and finance rather than the narrow logic of stagnant service productivity. Food and energy face different constraints again: input-market volatility, geopolitical risk, and physical infrastructure limitations that make efficiency in one part of the system translate unevenly into consumer prices.

AI may reduce some cost pressure in essential services, primarily through administrative efficiency — documentation, scheduling, triage, billing — rather than the labor-intensive core of care or instruction. These are real gains. But a hospital’s cost structure does not approach zero marginal cost because its electronic health records system uses machine learning, and a classroom does not become free because an AI can grade essays. The labor, facility, equipment, and regulatory costs that dominate essential-service spending are not primarily located in the administrative overhead that AI is best positioned to reduce 18, 26.

There is also a distributional question. Even where AI-driven efficiency gains do occur in healthcare or education administration, those gains do not automatically flow to patients or students as lower prices. They may instead reduce costs for providers and institutions without being passed through — a point the later sections develop through the lens of market structure and distribution.

Labor After AI: White-Collar Compression and Blue-Collar Scarcity

Previous waves of automation primarily affected routine physical and clerical tasks. Generative AI changes that pattern: it is most capable in tasks involving language, pattern recognition, synthesis, and structured reasoning — precisely the kinds of work that define white-collar, document-intensive occupations. Drafting, coding, analysis, translation, legal research, and financial modeling all involve the cognitive processing that AI tools can increasingly replicate or augment at scale.

According to McKinsey and Goldman Sachs, roughly two-thirds of U.S. occupations face some degree of exposure to AI-driven task automation, and within exposed occupations, one-quarter to one-half of tasks could potentially be automated 16, 17. These projections describe exposure under favorable adoption assumptions, not observed displacement. Acemoglu’s macroeconomic analysis estimates realized aggregate TFP effects over the next decade at no more than around 0.66 percent cumulative — considerably below headline scenarios 18. What does seem likely is a compression of economic value in the middle tiers of white-collar employment, even without mass layoffs.

Some forms of work are considerably harder for AI to automate: physical labor requiring fine motor adaptation in unpredictable environments, and interpersonal service work involving relational quality and situational responsiveness 16, 17. This relative insulation does not, however, automatically translate into wages. Many of the jobs hardest to automate are also among the lowest-paid. The labor market does not pay primarily for being difficult to replace by machines; it pays based on supply, demand, institutional arrangements, and bargaining power. Scarcity and prosperity are not the same thing.

The more important question is not which workers are exposed, but who captures the value generated by AI-driven productivity changes. Acemoglu argues that AI is more likely to widen the capital-labor income split than to compress it, unless offset by new task creation or institutional intervention 18. McKinsey and Goldman Sachs are less pessimistic, but even they acknowledge that favorable distributional outcomes depend on adoption speed, worker transition support, and institutional capacity to channel gains broadly 16, 17. A more productive economy can coexist with stagnant real compensation for many workers if gains are captured by capital, by top-quartile earners, or by the relatively small share of the workforce that has high AI complementarity and sufficient bargaining power.

The Hidden Costs of AI Infrastructure

When a user submits a prompt to an AI tool and receives a response in seconds, the transaction can feel nearly costless. That user-facing price, however, does not describe the full economic cost of the system.

Electricity and scale

According to IEA data published in 2024, global data centers consumed approximately 415 terawatt-hours of electricity in 2024 — roughly 1.5 percent of world electricity consumption — with projections of approximately 945 TWh by 2030 36. In the United States, LBNL reports data-center electricity demand rising from roughly 58 TWh in 2014 to approximately 176 TWh in 2023 37. AI workloads account for roughly 24 percent of server electricity use and about 15 percent of total data-center electricity, and are a primary driver of incremental growth 36, 37. These are real operating costs that do not disappear simply because they are not visible in the per-query price a user pays.

Water, hardware, and capital

Cooling AI-era data centers requires water management at meaningful scale. Microsoft’s 2024 sustainability disclosures note that newer data centers are being designed to consume zero water for cooling in some configurations — an acknowledgment that water has become a real operational constraint 38, 39. Public water accounting remains inconsistent across the industry 38.

