The Ownership Question: If Machines Do the Work, Who Owns the Machines?

Part 2 of Follow the Money. The labor share is collapsing and AI may finish the job — which means the most important economic variable of the next twenty years isn’t wages. It’s who holds the equity. With the math, the models, and the code.

TL;DR — The Findings

Wages are losing their role as the economy’s distribution mechanism. Part 1 documented labor’s share of nonfarm business income at a record-low 53.8% and showed that if AI makes capital a strong substitute for labor (σ > 1), capital deepening pushes that share down indefinitely. When the wage channel weakens, who owns the capital becomes the whole game.
Capital ownership is more concentrated than income ever was. The top 1% of US households hold roughly 50% of all corporate equities — more than the bottom 90% combined. The top 10% hold about 87–93%, an all-time high. The bottom half of the country holds a rounding error. If AI’s returns flow to equity, the current cap table of America routes them almost entirely to one decile.
The compounding math is brutal. A simple wealth-dynamics model shows why: when returns on wealth (r) exceed income growth (g), and the wealthy earn higher r than everyone else (they demonstrably do), concentration is not a possible outcome — it’s the equilibrium. AI plausibly widens the r-gap further. Our simulation, calibrated so its baseline reproduces the observed 1989–2025 drift, shows the top-1% wealth share rising from ~31% today toward ~48% within a generation under AI-concentration assumptions — Gilded Age territory.
The ownership question just went mainstream — from both flanks. In June 2026, Bernie Sanders announced the American A.I. Sovereign Wealth Fund Act: a one-time 50% tax on large AI companies, paid in stock, funding an estimated $7 trillion public fund paying every American a dividend over $1,000/year. Weeks later, the FT reported OpenAI had floated giving the US government a ~5% equity stake, and Sam Altman has publicly proposed a public wealth vehicle. Trump has mused about government stakes. When Bernie Sanders, Sam Altman, and Donald Trump circle the same policy, something structural is happening.
Working precedents exist and have for decades. Norway’s oil fund holds over $2.1 trillion — 1.5% of every listed company on Earth, roughly $390,000 per citizen. Alaska has paid every resident an annual dividend since 1982 ($1,000 in 2025) from a $88 billion oil fund. The template — convert a windfall into a diversified public endowment, distribute the returns — is not utopian. It is Republican-governed Alaska’s most popular institution.
The design space is wider than “nationalize it.” We model four mechanisms — sovereign wealth funds, baby bonds, employee ownership, and compute/token taxation — and simulate their 30-year effects on the wealth distribution. The differences are large: a seeded universal capital account compounds; a cash dividend doesn’t.
The honest counterargument gets a full hearing. Frontier AI firms are mostly unprofitable and valued on expectations; a 5% cash dividend obligation against paper valuations could force fire sales. Government equity stakes create conflicts (would a state that owns the AI industry regulate it?). Timing, valuation, and governance are where these schemes live or die — and where most of them have historically died.

Part 1: Why Ownership Is Suddenly the Whole Question

Every economy needs a mechanism that converts production into broad-based purchasing power. For two centuries, that mechanism has been the wage. Roughly 60% of national income arriving as paychecks meant that growth, however unequal, was automatically distributed — imperfectly, but at scale, without anyone designing it.

The previous article in this series documented the erosion of that mechanism: labor’s share of US nonfarm business income at 53.8%, the lowest since records began in 1947; after-tax corporate profits near record highs at 12.4% of GDP; and an AI capex boom (roughly $700 billion planned for 2026 by the hyperscalers alone) that drove the large majority of US GDP growth in a year the economy added just 584,000 jobs.

The CES model from Part 1 showed why AI threatens to make this permanent. If the elasticity of substitution between capital and labor exceeds one — if machines can genuinely replace workers rather than merely assist them — then every dollar of capital deepening lowers labor’s share, without floor. In the limiting case where AI is a near-perfect substitute for cognitive work, production becomes:

Y = A_K·K + A_L·L

and wages are pinned to the cost of the machine that can do the same task. Output keeps growing; the wage channel stops distributing it.

