Follow the Money: Where America’s Dollars Actually Go, Who Keeps Them, and What AI Is About to Do to the Split

A data-driven walk through the US income machine — with the math, the code, and an honest attempt to find the “greedy agents” soaking up value without returning it.

TL;DR — The Findings

• Labor’s slice of the pie just hit an all-time low. The share of US nonfarm business income going to workers fell to 53.8% in Q3 2025 — the lowest since records began in 1947, down from ~67% as recently as 1990. Meanwhile after-tax corporate profits reached 12.4% of GDP, near the highest ever recorded.
• The transfer is worth roughly $1.8 trillion a year. If labor’s economy-wide share had simply held its postwar norm, about $11,000 more per year would flow to each member of the labor force. A RAND study using tax data independently reaches the same order of magnitude.
• Flows have compounded into a wealth ratchet. The top 1% of households now hold ~31% of all US net worth (up from ~23% in 1989); the bottom 50% hold under 3%. Every group below the top 1% has lost share since 1989.
• Yes, “greedy agents” exist — but they’re structures, not villains. Formally, they’re rents: income without corresponding contribution. The audit finds them concentrated in ~$1T/year of health-care administration (~$265B of it pure waste), zoning-protected land rents, a finance sector whose unit cost hasn’t fallen in a century despite total automation, drug-patent monopolies, and licensing moats. Firm-level data shows average markups rose from 21% above cost (1980) to 61% (2016) — driven almost entirely by a few hundred firms at the top; the median firm’s pricing power never changed.
• Record buybacks (~$1.7T/year with dividends) are the smoke, not the fire. A buyback is a transfer, not a leak — but its scale is evidence that firms earn returns so far above competitive levels they have nothing better to do with the cash.
• AI is the stress test. AI capex (~$380B in 2025, ~$700B planned for 2026) drove the large majority of all US GDP growth in a year the economy added only 584,000 jobs — growth without hiring, visible in the national accounts. Whether AI crushes or rescues the labor share hinges on one parameter: the elasticity of substitution between capital and labor (σ). If σ > 1 — and early evidence leans that way — every dollar of AI capital deepening pushes labor’s share lower.
• The wildcard: AI is also the greatest potential rent-destroyer ever built. Administrative rents are made of exactly the cognitive paperwork LLMs automate at near-zero cost. Whether those savings reach households or get re-captured as margin depends on market structure and policy — not on the technology.

In the third quarter of 2025, the share of US nonfarm business income going to workers fell to 53.8% — the lowest reading since the Bureau of Labor Statistics started keeping score in 1947. In the same stretch, after-tax corporate profits hit 12.4% of GDP, the second-highest quarterly reading on record. S&P 500 companies returned roughly $1.7 trillion to shareholders over twelve months via buybacks and dividends. And a handful of technology firms committed close to $700 billion of 2026 capital spending to AI infrastructure — spending that, by some measures, accounted for the large majority of all US GDP growth.

Those four facts are not a coincidence. They are one system, observed from four angles. This article tries to map that system: where money flows in the US economy, how much labor gets, how that has changed over 75 years, and whether there are agents in the system extracting money without returning commensurate value — what economists call rents, and what you called “greedy agents.” Then we ask the uncomfortable question: what does AI, the biggest capital-deepening event since electrification, do to all of it?

Everything here is reproducible. The code pulls directly from FRED (the St. Louis Fed’s data service), and every claim is referenced at the bottom.

Part 1: The Accounting Identity Nobody Can Escape

Start with the one piece of math that makes everything else legible. Every dollar of output produced in the US is, by construction, a dollar of income paid to someone. GDP measured by production equals GDP measured by income:

Y = W·L + r·K + Π + T

where Y is total output, W·L is total labor compensation (wages, salaries, benefits), r·K is the return paid to capital (interest, depreciation allowances, rental income), Π is economic profit above the normal return to capital, and T is taxes on production net of subsidies.