Beyond electricity and water, the AI buildout requires massive capital investment. Amazon has publicly stated its expectation of approximately $100 billion in capital expenditures in 2025, with a significant portion directed toward AI infrastructure 40. Capital expenditure on data centers and hardware is committed upfront and recovered over time through service pricing. Low per-unit prices at scale do not mean low total system costs; they mean the costs are distributed differently.

The accounting argument

The price a user pays for AI access is not the same as the cost of producing and delivering that access. The gap is covered by capital markets, by pricing models designed for adoption rather than full cost recovery, and in some cases by costs borne externally — by energy grids, by local water systems — rather than appearing in the service price. The economy as a whole is not getting AI for free. It is getting AI at a price that is structured to look low at the point of use.

The capital required to build and operate AI at scale accrues as assets and returns primarily to the firms and investors who own that infrastructure. When AI produces genuine productivity gains, those gains flow first through the companies that made the capital commitment — as higher margins, greater competitive advantage, or returns to shareholders — rather than directly to the workers or households whose tasks are being made more efficient.

Market Power: When Efficiency Does Not Become Lower Prices

The textbook mechanism connecting efficiency to lower prices depends on competitive pressure. In a competitive market, firms that become more efficient face pressure to pass savings to consumers or lose market share. That mechanism is conditional, and the degree to which it operates depends heavily on market structure.

Research published by the Federal Reserve Bank of Boston finds that increases in concentration can materially alter cost-price pass-through dynamics: in more concentrated markets, firms can respond to cost changes in ways that preserve their margins rather than immediately adjusting prices 41. IMF analysis similarly links market power to pricing dynamics, noting that firms with greater pricing power may be better able to absorb favorable cost movements rather than passing them through to consumers 42. NBER work on wage-price relationships also examines how market power interacts with cost-signal propagation, though its primary focus is pass-through from labor costs rather than efficiency gains 43. These sources examine different aspects of pass-through and do not all prove the same directional mechanism — the broader structural point they collectively support is that when competitive pressure is weaker, the channel from lower costs to lower prices becomes less reliable.

U.S. corporate profit measures tracked by BEA and FRED show profits running near high levels in the post-pandemic period, even after the inflation surge began to ease 44, 45, 46. High profits can reflect demand, cyclical conditions, or product mix, and do not by themselves prove pricing abuse. But their persistence alongside high consumer prices is consistent with a structural story in which efficiency gains are retained rather than competed away 44, 46, 47, 48.

A specific version of this argument — “greedflation” — claims that firms were primarily responsible for the 2021–2023 inflation by expanding margins. In its strong form this is contested and should not be treated as settled. The more defensible version is narrower: firms with preexisting market power may have been better positioned to preserve or expand margins during a shock-heavy episode, amplifying price increases and potentially slowing disinflation as input costs fell 42, 49. That framing is supported by IMF and Fed research; stronger versions require more caution 50.

The market power argument applies directly to AI. If AI reduces the cost of delivering certain services, firms in concentrated markets that adopt AI are not compelled to pass those savings to consumers. They may retain gains as margin, return them to shareholders, or deploy them toward further competitive advantage. Whether AI makes services cheaper depends as much on market structure as on the technology itself.

Who Captures the Gains from Automation?

Productivity gains do not distribute themselves automatically. The preceding sections have traced several structural filters: cost pressures in essential services, concentrated markets that weaken price pass-through, AI infrastructure costs borne by capital rather than users, and a labor market in which task-compression falls unevenly. This section draws those threads together and asks directly who, in practice, captures the gains.

The labor share and within-labor inequality

For much of the postwar period, labor’s share of nonfarm business output was comparatively stable, around the low 60s percent. Since the early 2000s that share has moved lower, reaching near 58 percent by 2024 5. The labor share decline has multiple contributing causes: technology, global economic integration, and changes in labor market institutions all appear to matter 6, 7, 8.

The picture looks worse still when the focus shifts from average to median compensation. OECD research and Stansbury and Summers find that real median compensation has decoupled from labor productivity more sharply than average compensation has 19, 20 — implying a two-layered problem: labor as a whole receives a shrinking share of aggregate output, and within that shrinking share, gains are unevenly distributed. Research confirms the productivity-pay link is meaningful but incomplete, shaped by institutions, bargaining, and within-labor inequality 51, 20.