Here is the pivot this article makes: in that world, the accounting identity still holds. Output is still income — for whoever owns the capital. The question “how will people earn a living?” transforms into “who holds the equity?” And that question has an empirical answer today, which is the most alarming chart in this entire series.

Part 2: The Cap Table of America

The Federal Reserve’s Distributional Financial Accounts and Survey of Consumer Finances let us draw the ownership distribution of exactly the asset class AI returns will flow through: corporate equities and mutual fund shares.

The numbers, as of the most recent data:

The top 1% of households hold approximately 50% of all US equities — roughly $27–35 trillion — more than the bottom 90% of the country combined. The top 10% hold between 87% and 93% depending on measurement choices, the highest concentration on record. The top 0.1% — about 134,000 households — hold roughly a quarter of the market by themselves. Meanwhile, only 37% of American adults hold any stock outside a retirement account, and the bottom 50% of households hold low single digits of percent of the total.

Two things make this worse than it first appears.

First, equity concentration exceeds wealth concentration. The top 1% hold ~31% of total net worth but ~50% of equities, because middle-class wealth is houses and the wealthy’s wealth is businesses. AI returns will not flow through home equity. They will flow through exactly the asset the top decile has cornered.

Second, the concentration is rising, not stable. The top 1%’s share of public equities rose from 40% in 2002 to 54% by 2023. The 2020s bull market — increasingly an AI market, with a handful of AI-exposed mega-caps comprising roughly a third of the S&P 500 — compounded holdings at the top faster than anywhere else.

Put the two articles together and the mechanism is complete: AI shifts income from labor (broadly held) to capital (narrowly held), and the capital it shifts income to is the most concentrated major asset class in the economy. Absent intervention, the routing table is already written.

Figure 1. Household equity ownership by wealth group. The top 1% hold roughly as much stock as the bottom 99% combined; the bottom half of the country holds ~1.5%. Sources: Fed DFA/SCF-based estimates.

Part 3: The Compounding Machine — A Model You Can Run

Why does concentration keep rising? Because wealth dynamics are a compounding process with three asymmetries stacked on top of each other. Here’s the minimal model. Each household’s wealth evolves as:

W(t+1) = W(t)·(1 + r_i) + s_i·y_i(t)

wealth grows by its own return r_i plus savings out of income. Concentration rises when:

r > g (Piketty’s inequality): returns on existing wealth outpace growth of labor income, so inherited/accumulated wealth outruns earned wealth.
r is increasing in W: the rich earn higher returns than the poor — documented meticulously by Fagereng et al. in Norwegian registry data (the top wealth decile earns several points more per year than the bottom, via equity exposure, private business, and access to better vehicles). The poor hold deposits; the rich hold equity.
s is increasing in y: the rich save a far higher fraction of income (top-1% saving rates ~40%+; bottom-half saving rates near zero), so even equal returns would concentrate wealth.

AI enters the model in two places: it raises the aggregate capital share (more of Y flows through the r·W term and less through y), and it plausibly widens the return gap (the highest AI returns are in private markets — venture rounds, private credit — accessible only above wealth thresholds).

The simulation below implements this with realistic starting values and lets you toggle the AI scenario:

import numpy as np

N_YEARS = 30
# Stylized US: 4 groups with 2025 starting shares (Fed DFA)
groups        = ["Top 1%", "Next 9%", "Next 40%", "Bottom 50%"]
wealth_share0 = np.array([0.308, 0.364, 0.303, 0.025])   # Q4 2025 DFA
income_share  = np.array([0.20, 0.30, 0.38, 0.12])       # approx. pre-tax income
TOTAL_W, TOTAL_Y = 160e12, 30e12   # $160T household wealth, $30T GDP

def simulate(r, g, drift=0.0, sr=None, seed=None):
    """r: per-group return on wealth. g: labor income growth.
    drift: share of GDP shifting annually from labor- to capital-weighted
    income (the AI effect). sr: per-group saving rates. seed: annual $
    deposited into each group's accounts (public-fund payouts invested
    into seeded capital accounts rather than consumed)."""
    W, inc = wealth_share0 * TOTAL_W, income_share * TOTAL_Y
    out = [W / W.sum()]
    for t in range(N_YEARS):
        inc_t = inc + drift * TOTAL_Y * (wealth_share0 - income_share)
        W = W * (1 + r) + sr * inc_t
        if seed is not None:
            W = W + seed
        inc = inc * (1 + g)
        out.append(W / W.sum())
    return np.array(out)

