Two Scenarios, One Axis

Scenario one and scenario two are not opposites. They are the same argument, told from opposite ends of a single axis: how capable AI becomes. The more useful question sits on a second axis entirely: when AI does the work, who keeps the income it generates? The six steps below build that second axis from first principles, rather than asserting it.

Six Steps to a Testable Condition

The Result

ρ itself is not frictionless. Which instrument raises it, ownership-broadening versus distortionary taxation, determines how much of the nominal reattachment rate actually reaches consumption power. Call τ ∈ [0,1] the distortion coefficient of the chosen instrument (τ→0 for equity-broadening tools, τ>0 for punitive taxation, per the Policy-Growth Frontier below), and define the effective reattachment rate accordingly:

Substituting the effective reattachment rate into the consumption function gives the effective propensity to consume out of capital income:

The gap between the two propensities to consume then collapses to a single term:

Writing aggregate consumption as C(t) = Y(t)·Λ(s), and differentiating with respect to time, aggregate demand does not contract as automation proceeds, in a regime where displacement dominates reinstatement (s′(t) < 0), if and only if:

φ and α sit on both sides of this inequality: capability drives g_Y, the constructive growth force, through the productivity effect, and it drives |ds/dt|, the destructive force, through displacement, at the same time. In the regime the empirical literature currently describes, where g_Y is modest (Acemoglu's own estimate), capability's net effect on the sign of the inequality is small, and ρ is what actually moves it. A sufficiently explosive g_Y could in principle rescue aggregate solvency even at ρ_eff=0, but that only satisfies Step 5, not Step 6: an economy can stay macro-solvent while θ keeps the gains from ever reaching a typical household. Call the difference between a merely solvent economy and a genuinely broad-based one the Solvency-Abundance gap. It is closed by structural ownership-broadening, not by an aggregate growth number alone.

A pointed objection follows directly from that caveat: if capital owners save (1−c_K) of their AI-linked income rather than spending it, that saving does not vanish from the economy, it funds investment. Hundreds of billions of dollars in data-centre, energy, and silicon capex are exactly this channel in motion, and intermediate investment demand can hold aggregate output Y up even while the consumption share C/Y this model tracks is falling.

That is a real, near-term buffer, not a refutation. The Solvency Condition is a statement about C, and I is a genuine substitute for C in the short run. But it is not a permanent substitute: a data centre is built to serve future demand, not to be demand itself. If ρ never rises and the consumption channel never reattaches, the investment cushion faces the same overcapacity problem every historical capex supercycle has faced, capital gets built ahead of the consumption base meant to eventually justify it, and once the buildout phase matures, the correction shows up on the I side of the identity precisely because it never showed up on the C side. The investment channel does not change the Solvency Condition's verdict; it changes its timing, deferring the reckoning by however many years the capex cycle runs.

The model above treats ρ and g_Y as independent sliders, and the τ term introduced above is exactly where that independence breaks: τ and g_Y move together, driven by the same policy choice. If ρ is raised through highly distortionary instruments, punitive capital-gains or wealth taxation that falls on the same capital pool funding AI research and infrastructure, τ rises, ρ_eff falls short of the nominal ρ, and the expected after-tax return on that investment falls, so g_Y itself can drop as a direct consequence (Korinek & Stiglitz's own framework, already cited above, explicitly distinguishes redistribution instruments by how much they distort investment incentives). A policy that raises ρ by ten points while cutting g_Y in half has not obviously improved the Solvency Condition; it may have moved both sides of the inequality in the same direction. This is a specific instance of the equity-efficiency frontier Arthur Okun named the "leaky bucket" nearly half a century ago: every transfer from capital to broad consumption power leaks some growth on the way, and the size of the leak depends entirely on the instrument chosen.

This is also where the model earns its keep as more than a diagnostic. ρ is not one policy lever but a family of them, sitting at very different points on the distortion frontier. A sovereign wealth fund that buys equity in AI-linked capital, an employee stock ownership programme, or a broad-based AI dividend funded from a growing capital base each raise ρ by broadening who owns the asset, without directly taxing the return on it; the capital stock still earns what it would have earned, only more of the register of owners has changed. A wealth or capital-gains tax, or a "robot tax" levied on displacement itself, raises ρ by taking a larger share of the return after the investment decision is made, precisely the channel through which g_Y can be depressed. The practical mandate the model implies is not "raise ρ," it is: raise ρ through ownership-broadening instruments first, and treat direct taxation of the return as the higher-leakage, second-best tool.

A model this compressed earns scrutiny on where its own parameters stop being parameters and start being outcomes of the system itself. Three deserve to be named rather than left implicit.

