UQO Working Paper No. 24· v1.1· November 2026· 90 pages
From Labor to Capital: Artificial General Intelligence, General Equilibrium, and Asset Valuation
Under AGI the labor share falls to 0.283 by 2050 and toward zero, but the income labor loses is worth little as compute capital: the market value of compute peaks at 0.30 of GDP, and land ends up capturing 0.605 of income by 2100.
The model keeps Restrepo's task economy and makes compute a priced, depreciating capital good that needs energy and works with structures, land and energy sites. The labor-share bound carries over given unbounded compute; the wage-level result does not when the complements are gross complements. Every path is a perfect-foresight conditional scenario, not a forecast, and the two catalogue cells with an elasticity of substitution of one or more were attempted and not solved.
Simon-Pierre Boucher — contact@spboucher.ai
0.56 → 0.283
Labor share in the baseline scenario, 2026 to 2050; below one in a thousand after 2075
40.5 %
Peak real return on wealth, in 2060, along a perfect-foresight path with no risk premia
4.4 → 3.0 → 32.7
Wealth-to-income ratio: 2026, trough in 2042 while the capital share rises, and 2100 once land captures 0.605 of income
0.30 of GDP
Peak market value of compute, in 2045 (0.092 of wealth in 2044); across the solved catalogue the peak reaches 0.67 of GDP
0.79 → 0.64 → 0.78
Wealth Gini in the heterogeneous-agent economy, 2026, 2050 and 2100 (0.83 in 2150): non-monotonic, and a citizen fund holding one fifth of all claims does not lower it after 2040
3.51×
Peak of the higher skill's wage relative to 2026, while its expert task is the last one compute cannot do cheaply; the wage then collapses
24 of 26
Scenarios of the catalogue solved to tolerance; the two cells with an elasticity of substitution of 1 and 1.5 were attempted and are reported as not solved
Abstract
Restrepo (2025) shows that once compute can perform every task, wages converge to the value of the compute a worker saves and the labor share goes to zero. We add the other side of the national accounts: compute is a priced, depreciating, energy-using capital good that works with structures, land and energy sites, and households own the claims to income. His labor-share bound carries over, given unbounded compute; his wage-level result does not. When the physical complements are gross complements and compute becomes cheap, wages fall in level and the income that labor loses goes to the owners of the scarcest complement; with a floor on the compute price, the wage is the user cost of the replacement compute. In a calibrated economy solved along perfect-foresight paths under conditional scenarios, the labor share falls to 0.283 by 2050, the real return peaks at 40.5%, the wealth-to-income ratio falls to 3.0 before land raises it to 32.7 by 2100, the market value of compute peaks at 0.30 of GDP, and wealth inequality across households is non-monotonic. In the model, a compute tax raises the limit wage with the tax-inclusive user cost, a land tax rebated to the claim owners changes no allocation, and a citizen fund bounds workers' expenditure share from below.
JEL: E22, E24, E25, G12, J23, J24, O33, O41
Key Results
The labor-share bound carries over
Given unbounded compute, the labor share goes to zero under any constant-returns production structure, with or without complements and whether compute is an endowment or a priced capital good, because the bound needs only that labor and compute be perfect substitutes within tasks. The paper presents this as Restrepo's own bound restated.
The wage-level result does not
If the task composite and the physical complements are gross complements and compute becomes cheap, the price of compute and wages go to zero in level and the lost income goes to the owners of the scarcest complement. With a floor on the compute price, wages converge to the user cost of the replacement compute, inclusive of its energy (0.40 of the total on the baseline path in 2200).
Income shares and asset values are different objects
On a balanced path compute, which depreciates at 25 % a year and cheapens at 15 % a year, is capitalized at a price–earnings ratio of 2.65 against 8.2 for structures and 50 for land, so capital's share can rise while wealth-to-income falls. The paper labels this an organising identity rather than a new result.
A calibrated transition to 2300
In the baseline the routine task is automated by 2029 and the expert task by 2057; output rises 3.7-fold by 2050 and 233.0-fold by 2100, after which growth returns to its trend because land binds.
Households and the distribution of wealth
A two-class economy prices every claim, and a heterogeneous-agent economy with two skills, three discount types and five productivity states, solved from one pre-news population, traces the wealth Gini, stable within 0.0014 across 7 discretisations.
Policy in general equilibrium
In the model a compute tax raises the limit wage with the tax-inclusive user cost at a cost in output, a land tax rebated to the claim owners changes no allocation (rebated to workers it moves output by 0.4 % in 2050), a citizen fund keeps workers' expenditure share bounded away from zero, and an AI monopoly protects wages like a compute tax with the rents going to the monopolist.
Quality control
This paper went through a multi-stage quality control: 7 quality-control reports and 112 points checked. The version presented on this page is the revised manuscript. Scores are on a 0–100 scale and are a diagnostic of the manuscript, not an acceptance probability.
Score after revision
70.2/100
60.1 before quality control · +10.1
Quality control dated November 7, 2026
- Quality-control decision
- major revision
- Scientific integrity
- PASS
- Reproducibility
- R1reproduced from documented analysis data
- Maturity
- Bone revision cycle from submission
- Quality-control reports
- 7
- Points checked
- 112
Scores by dimension
Before quality controlAfter revision
- Originality & contribution64.4 → 65.0
- Scientific rigor56.3 → 70.0
- Identification & methodology53.0 → 66.0
- Data & reproducibility67.1 → 80.0
- Robustness & validation56.9 → 70.0
- Writing69.3 → 70.0
- Figures & tables57.1 → 75.0
- Literature positioning65.3 → 74.0
Revised manuscript90 pages11 figures23 tables15 equations133 references39,016 words
What quality control changed
Quality control made the destination of labor's lost income conditional: the statement that it goes to the scarcest complement now holds for gross complements and a compute price with a positive floor, and the two scenario cells outside that region are reported as unsolved. Several headline numbers were corrected, and the heterogeneous-agent economy was re-solved from a single pre-news population, which produced a non-monotonic path of wealth inequality.