# A: baseline — calibrated so 30-yr drift matches the observed 1989–2025 pace
base = simulate(np.array([0.058, 0.053, 0.048, 0.042]), g=0.035,
                sr=np.array([0.35, 0.22, 0.08, 0.02]))

# B: AI concentration — wider return gap + 0.4% of GDP/yr shifts to capital
ai   = simulate(np.array([0.078, 0.060, 0.046, 0.036]), g=0.025, drift=0.004,
                sr=np.array([0.35, 0.22, 0.08, 0.01]))

# C: AI + broad ownership — a $350B/yr public-fund payout (the Sanders-scale
# dividend) deposited into seeded accounts, 80% to the bottom half, which
# then earn the pooled-fund return instead of a deposit rate
seed = np.array([0.0, 0.0, 0.2, 0.8]) * 350e9
ubc  = simulate(np.array([0.074, 0.060, 0.054, 0.052]), g=0.025, drift=0.004,
                sr=np.array([0.35, 0.22, 0.10, 0.05]), seed=seed)

for name, s in [("Baseline", base), ("AI concentration", ai),
                ("AI + broad ownership", ubc)]:
    print(f"{name}: top-1% {s[0,0]:.1%} -> {s[-1,0]:.1%}, "
          f"bottom-50% {s[0,3]:.1%} -> {s[-1,3]:.1%}")

Representative output (your exact numbers will vary with parameters — that’s the point; change them and argue):

Baseline:             top-1% 30.8% → 39.0%,  bottom-50% 2.5% → 1.6%
AI concentration:     top-1% 30.8% → 48.3%,  bottom-50% 2.5% → 0.9%
AI + broad ownership: top-1% 30.8% → 42.4%,  bottom-50% 2.5% → 3.3%

Three lessons fall out of the model, and they’re robust to reasonable parameter choices. Even the baseline concentrates — the return-and-savings gradients alone push the top-1% share steadily upward, which is exactly what the 1989–2025 data showed. The AI scenario doesn’t create a new mechanism; it turbocharges an existing one, plausibly taking the top 1% toward half of all US wealth within a generation. And the only lever that changes the bottom of the distribution is changing what the bottom owns — because with near-zero savings rates, the bottom 50% cannot compound their way up through income. They can only be dealt into the capital structure directly.

Figure 2. Output of the simulation above. Left: the top-1% share under three scenarios, with the Gilded Age peak marked. Right: the bottom-50% share — note that only the broad-ownership scenario bends it upward. The baseline is calibrated to reproduce the observed 1989–2025 concentration pace.

That last finding is the entire case for what economists are starting to call predistribution: not taxing returns after they concentrate, but changing who holds the assets before the returns arrive.

Part 4: The Ownership Question Goes to Washington

For decades this was a seminar-room debate. In the space of one news cycle in mid-2026, it became live politics — with proposals from ideological poles that are startlingly convergent in structure.

The Sanders bill. On June 1, 2026, Senator Bernie Sanders announced the American A.I. Sovereign Wealth Fund Act. The mechanism: a one-time 50% tax on the largest US AI companies, paid in stock rather than cash, transferring roughly half the equity of firms like OpenAI, Anthropic, and xAI into a public fund. At current valuations the fund would hold an estimated $7 trillion. It would be required to distribute a 5% annual dividend — projected at over $1,000 per American per year — with returns also flowing to healthcare, education, and housing. Companies later crossing the revenue threshold would face the same equity transfer, and conglomerates would have to separate AI from non-AI operations so the public stake attaches to the AI business. Sanders’ framing: AI is built on the collective knowledge of humanity, so its wealth must return to humanity — and if valuations fall, the companies, not taxpayers, bear the loss.