Concentration and superstar firms

Research by Autor and coauthors identifies a specific mechanism: industries with larger increases in concentration also tend to show larger declines in labor share 52, 53. Highly productive, high-markup firms dominating their markets tend to have lower labor shares than smaller competitors — as they grow their output share, they pull aggregate labor share downward without any deliberate decision to suppress wages. It is a consequence of which firms grow, not just how any particular firm behaves.

AI and the distributional channel

AI fits naturally into this framework. Acemoglu argues that AI is unlikely to reduce labor income inequality and more likely to extend capital-biased automation unless offset by new task creation or institutional intervention 18: AI primarily reduces labor inputs in existing tasks, and the resulting productivity gains flow to capital owners and high-skill complements rather than to labor broadly. McKinsey and Goldman Sachs are less pessimistic, but condition favorable outcomes on adoption speed, worker transition support, and institutional capacity to channel gains broadly — conditions that are not guaranteed 16, 17.

The capital intensity of AI infrastructure strengthens this argument. The AI buildout requires very large capital expenditure committed well before any individual user interaction 40. That capital is owned by corporations and their shareholders. AI services also depend on substantial ongoing infrastructure — electricity, data-center capacity, hardware — whose costs are borne primarily at the system level 36, 37. The gains from that cost structure flow primarily to the firms providing the infrastructure.

Even in sectors where AI might reduce some costs, the distributional benefit to households is not assured. In healthcare, Baumol-type pressures mean core labor costs are structurally resistant to automation 27, 24, 25, 26. When AI does reduce administrative costs, savings may accrue to healthcare institutions rather than flowing through to patients or workers 13, 23.

The default distributional pattern, under current institutional conditions, is that capital owners and shareholders of technology and platform firms receive the most direct and immediate benefits; high-skill workers who complement AI tools may benefit, though their gains depend on bargaining position; and median workers and households face continued cost-of-living pressure in the sectors that AI does not make cheaper. History shows that productivity gains have at times diffused broadly — but that diffusion has required deliberate institutional conditions, not just efficient technology.

Does AI Make People Richer, or Just Faster?

There is a difference between being fast and being secure. There is a difference between doing more in less time and having enough money left at the end of the month. This section asks whether AI, as it is currently operating in the economy, is producing the second of each pair.

What faster work does and does not deliver

AI tools are genuinely improving the speed and scope of certain tasks. Documents that once required hours can be assembled in minutes; code can be scaffolded and debugged with AI assistance; translation, summarization, and data analysis are faster and cheaper in digitally scalable applications. These gains are real.

But faster work is not automatically higher pay. The long-run U.S. record shows that labor productivity has risen much faster than real hourly compensation since the early 1970s 1, 2, and median compensation has decoupled from productivity considerably more sharply than averages suggest 19, 20. The worker who is doing more, faster, is not automatically earning proportionately more — and has not been for decades. AI may accelerate task completion, but if productivity gains continue to flow disproportionately to capital and to the upper end of the income distribution — as labor share, median compensation, and capital intensity data suggest is the structural tendency 5, 6, 7 — AI speed does not automatically convert into household prosperity.

The cost side: where households actually feel the economy

Whatever AI does on the output side, household financial wellbeing depends equally on what the world costs. Overall prices rose roughly 75 to 80 percent from 2000 through mid-2024, while typical earnings improved only modestly in real terms 9, 10. Shelter, medical care, and health insurance premiums have all risen far faster than wages 9, 11, 12, 13, 23. Published college tuition outpaced general inflation substantially 14, 15. Food and energy added cumulative pressure 9, 22, 28.

The structural reasons for these pressures are distinct from the parts of the economy AI is good at improving. Housing reflects land, regulation, and finance. Healthcare and education carry Baumol-type pressures in their labor-intensive cores — AI may ease administrative overhead, but not the fundamental labor intensity of clinical care, teaching, or interpersonal service work 27, 18, 24, 25, 26. The gap between what AI makes faster and what AI makes cheaper is precisely the gap between the digital economy and the household economy.