The labor share is simply:

s_L = (W·L) / Y

For most of the 20th century, s_L was so stable that economists called it a "stylized fact" — Keynes described the constancy of labor's share as something of a miracle, and Nicholas Kaldor enshrined it as one of his famous growth facts. In the standard Cobb–Douglas production model:

Y = A · K^α · L^(1−α)

competitive markets pay each factor its marginal product, and the labor share equals exactly 1−α, a constant (~0.62–0.66 in postwar US data), regardless of how much capital accumulates. That's why the stability seemed natural: it fell out of the math.

The stability broke. Around 2000, and again sharply after 2010, the US labor share fell off its historical plateau and never came back. Whatever is happening, the Cobb–Douglas world — where technology and capital accumulation can’t change the split — no longer describes the US economy. Something else entered the equation. The rest of this article is about identifying that something.

The identity has a second life: the household ledger

There’s a parallel identity on the spending side. Every dollar a household spends is a dollar of revenue to some sector. So “where money flows” has two answers: the functional distribution (labor vs. capital, above) and the sectoral destination (which industries capture household spending). Both have moved dramatically:

Health care took about 5% of US GDP in 1960. It takes about 18% today — $5.3 trillion in 2024, roughly $15,500 per person. Housing costs have risen much faster than incomes in most metros, so a larger slice of the median paycheck flows to mortgage interest and rent. These sectoral shifts matter for the “greedy agents” question because a dollar that flows to a sector with weak competition and heavy intermediation behaves very differently from a dollar that flows to a competitive one.

Part 2: Pull the Data Yourself

Here is a complete, runnable script that reconstructs the core charts of this article. You’ll need a free FRED API key (fred.stlouisfed.org/docs/api/api_key.html).

# pip install fredapi pandas matplotlib numpy
from fredapi import Fred
import pandas as pd
import matplotlib.pyplot as plt

fred = Fred(api_key="YOUR_KEY_HERE")

SERIES = {
    "labor_share_nfb":   "PRS85006173",     # BLS labor share, nonfarm business (index)
    "labor_comp_gdi":    "W270RE1A156NBEA", # Compensation share of gross domestic income
    "corp_profits":      "CP",              # Corporate profits after tax (SAAR, $bn)
    "gdp":               "GDP",             # Nominal GDP (SAAR, $bn)
    "top1_wealth":       "WFRBST01134",     # Top 1% share of net worth
    "bottom50_wealth":   "WFRBSB50215",     # Bottom 50% share of net worth
    "median_wage_real":  "LES1252881600Q",  # Real median weekly earnings
    "productivity":      "OPHNFB",          # Output per hour, nonfarm business
}

df = pd.DataFrame({name: fred.get_series(code) for name, code in SERIES.items()})

# --- Chart 1: labor share vs. profit share -----------------------------
df["profit_share"] = 100 * df["corp_profits"] / df["gdp"]

fig, ax1 = plt.subplots(figsize=(11, 6))
ax1.plot(df.index, df["labor_comp_gdi"], color="tab:blue",
         label="Labor compensation share of GDI (%)")
ax1.set_ylabel("Labor share (%)", color="tab:blue")
ax2 = ax1.twinx()
ax2.plot(df.index, df["profit_share"], color="tab:red",
         label="After-tax corporate profits / GDP (%)")
ax2.set_ylabel("Profit share (%)", color="tab:red")
plt.title("The mirror image: labor share down, profit share up")
plt.tight_layout(); plt.savefig("shares.png", dpi=150)

# --- Chart 2: the productivity-pay wedge -------------------------------
base = "1979-01-01"
prod = 100 * df["productivity"] / df.loc[base:, "productivity"].iloc[0]
pay  = 100 * df["median_wage_real"] / df.loc[base:, "median_wage_real"].iloc[0]

plt.figure(figsize=(11, 6))
plt.plot(prod.loc[base:], label="Productivity (output/hour, indexed)")
plt.plot(pay.loc[base:],  label="Real median weekly earnings (indexed)")
plt.legend(); plt.title("Productivity vs. median pay, 1979 = 100")
plt.tight_layout(); plt.savefig("wedge.png", dpi=150)

# --- Chart 3: wealth concentration -------------------------------------
plt.figure(figsize=(11, 6))
plt.plot(df["top1_wealth"],     label="Top 1% share of net worth")
plt.plot(df["bottom50_wealth"], label="Bottom 50% share of net worth")
plt.legend(); plt.title("Household wealth shares (Fed Distributional Financial Accounts)")
plt.tight_layout(); plt.savefig("wealth.png", dpi=150)

When you run this, three pictures emerge.

