1.2.3 Capital vs. Labor Returns

Drive an hour west of most American cities and you will eventually find one: a low, windowless building the size of a shopping mall, ringed by fencing and cooling towers, humming with a sound like a distant highway. Inside are tens of thousands of graphics processors — the engines of the AI economy — drawing enough electricity to power a small town. The county that landed it was told the project represented a billion-dollar investment. That part was true. What the brochure was quieter about was the staffing: once construction crews leave, a facility that cost as much as a mid-sized hospital typically runs with a few dozen full-time technicians. The capital arrived by the billion. The jobs arrived by the dozen.

That ratio — enormous investment, minimal payroll — is the whole story of this chapter compressed into a single building. It is also the sharpest expression yet of a shift that has been building for four decades: the steady migration of income away from the people who work and toward the people who own. In the third quarter of 2025, the share of American national income going to labor fell to 53.8 percent, the lowest figure since the Bureau of Labor Statistics began tracking it in 1947 (BLS, 2025). When the government first measured it after the war, labor claimed roughly 70 percent. Nearly a fifth of the national pie has quietly slid off workers' plates and onto the balance sheets of capital. The question this chapter asks is simple and urgent: as AI arrives, is that slide about to accelerate — and if so, why, and what follows?

The Broken Promise

For a generation after the Second World War, pay and productivity rose together like two hands clasped. When workers produced more per hour, they were paid more per hour, almost in lockstep, from the late 1940s through the 1970s. This was not a law of economics. It was the visible result of a particular arrangement of institutions — strong unions, a rising minimum wage, tight labor markets, and a corporate culture that still imagined itself as answerable to workers and communities as well as shareholders.

Then, around 1979, the two hands let go. The Economic Policy Institute has tracked the divergence with almost painful clarity: from 1979 to 2025, productivity grew by roughly 90 percent while the pay of the typical worker grew by only about 33 percent (EPI, 2025). Productivity has risen close to three times as fast as compensation. The economy kept getting more efficient; the median paycheck barely moved. The wealth those extra hours of output generated did not evaporate. It accrued, in the EPI's phrasing, "somewhere in the economy besides the paychecks of typical workers, mostly in the pockets of extraordinarily highly paid managers and owners of capital."

Consider what this looks like from the inside. A worker who spent thirty years on a factory floor watched robots take over the heavy lifting, watched computerized systems double the plant's output, watched quarterly profits climb to records — and received, year after year, a raise that just barely kept pace with the price of groceries. The company kept its side of a bargain it had privately stopped believing in. Productivity was the worker's contribution. The gains were someone else's to keep.

The rupture of 1979 had no single author. Four forces pulled in the same direction and reinforced one another over decades.

The first was the collapse of collective bargaining. When roughly a third of private-sector American workers carried a union card, labor could credibly demand a share of the productivity it created. As union density fell toward single digits, that leverage drained away. The second was globalization: once a firm could source components or whole processes from lower-wage economies, the mere threat of offshoring capped what domestic workers could demand. The third was the rise of shareholder primacy — the doctrine, ascendant from the early 1980s, that a corporation's sole obligation is to maximize returns to its owners. Executives were increasingly paid in stock, which aligned their incentives not with their workforce but with their share price. The fourth was automation itself: as machines began to substitute for workers rather than merely assist them, the returns from each productivity gain flowed to whoever owned the machine.

AI did not invent any of this. It inherited a system already tilted steeply toward capital, its institutional guardrails already dismantled. What AI changes is the slope of the tilt.

What Makes AI Different

Every previous wave of automation had a cost that grew with its use. A robot on an assembly line is a physical object: it must be manufactured, shipped, installed, powered, and maintained, and to double your output you generally have to buy a second one. Physical capital scales, but it scales expensively, one unit at a time.

Software breaks that rule, and AI breaks it most dramatically of all. Training a frontier model costs an astonishing amount — hundreds of millions of dollars in computing time, plus the data centers described at the start of this chapter. But once the model exists, each additional use costs almost nothing. The same trained system can draft one legal memo or ten million, answer one customer or a continent's worth, at a marginal cost approaching zero. The economics are brutally lopsided: colossal upfront capital, negligible ongoing labor. That asymmetry is the technical heart of why AI concentrates returns in capital. When the marginal worker adds little and the marginal server adds nearly infinite scale, the value of ownership rises and the value of labor falls.

graph LR
  A[Huge upfront<br/>capital: models +<br/>data centers] --> B[Near-zero marginal<br/>cost per use]
  B --> C[Output scales without<br/>hiring more workers]
  C --> D[Value shifts from<br/>operating to owning]
  D --> E[Capital share rises,<br/>labor share falls]

Economists studying this directly have begun to put numbers on it. Research tracking AI deployment across European regions found that for every doubling of regional AI adoption, labor's share of income fell by between 0.5 and 1.6 percentage points (ScienceDirect, 2025). A separate body of work reaches the same destination by different roads: AI raises the capital intensity of production, lifting returns to capital while eroding the bargaining position of workers (Taylor & Francis, 2024). The more the technology is deployed, the more the income it generates flows to those who own it rather than those who labor beside it.

