Plenty, Unevenly

Line every household up by what it consumes in year 20 and split the line into fifths. Without AI each rung moves a little. Under Extreme AI with the app’s default policy, even the poorest fifth ends up consuming 26× today’s median, while the richest fifth averages 510×.

Medians of 20 runs. Poorest-40% consumption under Extreme ranges 27–29× across seeds. Prices stay flat (0.1% a year), so these are real gains, not inflation.

Wealth concentrates quickly once adoption takes off. The wealth Gini starts at 0.57, reaches 0.71 by year 5 and 0.91 by year 20. The top 1% go from about 8% of household wealth to 75%. The bottom 40% hold 2.6%.

Adoption speed changes the timing, not the destination. With the adoption midpoint at year 15 instead of year 3, the Gini passes 0.8 in year 14 instead of year 7, and still ends at 0.93.

In the simulator, firm profits and equity are shared out by skill, and the tilt toward skill steepens as AI raises firm capacity. The “AI ownership concentration” slider sets how fast it steepens. Under Extreme AI, profits become almost all of household income, so the tilt decides nearly everything. The 1% most-skilled households end up holding 65% of all household wealth, against 7% with no AI.

Owning the AI agents themselves barely matters. Agent accounts hold a negligible share of wealth in these runs; the money flows through firm profits instead. Whether 10% or 95% of households own agents, and whether each owner has 2 or 50, the top 1% hold 74–75% of wealth.

One representative run (seed 19, the run whose median consumption is closest to the 20-run median) and three of its households. Because every setting with the same seed starts from the same people, we can follow the same families through different policies.

Under the best low-inflation mix (broad ownership tilt plus a 50% dividend, see 05), the rich owner consumes 2,078× instead of 7,096× and is still the wealthiest household in the economy. The median family rises to 63× and the struggling household to 47×. With the dividend switched off, the median family ends at 0.99× and the struggling household at 0.72×, lower than where it started (0.83×), while the rich owner still consumes 7,910×.

Median household across that seed’s population, year by year. In the no-dividend setup, the 20th-percentile household is at 1.00× in year 20; in the best mix it is at 57×.

Two dials matter most. The AI dividend is the share of the AI-produced slice of GDP paid out equally to every household (the app’s “UBI share”, default 25%). The ownership tilt sets how strongly profits skew toward high-skill households as AI grows (default 0.8). We crossed both under Extreme AI.

With no dividend, the bottom 40% consume about 1.1× today’s median after 20 years, against 0.6× in year 1, while output grows 118-fold. The default 25% dividend lifts them to 28×, and 50% to 47×. Prices stay calm up to a 50% dividend (3.3% a year at the default tilt). Between 50% and 75% there is a cliff:

The best mix for the bottom 40% at 5% inflation or less is ownership tilt 0.1, dividend 50%: the bottom 40% consume 57× today’s median (range 56–58×), the median household 74×, at 2.8% inflation. The top 1% hold 18% of wealth instead of 75%, and the median household’s well-being index is 4.01 vs 3.22. Pushing the dividend to 100% lifts the bottom 40% further, to 77×, but at 52% inflation a year.

The simulator’s income tax is flat, and any budget surplus is rebated in proportion to income. So raising the income tax doesn’t move money down the ladder. At a 35% rate, the top 1% end up with more of the wealth (84% vs 75%), and the bottom 40%’s consumption barely moves (26× vs 28×). More public spending helps a little: with the government buying 35% of output, the bottom 40% get 32×. Central-bank settings change little on their own.

Is there a trade-off between total wealth and fairness? Not in output. Across every policy mix under Extreme AI, real output ends at 106–137× start, and the mixes with large dividends sit at the top of that range. Only shrinking the state costs much output: 106× for the small-state budget, 115× with government buying 10%. Average real household wealth tells a similar story: 37× start under the best mix and 38× under the defaults. It is lower in the high-dividend mixes where prices rise fast (24× at a 75% dividend, 31% inflation) and under high flat taxes (15× at 50%), which are less fair, not more.

The capability ceiling caps how much bought compute can raise a firm’s capacity (1 = it can double a firm). It scales everyone’s outcome almost in proportion. The bottom 40% go from 14× with a ceiling of 0 (compute adds nothing) to 28× at 1 and 68× at 4. The wealth Gini hardly moves (0.93 to 0.90). The singularity mix (ceiling 4, faster capability growth, 12-month cost half-life) adds the dividend story on top: with a broad tilt and 75% dividend the bottom 40% reach 167×, but at 30% inflation.

Agent prices. Under Extreme AI the price of agents falls fast enough that the pace hardly matters. With a cost half-life anywhere from 6 to 240 months, the bottom 40% get 27.6–28.0×. Price matters only when agents start expensive (half a wage) and fall slowly with no energy breakthrough: then the bottom 40% get 15×. The month-48 energy step (a fusion-like cut in compute cost) restores the full result (28×).

Recessions. With large, frequent shocks (30% chance a year, size 0.2), the median household’s worst 12-month consumption drop grows by 3.0 points under Extreme AI against 9.2 points with no AI (paired by seed). With jobs no longer the main income source, shocks barely reach household consumption.