The labor share plateau and cliff. Roughly flat (with cyclical wiggles) from 1947 to about 2000, then a step down after the dot-com recession, another after the Global Financial Crisis, and a fresh slide since 2023, culminating in the 2025 record low. Researchers at the New York Fed note a change in the dynamics too: before 2000, labor’s share reliably rebounded late in expansions; after 2000, it declines through expansions and stays down. The escalator broke.

Figure 1. Labor’s share of US nonfarm business income at published anchor points, 1947–2025. Interpolation between points is stylized; run the code above for the full quarterly series. Source: BLS via FRED (PRS85006173).

The wedge. From 1948 to the mid-1970s, productivity and median compensation grew together — the definition of broadly shared growth. Since 1979, productivity has grown several times faster than median pay. Output per hour keeps climbing; the median worker’s claim on that output does not.

The wealth ratchet. The Fed’s Distributional Financial Accounts show the top 1% of households holding roughly 31% of all US net worth in 2025, up from under 23% in 1989, while the bottom 50% of households — over 65 million families — hold roughly 2.5–3%. Strikingly, every group below the top 1% has lost share since 1989, including the 90th–99th percentiles. Flows compound into stocks: a persistent 8–14 percentage-point transfer of annual income from labor to capital, sustained for decades and reinvested at equity returns, produces exactly this wealth distribution. The flow chart and the wealth chart are the same chart at different time horizons.

Figure 2. Every group below the top 1% has lost share of US household net worth since 1989. Source: Federal Reserve Distributional Financial Accounts.

How big is the transfer in dollars?

A back-of-envelope worth doing explicitly. US GDP is roughly $30 trillion. The labor compensation share of gross domestic income has fallen by roughly 5–7 percentage points from its 1970s norm (the BLS nonfarm-business measure has fallen much more — from ~67% in 1990 to 53.8% in 2025 — but it covers a subset of the economy and embeds some measurement choices, so the economy-wide GDI measure is the conservative one). Take 6 points as a midpoint:

ΔFlow = 0.06 × $30T ≈ $1.8 trillion per year

That is the annual flow that would be going to paychecks if labor’s share had merely stayed at its postwar norm — roughly $11,000 per year per member of the US labor force. A RAND Corporation study approaching the question from tax data reached the same order of magnitude: the bottom 90% of earners captured 67% of taxable income in 1975 but under 47% by 2019, a shortfall on the order of tens of thousands of dollars per worker per year relative to the 1975 distribution. Two independent methods, one answer: the redistribution from labor to capital is measured in trillions annually.

Part 3: What Changed? The Suspect List

The decline of the labor share is one of the most-studied puzzles in modern economics, and no single cause survives scrutiny alone. The honest answer is a portfolio of causes, each with real evidence:

1. Cheap capital goods and automation. Karabarbounis and Neiman (2014) documented that the labor share fell in most countries worldwide, coinciding with a steep decline in the relative price of investment goods (computers especially). If capital gets cheaper and firms can substitute capital for labor easily — elasticity of substitution σ > 1, more on this in Part 5 — the capital share rises mechanically. The global breadth of the decline is this theory’s strongest card: US-specific villains can’t explain Germany.

2. Globalization and the China shock. Autor, Dorn, and Hanson showed that import competition from China after 2001 destroyed manufacturing employment in exposed regions with little offsetting job creation. Elsby, Hobijn, and Şahin found labor-share declines concentrated in import-exposed industries. Offshoring also changes bargaining: the credible threat to move production disciplines wage demands even where nothing moves.