Crucially, AI aims at a different class of worker than the last automation wave did. Factory robots displaced hands. Generative AI displaces cognition — the accountant, the paralegal, the junior analyst, the customer-service representative, the entry-level coder. These were precisely the "safe," white-collar destinations that displaced manufacturing workers were told to retrain into. The ladder people were urged to climb is the one now being sawed off.

Where the Money Goes — and a Twist

For most of the past decade, the answer to "where do the profits go?" was blunt: to shareholders, through dividends and especially share buybacks, in which a company uses its cash to purchase its own stock and drive up the price. Over the ten years to the mid-2020s, S&P 500 companies routed the overwhelming majority of their profits into buybacks and dividends rather than into wages, training, or new productive capacity. In 2025 the machine was still running at full speed: buybacks alone were on track to hit a record $1 trillion for the year, with the first quarter setting an all-time quarterly record of $293 billion and total shareholder payouts reaching $457.6 billion in that quarter alone (S&P Dow Jones Indices, 2025). Dividends over the twelve months to September 2025 set their own record at $664.9 billion. A modest 1 percent federal excise tax on buybacks did nothing to slow the tide.

But here the AI era introduces a genuine and instructive twist — one that complicates the simple "profits go to buybacks" story and, on inspection, makes the capital-concentration argument stronger rather than weaker. At the AI hyperscalers specifically, the buyback era is ending. Combined buybacks by the largest tech firms fell to $12.6 billion in the fourth quarter of 2025, a 74 percent collapse from their 2021 peak (Yahoo Finance, 2026). Of the four biggest AI spenders — Alphabet, Microsoft, Meta, and Amazon — only Microsoft bought back any meaningful stock in early 2025. Why? Because the cash is being redirected into an even more capital-intensive channel. Those four companies plan to spend roughly $725 billion on capital expenditure in 2026, up about 77 percent from an already staggering $410 billion in 2025 — nearly all of it on GPUs, custom silicon, and data centers (ValueAdd, 2026). Alphabet has floated an $85 billion equity sale to help fund the build-out. Apple ($100 billion) and Nvidia ($80 billion) remain buyback machines, but the frontier labs have swapped one form of capital allocation for another.

This matters for our question. When a company pours a hundred billion dollars into data centers instead of returning it to shareholders, that is not a victory for labor. The money flows from one pocket of capital (existing shareholders) into another (the physical infrastructure of AI and the chipmakers who supply it — Nvidia's market capitalization has sailed past $3 trillion). Either way, the worker is a bystander. Buyback or capex, the productivity gains circulate within the capital economy. The reinvestment that reformers long called for is finally happening — but it is reinvestment in machines that need very few people to run, not in the people themselves.

The result at the aggregate level is a market whose earnings are extraordinarily concentrated. As of mid-2025, a small cluster of AI-focused giants accounted for more than a fifth of all S&P 500 profits; the "Magnificent Seven" alone made up roughly a third of the index by market value, with a net profit margin near 26 percent — nearly double the 13 percent margin of the other 493 companies (FactSet, 2025). A handful of firms now capture a share of national corporate profit that would have been unthinkable a generation ago.

Two Tracks, Not One

The story so far risks sounding uniform: capital up, labor down, everywhere. The reality has a crucial fault line running through it, and honesty requires drawing it clearly. Not all workers are losing. A narrow band of them are winning spectacularly.

PwC's Global AI Jobs Barometer, which analyzes hundreds of millions of job postings, has documented the split in successive editions. In its 2025 report, wages were rising more than twice as fast — 16.7 percent versus 7.9 percent — in the industries most exposed to AI compared with the least exposed. Workers who could credibly claim AI skills commanded a wage premium of 56 percent, up from 25 percent a year earlier. Productivity in the most AI-exposed sectors, such as software and financial services, had nearly quadrupled (PwC, 2025). By the 2026 edition, the premium for AI skills had climbed again to 62 percent, and PwC described a "two-track" labour market: "professionalised" roles, in which AI handles the routine and humans supply judgment, were seeing twice the job growth and 42 percent faster salary growth than "democratised" roles that AI had made easy enough for anyone to do (PwC, 2026).