The industry’s own version. Weeks later, the Financial Times reported that OpenAI had floated giving the US government an equity stake of roughly 5%, following an April 2026 OpenAI concept paper for a “public wealth fund” letting every citizen benefit from AI-driven growth — an idea Sam Altman reportedly pitched to Trump as early as 2025, explicitly modeled on the Alaska Permanent Fund’s oil-rent logic. Anthropic’s Dario Amodei has written that taxes on AI companies could fund broad income support. Elon Musk has proposed “universal high income.” Trump himself has publicly likened government stakes in AI developers to a partnership with the American people — and in 2025 signed an executive order directing officials to plan a US sovereign wealth fund.

Note what has happened here: the disagreement is now about magnitude and control — 5% versus 50%, voluntary versus taxed, industry-designed versus state-imposed — not about whether the public should own a piece of the machine. Sanders himself drew the line sharply, dismissing the voluntary versions as the wealthy “buying off the public.” But structurally, everyone is negotiating over the same instrument: public equity in the AI capital stock, paying a citizen dividend. That is a remarkable convergence, and it’s the reason this article exists now.

Part 5: The Precedents — This Machine Has Been Built Before

The strongest argument for public capital funds is not theoretical. Two have been running for decades, in plain sight, in politically opposite jurisdictions.

Norway. The Government Pension Fund Global, seeded with North Sea oil revenue starting in 1990, now exceeds $2.1 trillion — the largest sovereign wealth fund on Earth, holding roughly 1.5% of every listed company on the planet across ~7,200 firms. That’s about $390,000 per Norwegian citizen. It returned 15.1% ($248 billion) in 2025 alone. Norway pays no direct dividend; instead a fiscal rule lets the government spend ~3–4% of the fund annually, financing the welfare state permanently, long after the oil is gone. The design lesson: an arm’s-length professional manager, a hard fiscal rule, and index-like diversification away from the windfall asset.

Alaska. The Permanent Fund, created by constitutional amendment in 1976 under a Republican governor, converts oil royalties into a diversified $88 billion portfolio and has paid every resident — every man, woman, and child — an annual dividend since 1982. Payments have ranged from $331 (1984) to $3,284 (2022); the 2025 dividend was $1,000. It is routinely described as the world’s only running basic income, and it is politically untouchable across party lines. The design lesson: a universal, individual, visible payment creates a constituency that protects the institution for fifty years.

Figure 3. The running precedents against the proposal. Note the right panel’s log scale and the stock/flow distinction: Norway’s figure is accumulated fund value per citizen; Alaska’s and the Sanders projection are annual dividends.

The analytic bridge to AI is the concept of rent, which ties this series together. Oil funds work because oil income is substantially rent — value from a resource nobody produced, so taxing it heavily doesn’t destroy the incentive to produce. The Altman/Alaska analogy asserts AI income is similar: built on decades of public research, trained on the accumulated text of humanity, and (per Part 1 of this series) likely to earn superstar-firm markups well above competitive returns. To the extent AI profits are rents — returns above what’s needed to elicit the investment — the classic public-finance result applies: rents are the ideal tax base, because taxing them is non-distortionary. The entire fight is over what fraction of AI profits will turn out to be rent versus necessary return on genuinely risky capital. Nobody knows yet. Which is itself an argument for taking payment in equity — equity automatically scales with the answer.

Part 6: The Design Space — Four Mechanisms, Modeled

“Broaden capital ownership” is a family of very different machines. Here are the four serious contenders, with the simulation results from the Part 3 model applied to each.

Mechanism 1: Sovereign wealth fund with citizen dividend (the Alaska/Sanders model). Public fund holds equity; citizens receive cash annually. In the model this raises bottom-50% income meaningfully (a $1,000+ dividend is real money at the median) but barely moves bottom-50% wealth, because near-zero saving rates mean the dividend is consumed. That is not a flaw — consumption by the bottom half is the point of a dividend — but it means the wealth distribution keeps concentrating underneath. Scale check: $7T at 5% is $350B/year ≈ $1,030 per American. For comparison, Norway’s per-citizen fund value took 35 years of disciplined contribution to reach $390,000.