- "The market value of compute never exceeds one fifth of GDP" → it peaks at 0.30 of GDP in 2045 and at 0.092 of wealth in 2044; across the solved catalogue the peak reaches 0.67 of GDP
- Wealth-to-income ratio falling to 2.7 and reaching 31.0 by 2100 → falling from 4.4 to 3.0 in 2042 and reaching 32.7 by 2100, as land captures 0.605 of income
- "The labor-share result survives any constant-returns production structure" → the bound carries over given unbounded compute, and is presented as Restrepo's own bound restated
- 25 scenarios all reported as solved → 26 scenarios, 24 solved to tolerance; the cells with an elasticity of substitution of 1 and 1.5 were attempted and are not solved, so the dynamic results are stated for gross complements only
- Heterogeneous-agent economy re-solved from one pre-news population: the wealth Gini is 0.79 in 2026, 0.64 in 2050 and 0.78 in 2100, and a citizen fund holding one fifth of all claims does not lower it after 2040
- "A land tax redistributes without distortion" → a land tax rebated to the claim owners changes no allocation; rebated to workers it moves output by 0.4 % in 2050
Limitations stated in the paper
- The paths are perfect-foresight equilibria: there is no aggregate risk, no risk premium and no learning about the arrival of AGI, and nothing in the paper is a forecast.
- The level of wages after 2060 depends on two scenario inputs, the compute-price floor and the terminal compute per worker, and on the energy cost of compute, none of which historical data identify.
- The two catalogue cells with an elasticity of substitution of one or more, in which income would go to compute owners rather than to land, are not solved at the paper's tolerance.
- The heterogeneous-agent economy has no return heterogeneity and therefore no Pareto tail (top-1 % share 0.13 against 0.325 in the data), and the production structure has no endogenous innovation or new human tasks beyond one reinstatement scenario.
Data
Model economy in three layers
An exact replication of the source; a limit economy with one complement in which nine propositions are proved; and a quantitative economy with four capital stocks, energy and housing, solved along 275-year perfect-foresight paths to 2300.
Calibration to the United States in 2026
Public, key-less inputs retrieved on 2026-10-27: 92 tidy series and 162 calibration targets, of which 7 enter the calibration. The national accounts give the labor share (0.568 in 2023), residential structures (1.17 times GDP) and a value of compute capital of 6.2 % of GDP; the data-center share of electricity (4.4 % in 2023) pins the energy intensity of compute.
Catalogue of conditional scenarios
26 scenarios on technology, substitutability and scarcity, preferences, ownership, competition and policy. The baseline has capability improving at 10 % a year and compute prices falling at 15 % a year to one hundredth of today's level, with a floor on energy-efficiency gains; these are scenario inputs that no data identify.
Methodology
Replication and audit of the source
Restrepo's economy is reproduced exactly before anything is changed; Propositions 1–6 of the source hold in the benchmark to 10⁻⁶.
Propositions in a limit economy
Nine propositions on the labor share, wage levels, balanced growth, the energy bottleneck, valuation ratios, the discount-rate channel, the two-class distribution and the tax-inclusive wage, each compared with its antecedents in the literature.
Perfect-foresight transition
Each scenario is a 275-year path solved to a maximum residual of 10⁻⁹ with a horizon check; cells that do not meet the tolerance are printed as not solved and nothing from their paths is quoted.
Heterogeneous households
Households with different skills, discount factors and productivity histories hold the claims through a fund; general equilibrium is solved to a maximum log excess demand for claims of 9e-05, with the distribution simulated non-stochastically.
Valuation decomposition and empirical discipline
2026 valuations are split into cash-flow and discount-rate news as a Shapley average over the two orderings, because the split depends on the order. A final section states what would falsify the model's description of the current phase in the next decade.
Reproducibility
- The replication package reproduces every number, table and figure of the paper from a clean clone with two commands, one to install the frozen environment and one to run everything; a clean-clone run took 167 minutes of wall-clock on a 24-core machine.
- A test suite of 23 test functions checks the source's propositions in the benchmark, the accounting identities, the residuals and market clearing of every saved scenario, and that the file of quoted numbers regenerates identically.
- All solvers are deterministic and no random seeds are used; each solved scenario is stored with its parameter file, convergence metadata and path arrays.
- All empirical inputs are public. The raw files are not redistributed, but a versioned manifest lists every file with its URL, retrieval date, size and SHA-256 hash; the processed calibration targets are versioned with the package under CC BY 4.0, and the code is under the MIT licence.
Cite this paper
Boucher, S.-P. (2026). From Labor to Capital: Artificial General Intelligence, General Equilibrium, and Asset Valuation. UQO Working Paper No. 24, Université du Québec en Outaouais.
@techreport{boucher2026wp24,
author = {Simon-Pierre Boucher},
title = {From Labor to Capital: Artificial General Intelligence, General Equilibrium, and Asset Valuation},
institution = {Université du Québec en Outaouais, Département des sciences administratives},
number = {24},
type = {UQO Working Paper},
year = {2026},
month = {nov},
url = {https://www.spboucher.ai/research/wp24},
}More in Finance & Economics