Most AI-exposed Least AI-exposed
Wage growth (2025 barometer) 16.7% 7.9%
Productivity growth since 2018 34% 24%
AI-skills wage premium (2026) 62% —
Job-growth track "Professionalised": 2× growth "Democratised": slower growth

Source: PwC Global AI Jobs Barometer, 2025 and 2026 editions.

The catch is who fits in that first column. Model builders, senior engineers, and professionals whose expertise AI amplifies are riding a genuine boom. But they are a sliver of the workforce. The typical worker is not an AI engineer; the typical worker is in a role AI will either eliminate or commoditize. For them, the two-track economy means watching the express train pull away from the platform. The headline employment numbers can look calm — total nonfarm payrolls kept ticking up — even as the character of work diverges violently, with a small, highly paid AI-adjacent tier gaining and a broad middle stagnating or contracting.

How Sure Are We That It's AI?

This is the moment for intellectual honesty, because it is easy to blame AI for a trend that predates it. Labor's share began falling around 1979, four decades before ChatGPT. Unions were gutted, offshoring entrenched, and shareholder primacy enshrined long before a transformer model existed. A careful reader should ask: is AI actually causing the current decline in labor's share, or is it merely the latest passenger on a train that left the station in the Reagan era?

The honest answer is that AI is best understood as an accelerant, not an origin. The institutional conditions — weak labor, mobile capital, governance built around the share price — were already in place. What the regional-deployment studies suggest is that AI adds an incremental, measurable downward push on labor's share on top of those trends (ScienceDirect, 2025). Disentangling AI's specific contribution from the broader forces is genuinely hard, and anyone claiming precision here is overselling. What the evidence supports is a more modest and more defensible claim: AI is not reversing the four-decade slide, and there are strong mechanistic reasons — the near-zero marginal cost, the targeting of cognitive work — to expect it to deepen the slide rather than arrest it.

A second open question is whether the current concentration is durable or transitional. Optimists point out that general-purpose technologies historically diffuse: electricity and the computer eventually spread their gains widely, once the complementary skills, business models, and institutions caught up. On this view, today's Magnificent-Seven dominance is a transitional artifact of an early, expensive phase, and as models commoditize, the gains will broaden. Pessimists counter that AI's cost structure is unusually resistant to diffusion — the returns cluster around whoever owns the largest models, the most compute, and the most data, and those are natural monopolies that resist competition.

What evidence would tell them apart? Watch three signals over the coming years. First, the median real wage: if a general-purpose-technology story is right, median pay should eventually rise even as headline productivity climbs. Second, the concentration of AI profits: durable structural advantage would show up as the leading firms keeping their outsized margins rather than seeing them competed away. Third, the fate of the displaced cognitive workers: broad diffusion would create new, better-paid roles faster than AI eliminates old ones, while structural concentration would show up as a growing pool of workers whose skills have been commoditized. The optimistic and pessimistic scenarios make genuinely different predictions. We do not yet have enough post-2022 data to declare a winner — and saying so plainly is more useful than false confidence in either direction.

The Distributional Arithmetic

Suppose, though, that the pessimistic reading holds and labor's share keeps eroding at something like its recent pace. What does the country look like in ten or twenty years?

The mechanics are unforgiving. If the fruits of AI-driven productivity keep flowing to a shrinking set of owners and a thin tier of high-skill AI workers, inequality compounds along two axes at once. Wealth concentrates, because capital ownership is already far more unequal than income — the top tenth of American households own the overwhelming majority of stocks, so every dollar shifted from wages to shareholder returns is a dollar shifted up the distribution. And income polarizes within the workforce, as the two-track split hardens into two separate labor markets with little mobility between them. A plausible sketch of 2040, if nothing intervenes, is a society in which aggregate output is dramatically higher, a technology-owning class is extraordinarily rich, a professional AI tier is comfortable, and a large majority of workers occupy commoditized service roles with stagnant real pay — participating in the AI economy mainly as consumers renting access to the tools that displaced them.