3. The collapse of worker bargaining power. Union membership in the private sector fell from about a third of workers in the 1950s to about 6% today. The real federal minimum wage peaked in 1968. Non-compete agreements spread to tens of millions of workers, including fast-food employees. Each of these shifts moves the division of surplus within firms toward owners without changing anything about technology.

4. Superstar firms and reallocation. Autor, Dorn, Katz, Patterson, and Van Reenen showed that much of the aggregate labor-share decline comes not from every firm paying workers less, but from market share reallocating toward a small set of hyper-productive, low-labor-share “superstar” firms. Winner-take-most product markets concentrate revenue in firms that simply don’t need much labor per dollar of sales — a preview of the AI economy, as we’ll see.

5. Market power. This is where the “greedy agents” thesis gets its strongest formal support, and it deserves its own section.

Part 4: Hunting the “Greedy Agents” — A Formal Theory of Rent Extraction

First, precision. “Greedy” is a moral category; economics has a sharper one: economic rent — payment to a factor or agent in excess of what is needed to bring it into productive use. Rent is income without a corresponding contribution to output. A rentier isn’t necessarily greedier than anyone else; they’ve just found a position in the flow network where money passes through and some sticks, protected from the competition that would normally squeeze it away.

The test you’re asking for — “soaking up capital and money without a return for the system” — is precisely the rent test. Here is how to operationalize it.

The markup math

Define a firm’s markup μ as price over marginal cost:

μ = P / MC

Under perfect competition μ = 1: price equals the cost of producing one more unit, and there is no rent. Under market power μ > 1, and the gap is captured from consumers. There’s a beautiful duality (from the firm’s cost-minimization problem) that lets you estimate μ from accounting data without observing marginal cost directly:

μ = θ_v × (Revenue / Variable cost)

where θ_v is the output elasticity of the variable input. Intuition: if a competitive firm’s production technology says variable inputs should account for, say, 80% of revenue, but you observe them accounting for only 50%, the residual is markup.

De Loecker, Eeckhout, and Unger applied this to essentially every publicly traded US firm from 1955 to 2016 (QJE, 2020). The findings are the closest thing we have to a rent map of the economy:

• Aggregate markups were flat around 1.2 (21% above marginal cost) from 1955 to 1980, then climbed to about 1.6 (61% above marginal cost) by 2016.
• The rise came almost entirely from the upper tail: the median firm’s markup barely moved, while top-decile firms’ markups exploded. Rent extraction is not general — it’s concentrated in a few hundred firms.
• Average profit rates rose from about 1% of sales in 1980 to about 8% — an increase in excess of what rising overhead costs could explain.
• In their macro model, rising market power simultaneously accounts for the falling labor share, the rising capital share, falling low-skill wages, and declining labor-market dynamism. One parameter, four symptoms.

The measurement approach has serious critics (results are sensitive to what counts as a “variable” input, and Compustat covers only public firms), and you should hold the exact magnitudes loosely. But the direction is corroborated by independent methods: rising concentration in most industries, rising Tobin’s Q relative to measured capital, and the growing gap between corporate returns and safe interest rates.

A rent audit, sector by sector

If rents are concentrated, where? Here are the places where the evidence is strongest, with rough annual magnitudes. Treat these as overlapping estimates, not a sum.

Health care administration. The US spends roughly $1 trillion a year — around a fifth to a quarter of all health spending — on administration: claims processing, prior authorization, billing departments that exist to fight insurance departments. Comparative studies (Cutler and others) find US administrative costs far above every peer country as a share of GDP, and research syntheses suggest roughly $250–265 billion a year could be eliminated with no loss of care. That excess is close to a pure rent: real resources (millions of skilled workers, entire office towers) consumed by a zero-sum struggle over who pays, producing no health. It is also, importantly, someone’s income — which is why it persists.

Housing and land. Land is the original Ricardian rent: its supply is fixed, so payments to it don’t call forth more of it. Where zoning and permitting restrict construction, home prices and rents rise above construction cost, and the gap flows to incumbent owners. Hsieh and Moretti estimated that housing constraints in high-productivity cities lowered aggregate US growth substantially by preventing workers from moving to where they’d be most productive — a rent that doesn’t just transfer money but destroys output. Rising “housing wealth” in supply-constrained markets is largely capitalized rent, not created value.

