Real income per person, by class
Index, 2025 = 100. Wages + capital income + any AI dividend, after the scarcity premium on household basics.
Who owns the wealth
Share of household net worth. Starting point: Federal Reserve Distributional Financial Accounts, 2025–26.
Simulation: how sure is any of this?
Middle 40% — real income across simulated worlds
Top 1% — share of wealth across simulated worlds
Each bar is a share of simulated worlds. Red bars: the middle ends below its 2025 income.
Every layer is a toll booth. A model, a chatbot or a robot is the visible tip; beneath it sit lithography, foundries, memory, accelerators, power, orbit and command software. Control over any single layer is leverage; control over several is governance by other means. Figures are public estimates as of October 2026 and differ by source — treat the bars as orders of magnitude.
What this model is
A deliberately small, transparent system: four population slices, one economy-wide output, and a handful of levers. It exists to make an argument inspectable, not to predict a year or a number. Change an assumption you disagree with and watch what moves. On top of that sits a Monte Carlo simulation that shows how wide the range of outcomes is.
The four classes
| Class | Population | Wealth share (start) | Equity share | Labor-income share | Automation exposure |
|---|---|---|---|---|---|
| Top 1% | 1% | 31.6% | 50.0% | 4% | 0.00 |
| Next 9% | 9% | 35.4% | 37.2% | 26% | 0.75 |
| Middle 40% (50th–90th) | 40% | 30.5% | 11.7% | 48% | 1.00 |
| Bottom 50% | 50% | 2.5% | 1.1% | 22% | 0.70 |
Wealth and equity shares follow the Federal Reserve's Distributional Financial Accounts (top 1% ≈ 31.6%, bottom 50% ≈ 2.5%, top 10% hold ≈ 87% of corporate equities). Labor-income shares and exposures are assumptions: exposure encodes the idea that routine cognitive work in the middle is hit first, physical work in the bottom half later (humanoid robots), the upper tier partly, and owners not at all.
The mechanics
Automation a(t) = reach × normalised logistic, centred on the midpoint year Output Y(t) = 100 × 1.015^t × (1 + uplift × a(t)) Labor income of class i = Y × 0.58 × laborShare_i × (1 − a × exposure_i × (1 − reskilling)) Capital income K = Y − Σ labor income legacy part = 0.42 × Y, split by current wealth shares automation rent = K − legacy, split by AI-capital ownership AI-capital ownership = interpolation: broad (per head) → today's equity map → extreme (85% to top 1%) AI dividend: τ_eff = τ × (1 − capture × min(1, 2 × top-1% wealth share)) automation rent × τ_eff is pooled and paid equally per head Real income = income ÷ (1 + scarcity × a × basket_i) Wealth_i += saving_i × real income_i (saving rates calibrated so wealth/income is stable with no AI) Influence_i = (1 − capture) × population share + capture × wealth shareThe capture feedback is the loop the essay worries about: the more wealth the top 1% hold, the less of any redistribution a captured politics lets through, which lets them hold more.
The simulation behind the lines
The single lines are one world. To show how fragile or robust a conclusion is, the page runs a seeded Monte Carlo simulation: your scenario is re-run 800 times, each time with the inputs nudged by random draws, and the results are summarised as ranges (10th–90th percentile) and probabilities.
- What varies. The inputs below are drawn from normal distributions centred on your slider settings, truncated to sensible bounds. Annual growth gets correlated shocks, and the automation exposure of every class is scaled by one common random factor. The dividend is a policy choice, so it stays fixed. The Uncertainty slider multiplies every spread; at 0× all worlds are identical.
- Paired counterfactual. Each run is paired with a no-AI world that uses identical draws and shocks with automation switched off. "Worse off with AI than without" compares these pairs, so shared noise cancels instead of masquerading as an AI effect.
- Reproducible. A seeded generator (mulberry32): the same settings and seed give identical output. Re-roll the dice changes the seed. The seed and run time are shown under the verdict.
- What it cannot do. It measures uncertainty inside the model. The spreads are my judgement, not estimated from data, and if the model's structure is wrong, more runs will not fix it. Read the probabilities as sensitivity statements, not forecasts.
What it leaves out
- Population change, migration, trade, prices adjusting, inflation, interest rates, asset bubbles and crashes.
- New industries nobody can name yet. The Reskilling & new tasks slider is a crude stand-in for them. If you think history guarantees abundant new jobs, push it to 60% and see what remains.
- Politics that works: unions, antitrust, public compute, ownership funds, tax reform. Only the dividend lever is modelled.
- Anything physical about the supply chain. That is on the second tab, as evidence, not as dynamics.
What would show the pessimistic reading is wrong
- Labor share of output holding steady while AI adoption rises.
- Entry-level hiring in exposed occupations recovering rather than widening its gap with unexposed ones.
- Competition at the chokepoints (EUV alternatives, non-Nvidia accelerators, non-SpaceX launch) eroding the leader shares.
- AI-industry political money failing to change regulatory outcomes.
Sources behind the numbers
Secondary sources are cited where primary releases were not directly retrieved. Check the original filings and data series before relying on any figure. This page is commentary, not investment, legal or policy advice.