That is not a forecast; it is an extrapolation of a trend line, and trend lines bend. But it is the direction the current arrangement points, and it clarifies the stakes. The second-order consequences of such a path are not merely economic. History and political science are fairly consistent on what happens when aggregate prosperity rises while median living standards stall: legitimacy frays. People who are told the economy is booming while their own paycheck stands still stop believing the institutions that keep telling them so. The result tends toward political volatility, the search for someone to blame, and a receptiveness to leaders who promise to overturn the table. A durable gap between a soaring GDP and a stagnant median wage is not just unfair; it is destabilizing, and the destabilization is the mechanism by which such gaps have, historically, eventually been closed.

What History Suggests

We have run a version of this experiment before. In the early Industrial Revolution, the gains from mechanization flowed overwhelmingly to factory owners while workers labored brutal hours for subsistence wages. Real wages for many British workers stagnated for roughly half a century — the period economic historians sometimes call "Engels' pause" — even as output and owners' fortunes exploded. The gap did eventually close, but not because the technology grew kinder. It closed because workers organized, built political power, and forced the issue across generations: factory acts, the ten-hour day, the legalization of unions, the extension of the vote, and eventually the whole apparatus of the welfare state. Redistribution, when it came, was won, not granted.

The parallel offers both warning and, faintly, hope. The warning is that the interval of maximum unfairness lasted decades and was reversed only by sustained struggle; nothing about a transformative technology guarantees broadly shared gains, and the people profiting from concentration have no structural incentive to end it. The hope is that the reversal happened at all — that a period of grotesque maldistribution was, in the end, politically survivable and politically corrigible. The lesson is not that AI will eventually lift all boats on its own. It is that whether the boats rise depends on choices — about taxation, ownership, and power — that are still open.

The Choices on the Table

Because the outcome is not technologically fixed, it is worth being concrete about the alternatives, and honest about what each would cost.

At the level of who owns the machines, several models aim to broaden the base of AI ownership. Worker cooperatives that hold shared stakes in AI tools would distribute returns across a firm's employees rather than an external shareholder class; their obstacle is scale, since cooperatives struggle to raise the vast capital that frontier AI demands. Public ownership of foundational AI infrastructure — models or data centers held on behalf of citizens — could route productivity gains to a broad tax base; its obstacle is the state's mixed record as a technology operator and the political difficulty of building such institutions from scratch. The idea with the most sudden momentum is universal basic capital, which would give ordinary citizens an equity stake in the AI economy rather than merely a wage from it. In 2025 and 2026 the concept jumped from think tanks into live politics: Senator Bernie Sanders introduced an American AI Sovereign Wealth Fund Act proposing a one-time 50 percent stock levy on the largest AI firms to seed a roughly $7 trillion public fund paying out annually to citizens (The Hill, 2026), while figures around the industry floated a gentler version in which leading labs voluntarily contribute 5 percent of their equity to an Alaska-Permanent-Fund-style vehicle, and the administration signed an executive order calling for a national sovereign wealth fund. The obstacles are real — critics warn a 50 percent levy could deter investment, and any such fund raises hard questions of governance and political capture — but the striking fact is the bipartisan interest, a signal that the distribution question has entered mainstream politics.

At the level of policy levers, the menu is more familiar but no less contested. Strengthened labor rights and easier unionization would rebuild the bargaining power whose loss began the slide, though they run against decades of institutional erosion. Higher capital-gains taxation would claw back some of the returns flowing to owners, at the risk — its critics argue — of dampening the investment that funds the data centers. Antitrust enforcement against AI's natural monopolies could keep the gains from ossifying around a few firms, though regulators are chronically outpaced by the technology. None of these is a silver bullet; each carries a genuine trade-off between equity and the incentives that drive innovation, and pretending otherwise is how policy debates go wrong.

At the level of corporate governance, the deepest lever may also be the least discussed. Share buybacks did not reach a trillion dollars a year by accident. They are the logical output of shareholder primacy, of executive pay denominated in stock, and of a legal understanding of fiduciary duty that treats the share price as the near-exclusive measure of a company's obligations. A meaningfully different distribution of AI's gains would require altering those norms at the root — broadening the definition of fiduciary duty to include workers and communities, restructuring executive compensation so it does not reward extraction, and loosening the grip of the doctrine that a firm exists only for its owners. These are not technical fixes; they are changes to what a corporation is understood to be for. That is precisely why they are hard, and precisely why they matter.

How Should We Judge It?

Underneath all of this sits a question the data cannot answer on its own: by what standard should we decide whether the current split between capital and labor is acceptable? Four criteria recur, and they pull in different directions.