AI projections and the infrastructure cost

Major private-sector analyses project very substantial productivity and GDP gains from generative AI 16, 17. If those gains materialize and diffuse broadly, they could eventually improve household living standards. But those projections describe what could happen under favorable adoption conditions, not what has already occurred. More cautious macroeconomic analysis — notably Acemoglu (2024) — suggests that aggregate productivity effects may be considerably smaller once realistic diffusion assumptions are applied 18.

Even under optimistic scenarios, whether productivity gains become household prosperity depends on conditions that the technology does not determine: how gains are distributed between capital and labor 5, 6, whether competitive pressure passes cost savings through to lower prices, how AI infrastructure costs are allocated 36, 37, 40, and whether labor bargaining and regulatory structures support broad pass-through. The infrastructure behind AI’s apparent cheapness — electricity, cooling, capital expenditure running to hundreds of billions annually — is borne primarily by firms and capital markets, not by households 36, 37, 38, 39, 40. The capital invested in AI infrastructure accrues returns primarily to those who own it.

Faster is not richer. More efficient is not more secure. The current evidence does not show that AI has yet made them the same things for ordinary households.

Conclusion: Productivity Is Not Prosperity

This article began with a question that most people who use AI tools have not yet needed to ask: if automation is genuinely making work faster and some tasks cheaper, why doesn’t life feel correspondingly more affordable?

The answer is structural. The economy’s capacity to produce output is one thing. The conditions under which that output reaches ordinary households as wages, affordable prices, and economic security are another. Productivity and prosperity are related, but they are not the same thing, and the mechanisms that connect them are neither automatic nor neutral.

What the evidence shows

The historical record shows that productivity and compensation tracked reasonably closely in the postwar decades, then separated more clearly from the 1970s onward — a separation documented in official BLS data, confirmed in OECD cross-country research, and reinforced by the long-run decline in labor’s share of national output.

The cost side tells a parallel story. The categories of spending that dominate household budgets — shelter, healthcare, and education — have risen faster than overall inflation for most of the twenty-first century, while goods and services amenable to digital automation and scale have become cheaper. This creates a specific asymmetry: the things AI makes less expensive are mostly not the things that determine whether households are financially secure.

The structural reasons for this divergence are multiple and analytically distinct. Labor-intensive services face Baumol-type cost pressures. Housing faces land, regulatory, and financial constraints unrelated to productivity in any sector. Market concentration weakens the competitive pressure that would normally translate efficiency gains into lower prices. And the AI infrastructure that makes user-facing tools appear cheap is backed by enormous capital expenditure, electricity demand, and hardware costs borne primarily by the firms and capital markets that own the systems — not by the households that use them.

The central claim

Productivity is not prosperity.

This is not a claim that productivity is irrelevant. Economies that produce more per hour of work have a larger material base from which to raise wages, lower prices, and fund public goods. Productivity growth remains necessary for long-run improvements in living standards. The claim is about sufficiency, not relevance.

Productivity growth is necessary but not sufficient for broadly shared prosperity. Whether efficiency gains become rising wages, lower prices, shorter working hours, or greater economic security depends on a set of mediating conditions — the bargaining power of workers, the competitive structure of markets, the distribution of capital ownership, the regulatory design of critical sectors, and the degree to which gains are captured narrowly or passed through broadly. Those conditions are not fixed by the technology. They are shaped by institutions, by policy, by market structure, and by the accumulated choices of firms, governments, and labor organizations.

What makes this moment distinctive

Generative AI differs from previous automation waves in one important respect: it reaches into cognitive work that was previously insulated from automation pressure. That shifts the distributional question in ways not yet fully visible in aggregate data, but that the structural framework developed in this article suggests will follow established channels: gains accruing disproportionately to the owners of AI infrastructure and to high-skill workers with sufficient bargaining power, while median workers and ordinary households see some benefit in digital affordability but limited change in the costs that matter most to their financial security.

Whether AI eventually improves the material lives of ordinary households in a broad and durable way will depend not on the technical capability of the systems, but on whether the institutional conditions exist — or are created — to route efficiency gains more widely.