Mechanism 2: Universal basic capital / baby bonds (the compounding model). Instead of paying cash, seed every citizen an account holding equity — at birth or by lottery of time. Connecticut has run baby bonds since 2021 ($3,200 invested for every Medicaid-born child); Booker-style federal proposals scale to ~$50k by age 18 for poor kids. In the simulation this is the only mechanism that reverses the bottom-50% decline — lifting its wealth share from 2.5% to 3.3% over 30 years while every other scenario grinds it toward 1% — because it fixes the exact variable the poor cannot fix themselves: exposure to the compounding asset. The trade-off is patience — benefits arrive in decades, which is politically expensive.

Mechanism 3: Employee ownership (the ESOP model). Roughly 6,500 US ESOPs cover ~14 million participants today. Attractive because it needs no new federal institution — tax code tweaks suffice. But it has a fatal mismatch with the AI scenario: it ties workers’ capital income to their own employer’s equity, concentrating risk (ask Enron’s employees), and by definition covers only the employed — useless in precisely the high-displacement world it’s meant to insure against.

Mechanism 4: Compute/token taxation feeding a fund (the flow model). Rather than a one-time equity grab, tax the metered flow of AI usage — per-token, per-GPU-hour, or Steyer’s proposed California “token tax” — into a fund. Economically elegant (it scales automatically with AI diffusion, needs no valuation fights) but it’s a tax on the technology’s use, which is the one thing you don’t want to discourage, and it invites offshoring of inference.

The portfolio answer the model suggests: equity-funded endowment (take payment in stock, sidestepping the valuation problem), paying out through both channels — a modest cash dividend for legitimacy and immediate relief (Alaska’s lesson), and seeded individual capital accounts for the structural fix (the compounding lesson).

Part 7: Steelmanning the Other Side

The counterarguments are serious, and three of them are strong enough that any honest proposal has to answer them.

The valuation problem. Frontier AI firms are mostly unprofitable and valued on expectations — OpenAI is reportedly losing on the order of $14 billion a year even as revenue grows. A fund “worth $7 trillion” holds paper, not cash flow. A mandatory 5% cash dividend against paper valuations forces asset sales into whatever market exists at the time — potentially crystallizing losses at the worst moment. (Forbes’ critique of the Sanders bill lands hardest here.) The fix is Norwegian: dividend rules keyed to realized fund income with smoothing, not to headline valuation.

The governance problem. A state that owns half the AI industry faces an unresolvable conflict when regulating it — every safety restriction reduces its own fund’s value. Alaska and Norway solved this with radical arm’s-length design: professional managers, index-style diversification out of the source industry over time, and constitutional protection of the corpus from legislative raids. A US AI fund that stayed concentrated in AI equities would fail this test by construction; one that diversified into the global market (Norway’s actual practice) largely passes it.

The boundary problem. “AI company” is not a stable legal category. Every large firm is becoming an AI company the way every firm became an internet company; carving “AI operations” out of conglomerates (as the Sanders bill requires) invites a decade of definitional litigation. The clean alternatives tax things that are measurable — market capitalization above a threshold, compute consumption, or simply broad capital income — rather than trying to draw a line around a technology.

And beneath all three sits the deepest objection: perhaps the AI profit pool these schemes want to socialize won’t materialize on schedule. Part 3 of this series (the circular-financing article) will take that question head on. If the returns disappoint, an equity-funded scheme automatically shrinks — which, note, is an argument for equity funding and against fixed cash promises.

Part 8: The Stakes, Stated Plainly

Run the numbers one last time. If the AI-concentration scenario plays out — capital share rising, return gaps widening — the model puts the top 1% near 48% of US wealth within a generation, with the bottom half falling below 1%. No modern democracy has operated at that level of concentration; the closest historical analogues are the Gilded Age (top 1% ≈ 45% circa 1913) and pre-revolutionary agrarian societies. That is not a prediction. It is a default — the path requiring no new decisions.

The alternative path requires exactly one structural decision, made early, while the equity is still being issued: that some fraction of the machine’s cap table belongs to everyone. The precedents say it’s buildable (Norway), popular (Alaska), and cross-ideological (Sanders–Altman–Trump, however differently they’d build it). The models say the design details — equity versus cash funding, dividends versus seeded accounts, concentrated versus diversified holdings — change the outcome by more than the headline size does.