The propensities to consume are not fixed. Treating c_L and c_K as constants is a simplification the Lucas Critique (Lucas, 1976; Nobel Prize in Economic Sciences, 1995) warns against directly: consumption is forward-looking, so a change in the automation regime itself changes the parameters describing behaviour under it. If workers perceive their own tasks as highly automatable, precautionary saving rises and c_L falls exactly when the model most needs it to hold, the same buffer-stock behaviour Carroll et al. (already cited) calibrate for wealth-constrained households generally. If capital owners instead perceive AI as a permanent, economy-wide upgrade to future income (Friedman's permanent-income hypothesis, 1957), c_K can rise. Both effects push the Distributional Wedge Δ the same way, toward widening, which means this comparative-statics snapshot is more likely conservative on the collapse side of the ledger than overstated.

The wage-led regime is an assumption, not a law. Bhaduri & Marglin's own 1990 result (already cited) is precisely that demand regimes are conditional: whether a falling wage share depresses or expands output depends on how sensitive investment is to the profit share versus how sensitive consumption is to the wage share. This model assumes the historically dominant wage-led case. Should AI-driven automation push profit shares high enough to trigger an investment boom large enough to outweigh the consumption drag, a profit-led regime becomes possible in principle, one more reason the Investment Cushion above is a genuine channel rather than a hand-wave. Which regime currently applies is, in principle, measurable from the empirical wage-led/profit-led literature this synthesis draws on but does not itself estimate.

The displacement pace is not exogenous either. |ds/dt| is treated as a slider, but Acemoglu & Restrepo's own task-based framework (already cited) makes reinstatement an endogenous response to relative factor prices: as automation cheapens capital relative to labour, it becomes more profitable, not less, for entrepreneurs to invent new tasks where labour retains an edge, the same mechanism behind the historical wage-share stability the model cites in Step 2. Nothing here assumes that channel is switched off; it assumes only that its speed, historically measured in decades, may not clear at the pace AI compresses the transition to. The boundary condition is a race between two endogenous processes, not a fixed force against a fixed wall, and this synthesis does not resolve which one wins. It states only the condition under which the race is lost.

Try the Model Yourself

Move the sliders to see how the reattachment rate changes the verdict, holding everything else fixed. The ρ slider here represents the effective reattachment rate, ρ_eff, net of whatever instrument-specific distortion τ is discussed above; it does not model τ separately. The presets reflect the illustrative positioning discussed below, not measured national estimates.

Where Three Economies Sit Today

Capability diffuses faster than ownership does, so all three face roughly the same φ today. What differs is each economy's realized ρ and its exposure to displacement.

The Solvency Condition above is, implicitly, a closed-economy statement: g_Y and θ are treated as accruing wherever the capital sits. India's IT-BPM sector breaks that assumption in a way the other two geographies do not. When a US client automates code review or Tier-1 support previously delivered from Bengaluru or Pune, the displacement, |ds/dt|, lands on the Indian labour market in full, but the productivity gain, g_Y, and the capital ownership, θ, accrue to the US firm that built and deployed the automation, not to India. Domestic ρ policy, however aggressive, cannot redistribute capital income that never crossed the border to begin with. For India's export-facing workforce specifically, the lever is not domestic redistribution at all; it is capturing a larger share of θ itself, through onshore compute capacity and equity stakes in the automating firms, not taxing a foreign firm's gains after the fact.

The ρ values loaded by the model's three preset buttons, 25% for the United States, 15% for India, 8% for the rest of the world, are illustrative positioning, not measured national statistics, and the model says so at every point a reader can touch it. But "illustrative" should not mean arbitrary. Each preset is anchored, directionally, to the sourced proxy in its geo-card above: the US figure sits where it does because deep capital markets and widely diffused retirement and brokerage-account infrastructure make a materially higher ρ buildable without new institutions, even though realized θ today is severe (Federal Reserve concentration data, above); India's sits lower because retail equity participation is shallower relative to the size of its automation-exposed IT-BPM workforce (Nasscom data, above), partly offset by a large informal and agricultural sector that is structurally slower to automate and so not yet fully exposed to the fork; the rest-of-world figure sits lowest because, for most economies, the capital income θ concentrates around is earned by frontier-compute-owning nations and never crosses the border to be redistributed domestically in the first place. None of this substitutes for the national-accounts and wealth-survey estimation the Research Agenda below calls for. It is the difference between choosing a number and being able to say, out loud, why that number and not another.

Why "Sub-Optimal Forever" Is the Less Likely Half

A financial bust in AI is a real possibility, deployment cost overshoots and corrects, per Shiller. But a cost-of-capital reset is not a capability reset. The dot-com crash wiped out capital, not the internet's underlying trajectory: telecom and tech capex fell by more than half from 2000 to 2002-03, even though the fibre laid during the boom stayed in the ground.