By the criterion of efficiency, the current arrangement has a defense: concentrating returns on capital may be exactly what funds the enormous, risky investment that frontier AI requires, and redistribution that kills the investment helps no one. By the criterion of fairness, it fails badly: the workers whose data trained the models and whose jobs the models displace capture almost none of the value they helped create, while owners capture nearly all of it. By the criterion of democratic legitimacy, it is precarious: a distribution that a large majority experiences as rigged is unlikely to command consent for long, whatever its efficiency. And by the criterion of historical precedent, it looks like a familiar and dangerous phase — the early, maximally unequal stage of a transformative technology, before institutions caught up.

The book's own posture is not to pretend these criteria collapse into a single verdict. They do not. A reasonable person can weigh efficiency against fairness differently than her neighbor. But the evidence does support one firm conclusion, and it is the thread that runs through this entire chapter: the current distribution of AI's returns is not the neutral output of a machine. It is the product of decisions — about tax law, corporate governance, antitrust, and labor rights — that were made by people and can be made differently by people. The trillion dollars in buybacks, the data center with a dozen jobs, the record-low labor share: none of these fell from the sky. They were chosen. Which means a different distribution can be chosen too. Whether it will be is the open question of the next decade — and, if the Industrial Revolution is any guide, it will be answered not by the technology, but by whoever organizes to answer it.

Summary

The postwar bargain in which pay and productivity rose together broke in 1979, and labor's share of national income has been falling ever since — reaching 53.8 percent in the third quarter of 2025, the lowest on record since 1947. AI did not start this slide; it inherited an economy already stripped of the unions, tight labor markets, and stakeholder norms that once let workers claim a share of what they produced. What AI adds is a uniquely capital-concentrating cost structure: enormous upfront investment in models and data centers, near-zero marginal cost per use, and a direct assault on the cognitive jobs that were supposed to be safe. The gains show up as record shareholder payouts and, increasingly, as a $700-billion-plus surge of capital expenditure on AI infrastructure that employs very few people — either way, capital keeps the returns. A thin tier of AI-skilled workers is winning handsomely, commanding wage premiums above 60 percent, but they are a sliver of a workforce splitting into two disconnected tracks. How much of the trend is AI's specific doing remains genuinely uncertain, and whether the concentration proves durable or diffuses is the central open question. What is not uncertain is that the outcome is a matter of choice — of tax policy, ownership structures, antitrust, and corporate governance — and that history's one clear lesson is that a more equitable split, when it has come, was organized for and won, never simply granted.

Key Takeaways

  • Labor's share hit a record low. In Q3 2025 it fell to 53.8 percent of U.S. national income — the lowest since records began in 1947, down from roughly 70 percent then. Since 1979, productivity has grown about 90 percent while typical worker pay grew only about 33 percent (EPI, 2025).
  • AI is an accelerant, not the origin. The slide began in 1979, driven by union decline, offshoring, and shareholder primacy. AI adds an incremental, measurable downward push — for every doubling of regional AI deployment in Europe, labor's share fell 0.5–1.6 points (ScienceDirect, 2025).
  • AI's cost structure is uniquely capital-concentrating. Huge upfront investment plus near-zero marginal cost per use means output scales without hiring, shifting value from operating to owning. Unlike factory robots, AI targets cognitive jobs — the white-collar roles workers were told were safe.
  • The gains flow to capital either way. S&P 500 buybacks were on track for a record $1 trillion in 2025. At the AI hyperscalers, buybacks are collapsing (down 74% from their peak) — but only because cash is being redirected into roughly $725 billion of AI capex in 2026, infrastructure that employs very few people.
  • A two-track workforce. AI-skilled workers command wage premiums above 60 percent and "professionalised" roles are growing twice as fast as "democratised" ones (PwC, 2026) — but this narrow tier sits atop a broad middle facing stagnation or displacement.
  • The outcome is contested and unproven. Whether concentration is a durable structural feature or a transitional phase that will diffuse is genuinely open; the median wage, the persistence of AI profit margins, and the fate of displaced cognitive workers are the signals to watch.
  • Alternatives exist but face real obstacles. Worker cooperatives, public AI infrastructure, and universal basic capital — including live proposals like Sanders' AI sovereign wealth fund — could rebalance returns, but each carries genuine trade-offs between equity and investment incentives.
  • History's lesson is sobering. The Industrial Revolution concentrated gains for decades before organizing and reform redistributed them. Redistribution was won, not granted — and nothing about AI guarantees a kinder path.

Sources

Last updated: 2026-07-09

V2 (in progress) Previous: V1