Finance. Thomas Philippon’s landmark finding: the unit cost of financial intermediation in the US — what society pays per dollar of assets moved from savers to borrowers — has been roughly constant at about 2% for a century, despite computers, despite indexing, despite everything. An industry that grew from ~4% to ~8% of GDP without getting cheaper per unit of service is an industry where technology gains were captured by the intermediaries rather than passed through. Active management fees on funds that reliably fail to beat their benchmarks are the cleanest textbook rent in the modern economy: a measurable fee for a statistically nonexistent service.

Intellectual-property monopolies beyond the incentive optimum. Patents deliberately create temporary rents to reward invention. The rent audit question is whether the current scope exceeds what innovation requires — evergreening of drug patents, patent thickets, and the international extension of IP terms. US prescription drug prices run 2–3× peer countries for identical molecules; the gap over what would still fully fund R&D is rent.

Regulatory and licensing moats. Occupational licensing now covers roughly a quarter of US workers (up from ~5% in the 1950s), certificate-of-need laws restrict hospital competition, and dealer-franchise laws prohibit direct car sales. Each moat is small; collectively they are a broad tax paid by outsiders to insiders.

What about stock buybacks?

Buybacks are the most cited “greedy agent” in popular discourse, and the record numbers are real: over $1 trillion of S&P 500 repurchases in the twelve months through September 2025, and about $1.69 trillion including dividends. But be careful with the mechanism. A buyback is a transfer, not a consumption of resources — the cash goes to selling shareholders, who mostly reinvest it elsewhere. In a well-functioning capital market, returning cash from mature firms to be redeployed into young ones is exactly what should happen; the dot-com-era alternative of empire-building CEOs hoarding cash was worse.

The legitimate critiques are narrower and sharper: (1) buybacks timed around executive equity compensation create a genuine agency rent for insiders; (2) the sheer scale of payouts is evidence of the upstream problem — firms earning returns so far above their cost of capital, with so little competitive pressure to reinvest, that they have nothing better to do with $1.7 trillion. The buyback is the smoke; the markup is the fire.

The honest scorecard

Sum the defensible sector estimates — excess health-care administration, pricing power above competitive levels, land-supply rents, financial intermediation above efficient cost, licensing moats — and credible tallies land in the range of several hundred billion to a couple of trillion dollars per year, i.e., low-to-mid single-digit percentages of GDP. That is remarkably consistent with the ~$1.8T/year labor-share transfer computed in Part 2. It would be too strong to say the rent estimates explain the labor-share decline — automation and globalization are doing heavy lifting too — but the magnitudes rhyme, and De Loecker–Eeckhout’s model says market power alone can generate most of the observed symptoms.

So: yes, there are “greedy agents” in the formal sense — identifiable positions in the flow network extracting payment beyond contribution. They are not a cabal; they are structural features: an insurance-claims apparatus, a zoning board, a 2%-and-20 fee schedule, a patent thicket, a superstar platform with 60% margins. No one of them needs to feel greedy. The system routes money to them anyway.

Figure 3. Orders of magnitude. These estimates overlap and measure different things (a rent, a waste component, a transfer, a counterfactual flow) — the point is that they all live in the same trillion-dollar neighborhood.

Part 5: Now Add AI

Everything above was the pre-AI baseline. Here is what changed in the last three years, and why it stresses every mechanism in this article at once.

The capital flood

AI infrastructure spending is the largest concentrated private capital deployment in modern US history. The biggest hyperscalers spent roughly $380 billion on capex in 2025 and have guided to roughly $660–725 billion for 2026 — near-doubling in a year. Estimates of AI-related investment run from ~1.2–2% of GDP on narrow definitions (data centers, chips, networking) to ~5% on broad ones — the latter comparable to the peak of the late-1990s telecom buildout.