The right question to carry forward

The history of technological progress includes periods when efficiency gains did diffuse broadly. Real wages rose alongside productivity for several postwar decades. The gains from industrialization eventually reached large parts of the working class, though rarely without sustained labor organizing, political negotiation, and institutional change. The diffusion was not automatic, and it was not fast. But it happened.

Whether the current wave of AI-driven productivity will eventually produce similar broad diffusion — or primarily accelerate existing distributional trends — remains genuinely open. The more important question is not what AI can do, but what conditions would need to hold for what AI can do to become what ordinary people actually experience as a better life.

That question is not answered by faster models, larger datasets, or more impressive benchmarks. It is answered in labor markets, in housing policy, in the structure of health care financing, in the regulation of concentrated industries, and in the institutional arrangements that determine whether economic gains are shared or sequestered. Those are harder problems than building a better AI. They are also, in the end, the more consequential ones.

Productivity is not prosperity. The former is a technical achievement. The latter is a social one.

References

  1. U.S. Bureau of Labor Statistics. Labor Productivity and Costs. BLS. Source
  2. U.S. Bureau of Labor Statistics. “Labor Productivity and Real Hourly Compensation, Nonfarm Business Sector since 1973.” BLS Graphic, 2022. Source
  3. U.S. Bureau of Labor Statistics. “The Compensation-Productivity Gap.” The Economics Daily, 2011. Source
  4. U.S. Bureau of Labor Statistics. Table 2. “Nonfarm Business Sector: Labor Productivity, Hourly Compensation, Unit Labor Costs, and Prices.” Source
  5. Federal Reserve Bank of St. Louis. FRED: “Nonfarm Business Sector: Labor Share for All Workers” (PRS85006173). Source
  6. Jaumotte, Florence, and Carolina Osorio Buitron. “Understanding the Downward Trend in Labor Income Shares.” IMF Working Paper 2017/167. Source
  7. IMF Blog. “Drivers of Declining Labor Share of Income.” IMF, 2017. Source
  8. OECD. “OECD Employment Outlook 2012.” OECD, 2012. Source
  9. U.S. Bureau of Labor Statistics. Consumer Price Index Historical Data. BLS. Source
  10. U.S. Department of Labor. “Consumer Price Index – June 2024.” Source
  11. Harvard Joint Center for Housing Studies. “The State of the Nation’s Housing 2023.” Harvard, 2023. Source
  12. Harvard Joint Center for Housing Studies. “The State of the Nation’s Housing 2024.” Harvard, 2024. Source
  13. Kaiser Family Foundation. Employer Health Benefits Survey. KFF. Source
  14. College Board. “Trends in College Pricing and Student Aid.” College Board Research. Source
  15. College Board. “Trends in College Pricing and Student Aid 2024.” College Board, 2024. Source
  16. McKinsey Global Institute. “The Economic Potential of Generative AI: The Next Productivity Frontier.” McKinsey, 2023. Source
  17. Goldman Sachs Global Investment Research. “Generative AI Could Raise Global GDP by 7%.” Goldman Sachs, 2023. Source
  18. Acemoglu, Daron. “The Simple Macroeconomics of AI.” NBER Working Paper 32487, 2024. Source
  19. OECD Economics Department Working Paper No. 1373. “Decoupling of Wages from Productivity.” OECD, 2017. Source
  20. Stansbury, Anna, and Lawrence H. Summers. “Productivity and Pay: Is the Link Broken?” NBER Working Paper 24165, 2017. Source
  21. Herrendorf, Berthold, et al. “Revisiting Baumol’s Disease: Structural Change, Productivity Slowdown and Income Inequality.” Intereconomics, 2023. Source
  22. Federal Reserve Bank of St. Louis. FRED Economic Data: CPIENGSL and CPIMEDSL. Source