Part 1 of this series ended by observing that money flows are a routing table, rewritten roughly once a generation. The last rewrite, 1980–2000, moved six points of national income from labor to capital while almost nobody was watching. The next rewrite is happening now, in public, with bills on the floor and term sheets in the FT. The ownership question is open. It will not stay open long: cap tables, once written, compound.

Methodology Notes & Caveats

Equity ownership figures blend Fed DFA and SCF sources; “top 10% hold 87–93%” spans measurement choices (direct vs. including all indirect/pension holdings, household vs. total market cap). The top-1% ≈ 50% figure refers to household-held corporate equities and mutual fund shares.
The wealth simulation is stylized: four groups, fixed group membership (no mobility between groups), no taxes, no bequest splitting, deterministic returns. It’s a mechanism demonstration, not a forecast. All parameters are in the code — the invitation is to disagree with them quantitatively.
Return gradients (rich earn higher r) are empirically grounded in Fagereng, Guiso, Malacrino & Pistaferri’s Norwegian registry work; US gradients are likely similar or steeper given private-market access rules.
Sanders bill figures ($7T fund, $1,000+ dividend) are the sponsor’s estimates at announcement (June 2026); no CBO score existed at time of writing. The OpenAI ~5% stake report (FT, July 2, 2026) was unconfirmed by the company at time of writing.
Norway/Alaska figures: GPFG ~$2.05–2.1T (May 2026), ~$390k/citizen, 15.1% return in 2025; Alaska Permanent Fund ~$88B, 2025 PFD $1,000, dividend range $331–$3,284 since 1982.

References

Federal Reserve, Distributional Financial Accounts & Survey of Consumer Finances — equity ownership by wealth percentile (top 1% ≈ 50% of equities; top 10% ≈ 87–93%; top 0.1% ≈ 24%); Fed SHED 2025 (37% of adults hold stocks outside retirement accounts).
Inequality.org / IPS analysis of Fed data — top 1% share of public equities 40% (2002) → 54% (2023).
Piketty, T. (2014). Capital in the Twenty-First Century — the r > g framework.
Fagereng, A., Guiso, L., Malacrino, D., & Pistaferri, L. (2020). “Heterogeneity and Persistence in Returns to Wealth.” Econometrica, 88(1).
Saez, E., & Zucman, G. (2016). “Wealth Inequality in the United States since 1913.” QJE, 131(2) — historical top-share benchmarks.
American A.I. Sovereign Wealth Fund Act — Sanders announcement, June 1, 2026 (AP; Yahoo Finance, June 18, 2026; Newsweek, June 2, 2026; The Commons explainer, June 17, 2026): one-time 50% stock-paid tax, ~$7T fund, 5% dividend, >$1,000/person.
Financial Times (July 2, 2026), via Bloomberg/CNBC/Forbes — reported OpenAI proposal of ~5% government equity stake; OpenAI “public wealth fund” concept (April 2026); Altman–Trump discussions reported from early 2025.
Broughel, J., Forbes (June 22, 2026) — critique of the Sanders bill (valuation, dividend, and AI/non-AI separation problems).
Windfall Trust Policy Atlas — Universal Basic Capital and Sovereign Wealth Funds entries; Berggruen & Gardels on UBC; Connecticut Baby Bonds (2021, $3,200/child); Casey, Rockall & Roy “AI Bond”; Steyer token-tax proposal.
Norges Bank Investment Management / Wikipedia (May 2026) — GPFG AUM $2.05T+, 1.5% of world’s listed equities, ~$390k per citizen; CNBC (Feb 27, 2026) — 15.1% return, $248B profit in 2025.
Alaska Permanent Fund Corporation / Alaska Dept. of Revenue — fund ~$88B; 2025 PFD $1,000; dividend history 1982–present ($331 min, $3,284 max).
Visual Capitalist / Global SWF (May 2026) — world SWF league table (~$13T total).
Prior article in this series: “Follow the Money” — labor share, markups, rents, and the CES/σ framework this article builds on.

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.