The growth-accounting is more startling than the levels. Jason Furman calculated that information-processing equipment and software — only about 4% of GDP — accounted for roughly 92% of US GDP growth in the first half of 2025. By Q1 2026, analyses of the BEA data attributed on the order of three-quarters of all US growth to the AI buildout, in a year when the economy added only 584,000 jobs (versus ~2 million in 2024) and unemployment drifted up to 4.4%. Growth without hiring, in one chart: the “K-shaped economy” is now visible in the national accounts themselves.

Figure 5. Left: hyperscaler capital expenditure (2024 actual, 2025 actual, 2026 guidance midpoint). Right: net US payroll growth. Sources: company guidance via Futurum/BloombergNEF; BLS.

And note the labor content of the spending: a hyperscale data center employs on the order of tens of construction workers per hundred once built. Dollar for dollar, this is the least labor-intensive investment boom the US has ever run.

The math of substitution: why σ is the most important parameter in economics right now

Go back to the production function, but generalize it. The CES (constant elasticity of substitution) form:

Y = [ α·K^ρ + (1−α)·(A_L·L)^ρ ]^(1/ρ),    σ = 1/(1−ρ)

σ measures how easily capital substitutes for labor. The labor share under competitive pricing becomes:

s_L = (1−α)·(A_L·L / Y)^ρ

Three regimes:

• σ < 1 (complements): capital accumulation raises the labor share. More machines make workers scarcer relative to output, bidding wages up faster than output. This was arguably the 19th–20th century norm.
• σ = 1 (Cobb–Douglas): shares are constant forever. The postwar illusion.
• σ > 1 (substitutes): capital accumulation lowers the labor share without bound. Every unit of cheap capital displaces labor at the margin, and labor’s share asymptotes toward zero as capital grows.

Karabarbounis–Neiman’s estimates already put σ modestly above 1 for recent decades. The economic definition of transformative AI is that it pushes σ dramatically higher — for cognitive tasks, potentially toward the limit where capital is a perfect substitute for labor (σ → ∞). In that limit the model degenerates to:

Y = A_K·K + A_L·L

and wages are pinned at the rental rate of the machine that can do the same task: w → A_L/A_K × r·(unit compute cost). Labor income becomes a race between human wages and the falling price of inference. Every subsequent efficiency gain in AI flows entirely to capital owners and consumers, none to wages.

You can simulate the transition in a few lines:

import numpy as np
import matplotlib.pyplot as plt

def labor_share(K, L=1.0, alpha=0.35, sigma=1.4, A_L=1.0):
    rho = 1 - 1/sigma
    Y = (alpha*K**rho + (1-alpha)*(A_L*L)**rho)**(1/rho)
    # competitive labor share = (1-alpha) * (A_L*L / Y)^rho
    return (1-alpha) * (A_L*L/Y)**rho

K = np.linspace(0.5, 20, 400)   # capital deepening, e.g. the AI buildout
for sigma in [0.7, 1.0, 1.4, 3.0]:
    plt.plot(K, labor_share(K, sigma=sigma), label=f"σ = {sigma}")
plt.xlabel("Capital stock (relative)"); plt.ylabel("Labor share")
plt.legend(); plt.title("Capital deepening and the labor share, by substitutability")
plt.tight_layout(); plt.savefig("ces.png", dpi=150)

Run it and you’ll see the whole debate in one figure: at σ = 0.7 the AI buildout would raise labor’s share; at σ = 3 it grinds labor’s share down monotonically as the capital stock grows. The empirical question of the decade is which curve we’re on — and the 2023–2025 divergence (record capex, record-low labor share, near-zero net hiring) is early, suggestive, not-yet-conclusive evidence for the upper curves.

Figure 4. Output of the simulation above. The entire debate about AI and wages compresses into which curve the economy is on.

AI meets the rent map

Now overlay Part 4. AI interacts with each rent channel, and mostly not in labor’s favor:

Superstar dynamics on steroids. AI production has extreme fixed costs (training runs, data centers) and near-zero marginal costs (inference), the classic recipe for winner-take-most markets and high sustained markups. The superstar-firm mechanism that drove reallocation-based labor-share decline is the native market structure of AI.