My Read: Neither Fork, a Footrace

That reattachment is not new. Agriculture fell from roughly 40% of the US workforce to under 2% over a century, and demand did not collapse; new sectors and new forms of income absorbed the shift, a real-world instance of ρ rising gradually as capital deepened and social-insurance systems matured. What is different this time is speed. Past transitions ran across generations, long enough for ρ to catch up with α without anyone naming the race. AI is compressing a multi-generational shift into single-digit years. The real risk is not that no equilibrium exists, the model has one whenever ρ→1. It is that we hit a disorderly transition, α outrunning ρ, before any new equilibrium forms.

So if AI becomes too smart, the outcome is not collapse, it is a test of whether ρ rises fast enough to keep Δ near zero. If AI stays sub-optimal, the outcome is not a jobs windfall either: a cost-of-capital bust reprices capital, it does not reverse capability for long, so a bubble bursting buys time to raise ρ, it does not buy safety on its own. The third path is the more boring, more likely one: an uneven, geography-by-geography race between α and ρ, where the deciding fight is over who owns the gains, not how smart the model gets.

What This Means for Decision-Makers

A framework earns the right to be more than an essay when it changes a decision. Two audiences this model is built to be useful to, beyond the general reader:

Evaluate an asset's exposure not only by its technical AI capability, φ in this model's terms, but by how dependent its revenue is on middle-class discretionary consumption in a geography with a low realized ρ. A portfolio concentrated in consumer-facing businesses in an economy where capital income is not reattaching to broad spending power carries a distributional risk the balance sheet does not show. The geography section above is a starting checklist for where that risk currently concentrates, and where a shift toward ownership-broadening instruments, rather than distortionary ones, would most re-rate it.

The entry-level cognitive and business-process work that has functioned as India's ladder into the middle class for two decades sits exactly where realized automation reaches first, per the geography section above. That ladder is also the demand side of the premium K-12 and higher-education market: the upper-middle-class disposable income that funds premium school and university fees is substantially the same income the IT-BPM ladder generates. If that ladder shrinks faster than ρ reattaches, institutions built on fee growth from that segment face a demand-side risk sitting alongside the more obvious curriculum one. Curricula and institutional positioning built around preparing students for that ladder's lower rungs are preparing them for the segment of the labour market most exposed to displacement outrunning reattachment. The more durable positioning is toward high-agency, complex problem-solving and relationship-driven work, where reinstatement (Step 2 above) has historically concentrated, and where human comparative advantage survives longest even in the disorderly-transition scenario this model treats as the more likely one.

What This Would Take to Become an Actual Contribution

A model is not a forecast, and this one is offered as a framework for discussion, not a finished result. Turning it into one would require: estimating ρ and θ from national accounts and wealth-survey data by country, rather than positioning them illustratively; proving the full dynamic system rather than a comparative-statics snapshot, including whether a stable equilibrium exists or the system can oscillate; deriving falsifiable, out-of-sample predictions and testing them against panel data as AI adoption proceeds; and subjecting it to peer review against a literature this synthesis draws on but does not exhaust.

Sources

1. Acemoglu, D. & Restrepo, P. (2018/2019). "Automation and New Tasks: How Technology Displaces and Reinstates Labor." Task-based framework; displacement, productivity, and reinstatement effects.
2. Acemoglu, D. (2024). "The Simple Macroeconomics of AI." NBER Working Paper 32487. Cumulative 10-year TFP effect estimate of ≤0.66%.
3. Bhaduri, A. & Marglin, S. (1990). Post-Kaleckian model of wage-led versus profit-led demand regimes.
4. Carroll, C., Slacalek, J., Tokuoka, K. & White, M.N. (2017). "The Distribution of Wealth and the Marginal Propensity to Consume." Quantitative Economics.
5. Korinek, A. & Stiglitz, J.E. (2017). "Artificial Intelligence and Its Implications for Income Distribution and Unemployment." NBER Working Paper 24174.
6. Shiller, R. (1981); Nobel Prize in Economic Sciences, 2013. Excess volatility of asset prices relative to fundamentals.
7. Piketty, T. (2013). Capital in the Twenty-First Century. The r > g finding.
8. India IT-BPM industry revenue and exports, FY25: Nasscom Strategic Review, 2025.
9. US household equity concentration: Federal Reserve Board, Distributional Financial Accounts, Q1 2026.
10. Median age estimates: UN World Population Prospects and CIA World Factbook, 2025.
11. Lucas, R.E. (1976). "Econometric Policy Evaluation: A Critique." Carnegie-Rochester Conference Series on Public Policy. Nobel Prize in Economic Sciences, 1995.
12. Friedman, M. (1957). A Theory of the Consumption Function. Permanent-income hypothesis.
13. Okun, A.M. (1975). Equality and Efficiency: The Big Tradeoff. The equity-efficiency "leaky bucket."