  23. Health Affairs. “Health Benefits in 2023: Premiums Increase With Inflation and Workforce Shortages Remain a Challenge for Employers.” Health Affairs, 2023. Source
  24. Hartwig, Jochen. “Drivers of Health-Care Expenditure: What Role Does Baumol’s Cost Disease Play?” Social Science Quarterly, 2018. Source
  25. Hartwig, Jochen. “Drivers of Health Care Expenditure: Does Baumol’s Cost Disease Account for Health Care Cost Growth?” SSRN. Source
  26. Bock, Alexander, et al. “Is Baumol’s Cost Disease Really a Disease? Healthcare Expenditure Revisited.” MPRA Paper 120873, 2024. Source
  27. Baumol, William J. “Macroeconomics of Unbalanced Growth: The Anatomy of Urban Crisis.” American Economic Review, 1967. Source
  28. U.S. Department of Labor. “Consumer Price Index – January 2024.” Source
  29. U.S. Bureau of Labor Statistics. “Measuring Price Change in the CPI: Household Energy.” Source
  30. Goldfarb, Avi, and Catherine Tucker. “Digital Economics.” NBER Working Paper 23684, 2017. Source
  31. OECD. “Measuring the Digital Transformation: A Roadmap for the Future.” OECD, 2019. Source
  32. Federal Reserve Bank of St. Louis. FRED: “Consumer Price Index for All Urban Consumers: Information Technology, Hardware and Services in U.S. City Average” (CUSR0000SEEE). Source
  33. U.S. Bureau of Labor Statistics. “Measuring Price Change in the CPI: Computers, Peripherals, and Smart Home Assistant Devices.” BLS. Source
  34. Grimm, Bruce T., Brent R. Moulton, and David B. Wasshausen. “Information Processing Equipment and Software in the National Accounts.” BEA, 2002. Source
  35. World Trade Organization, OECD, IMF, and United Nations. “Handbook on Measuring Digital Trade: Second edition.” IMF, 2023. Source
  36. International Energy Agency. “Electricity 2024” and “Energy and AI.” IEA, 2024. Source
  37. Lawrence Berkeley National Laboratory. “Report Evaluates Increase in Electricity Demand from Data Centers.” Berkeley Lab / DOE, 2024. Source
  38. Microsoft. “2024 Environmental Sustainability Report.” Microsoft, 2024. Source
  39. Microsoft. “Datacenter Sustainability.” Microsoft. Source
  40. CNBC. “Amazon Expects to Spend $100 Billion on Capital Expenditures in 2025.” CNBC, 2025. Source
  41. Federal Reserve Bank of Boston. “Cost-Price Relationships in a Concentrated Economy.” Current Policy Perspectives, 2022. Source
  42. IMF Blog. “Taming Market Power Could (also) Help Monetary Policy.” IMF, 2021. Source
  43. Mongey, Simon. “The Missing Inflation Puzzle: The Role of the Wage-Price Pass-Through.” NBER Working Paper 27663, 2020. Source
  44. U.S. Bureau of Economic Analysis. “Corporate Profits: BEA’s Suite of Measures Are Essential Indicators.” BEA, 2024. Source
  45. U.S. Bureau of Economic Analysis. Corporate Profits Data. BEA. Source
  46. Federal Reserve Bank of St. Louis. FRED: “Corporate Profits After Tax (without IVA and CCAdj)” (CP) and “Corporate profits: Profits before tax” (A053RC1A027NBEA). Source
  47. Miller, Nathan H., et al. “Rising Markups and the Role of Consumer Preferences.” NBER Working Paper 32739, 2024 / Journal of Political Economy, 2025. Source
  48. Amiti, Mary, et al. “Markups and Cost Pass-through Along the Supply Chain.” NBER Working Paper 34110, 2024. Source
  49. Federal Reserve Bank of San Francisco. “Are Markups Driving the Ups and Downs of Inflation?” FRBSF Economic Letter, 2024. Source
  50. Roosevelt Institute. “Why Market Power Matters for Inflation.” Roosevelt Institute, 2024. Supplementary framing only. Source
  51. Bureau of Economic Analysis and U.S. Bureau of Labor Statistics. “Integrated BEA GDP-BLS Productivity Account.” Source
  52. Autor, David, et al. “Concentrating on the Fall of the Labor Share.” NBER Working Paper 23108, 2017. Source
  53. Autor, David, et al. “The Fall of the Labor Share and the Rise of Superstar Firms.” NBER Working Paper 23396, 2017. Source
Post Views: 9