Bargaining power. Even where AI doesn’t replace a worker, the credible option to do so shifts surplus division — the same mechanism as offshoring threats in the 2000s, now applicable to cognitive work, and to white-collar occupations that were previously sheltered.

A new rentier position: the compute layer. Chips, power, and data-center land are the choke points of the new economy, and choke points are where rents accrete. One company’s near-monopoly on AI accelerators, utilities repricing electricity, and municipalities auctioning grid interconnection are early sightings of the AI economy’s rent map forming in real time.

The countervailing case. Fairness requires it: AI could also be the greatest rent-destroyer ever built. Recall the audit — health administration, legal process, financial intermediation, compliance: these rents are made of expensive human cognitive labor performing zero-sum paperwork. That is precisely what language models do at near-zero marginal cost. If competition (or public deployment) forces the savings through to prices, AI could vaporize hundreds of billions in administrative rents. The distributional question is whether the savings are passed through to households or simply re-captured as margin by whoever deploys the AI. History’s answer — Philippon’s flat 2% cost of finance across a century of automation — is not encouraging. Technology destroys the cost of rent-producing activity reliably; it destroys the rent only when market structure forces it to.

A scenario table, honestly labeled as speculation

• σ stays near 1, AI diffuses like electricity (decades): productivity accelerates, labor reallocates, labor share stabilizes at a lower plateau. The 2025 record low turns out cyclical-plus-trend, not a regime change.
• σ > 1 for cognitive work, current market structure persists: labor share continues a stair-step decline; income becomes increasingly a function of asset ownership; the top-1% wealth share (31% and rising) accelerates, since AI-era income accrues to exactly the equity that group holds.
• σ >> 1, plus policy response: the flow identity still holds — output becomes income for someone — but the “someone” is decided by ownership structure and taxation rather than wage bargaining. Proposals in play range from sovereign wealth funds holding compute equity, to Baumol-aware public financing of care sectors, to competition policy aimed at the compute layer. The math doesn’t say which; it only says the wage channel stops doing the distribution job by itself.

Part 6: What Would Actually Change the Flows

If the diagnosis is “structural rent positions plus rising capital-labor substitutability,” the remedies map one-to-one, and none of them are exotic:

Competition policy aimed where markups actually rose — the upper tail — rather than at bigness per se. Administrative standardization in health care (the identified ~$265B/year of pure waste has a literature and a to-do list; it lacks a constituency). Land-use reform, the rare policy with support from both Georgists and growth economists, because land rents are both huge and purely extractive. Portable benefits and bargaining institutions rebuilt for a world where the AI-substitution threat does the wage-suppression work that offshoring threats used to. Fee transparency and default indexing in finance. And — if the high-σ scenario materializes — broadening capital ownership itself, because in that world the labor share stops being the main distribution channel and the ownership share becomes the whole game.

The deepest lesson of the accounting identity is that money never disappears from the system — it always flows to someone. “Greedy agents” aren’t leaks; they’re routers configured by law, market structure, and bargaining power. The configuration was rewritten once, roughly between 1980 and 2000, and labor lost about six points of national income. AI is a once-in-a-century opportunity to rewrite the routing table again. The direction of the rewrite is not determined by the technology. It’s determined by who’s at the table when the config file is open.

Methodology Notes & Caveats

1. Which labor share? The headline 53.8% figure is the BLS nonfarm-business measure (FRED: PRS85006173), which excludes government, farms, and imputed housing, and allocates proprietors’ income by assumption. Economy-wide compensation share of GDI falls less dramatically (~5–7pp from postwar norms). Both directions of movement agree; magnitudes differ. Use GDI-based numbers for conservative claims.
2. Self-employment imputation, depreciation, and housing are the three big measurement fights in this literature (see Elsby–Hobijn–Şahin and Rognlie). Rognlie’s famous finding: much of the net capital-share rise is housing — which strengthens, not weakens, the land-rent story.
3. Markup estimates are sensitive to specification (variable-input choice, Compustat selection). Treat 21%→61% as a robust direction with debated magnitude.
4. AI GDP attributions (92% of H1–2025 growth; ~74% of Q1–2026 growth) are derived by analysts from NIPA sub-components, not official BEA “AI” lines; they require the judgment that these categories are currently AI-driven.
5. Rent estimates overlap. Health-admin waste, markups, and finance costs are not additive; a dollar can appear in two audits.

References

• BLS, Productivity and Costs — Labor Share, Nonfarm Business (FRED: PRS85006173); Q3 2025 record low of 53.8% widely reported (The American Prospect, Jan 2026; Fortune/Yahoo Finance, Jan 2026).
• Federal Reserve Bank of New York, Liberty Street Economics, “The Post-COVID Decline in the Labor Share” (June 2026).
• De Loecker, J., Eeckhout, J., & Unger, G. (2020). “The Rise of Market Power and the Macroeconomic Implications.” Quarterly Journal of Economics, 135(2), 561–644.
• Autor, D., Dorn, D., Katz, L., Patterson, C., & Van Reenen, J. (2020). “The Fall of the Labor Share and the Rise of Superstar Firms.” QJE, 135(2).
• Karabarbounis, L., & Neiman, B. (2014). “The Global Decline of the Labor Share.” QJE, 129(1).
• Elsby, M., Hobijn, B., & Şahin, A. (2013). “The Decline of the U.S. Labor Share.” Brookings Papers on Economic Activity.
• Rognlie, M. (2015). “Deciphering the Fall and Rise in the Net Capital Share.” Brookings Papers on Economic Activity.
• Philippon, T. (2015). “Has the US Finance Industry Become Less Efficient?” American Economic Review, 105(4).
• Hsieh, C.-T., & Moretti, E. (2019). “Housing Constraints and Spatial Misallocation.” AEJ: Macroeconomics.
• RAND Corporation (Price & Edwards, 2020). “Trends in Income From 1975 to 2018.”
• BEA, Corporate Profits (Q1 2026: $4.42T annualized; after-tax profits 12.4% of GDP — Yahoo Finance, June 2026).
• Federal Reserve, Distributional Financial Accounts (FRED: WFRBST01134, WFRBSB50215).
• S&P Dow Jones Indices, S&P 500 Buybacks press releases (Q1–Q3 2025): 12-month buybacks $1.02T; buybacks + dividends $1.685T through Sep 2025.
• CMS, National Health Expenditure Fact Sheet (2024: $5.3T, 18.0% of GDP); Brookings/Hamilton Project, “A Dozen Facts about the Economics of the US Health-Care System”; McKinsey administrative-simplification estimates (~$950B admin baseline, ~$265B/yr savings); Penn LDI / E. Emanuel on ~$1T (~22%) administrative spend.
• Epoch AI, “Data center buildout share of US GDP” (June 2026: AI data centers, hardware, networking ≈ 1.4% of GDP in Q1 2026, up from 0.7%).
• Furman, J. — analysis of info-processing equipment & software: ~4% of GDP, ~92% of H1–2025 GDP growth (widely reported, Oct 2025); St. Louis Fed, “Tracking AI’s Contribution to GDP Growth” (Jan 2026); BEA Q1 2026 third estimate analyses attributing ~74% of growth to AI-related categories.
• Hyperscaler capex: ~$380B (2025) to ~$660–725B planned (2026) — Futurum Group (Feb 2026); BloombergNEF (Mar 2026); KKR Insights (May 2026).
• J.P. Morgan Asset Management, “Is AI already driving U.S. growth?” (Sept 2025) — on data centers’ low employment intensity.

All FRED series are public domain. Code in this article is MIT-licensed; adapt freely.

Brian James Curry is a Kansas City–based data scientist and applied Machine Learning and AI researcher working at the intersection of economic modeling, causal inference, and machine learning — with a standing preference for models where the economics is load-bearing. He is the founder of Vector1 Research, an independent lab exploring marketing economics, cognitive AI architecture, and AI-driven enterprise transformation, with research contributions including Papilon, PyCausalSim, and Memory-Node Encapsulation. His twenty-plus years of enterprise work span Koch Industries, Tractor Supply, Vail Resorts, Hallmark, Garmin, and AT&T.