Literature matrix — UQO Working Paper No. 24
"From Labor to Capital: AGI, General Equilibrium, and Asset Valuation"
Keys refer to paper/references.bib. Our paper, in one sentence: compute is a reproducible capital good with a falling relative price and depreciation (not an exogenous endowment); it enters a nested CES with scarce complements (structures, land, energy); heterogeneous households accumulate wealth under incomplete markets; an explicit stochastic discount factor (SDF) prices compute equity, structures, land and housing; we derive when the labor share → 0 survives, when wages → 0 versus a positive floor, and when valuation ratios fall despite higher output; and we compare a land tax, a compute tax and a citizen capital fund.
"What it does" lines are kept factual; where I did not read the paper closely enough to characterise it, the line says so.
1. Task models, automation and AGI
| Key | What it does | How our paper differs |
|---|---|---|
| Restrepo2025 | Theoretical chapter: with AGI, all economically valuable work can be done with compute; distinguishes bottleneck vs. accessory tasks; as compute expands, bottleneck work is automated, output becomes linear in compute and labor, growth is driven by compute expansion, wages converge to the compute cost of reproducing human work, and the labor share → 0 (abstract, p. 1). | Compute is produced (investment, falling price, depreciation) rather than an expanding exogenous stock; we add structures/land/energy as scarce complements in a nested CES, so output is not asymptotically linear in compute and the labor-share → 0 result holds only under conditions we state; we add households, an SDF and asset prices. |
| AcemogluRestrepo2018 | Task-based growth model where automation displaces labor and the creation of new labor-intensive tasks reinstates it; characterises balanced growth and factor shares. | We do not rely on new-task creation to sustain wages; wages are sustained (or not) by scarce complements and compute pricing, and we price the capital stock explicitly. |
| AcemogluRestrepo2019a | Chapter: AI as automation technology; displacement vs. productivity/reinstatement effects; discusses why AI need not benefit labor. | Same contrast as above; our contribution is general-equilibrium asset valuation and wealth distribution, not the labor-demand decomposition. |
| AcemogluRestrepo2019b | JEP exposition of the displacement/reinstatement framework with US evidence on task content. | We take the task framework as given and study the capital side (who owns compute, how it is priced). |
| AcemogluRestrepo2020 | Empirical: industrial robots and US commuting-zone employment and wages. | We are theoretical/quantitative GE; their estimates can inform our calibration of displacement, not our mechanism. |
| AcemogluRestrepo2022 | Task-displacement account of US wage inequality 1980–2016. | We model a representative task continuum with heterogeneous wealth, not heterogeneous skills; inequality in our paper is on the asset side. |
| AcemogluAutor2011 | Handbook chapter: the canonical task/skills framework replacing the canonical skill-biased model. | Framework source only; no asset pricing or capital accumulation there. |
| AutorLevyMurnane2003 | Routine-task substitution by computers; empirical and conceptual basis of task models. | Background; we go to the AGI limit where routine/non-routine distinction collapses. |
| Zeira1998 | Automation of tasks with machines in a growth model; growth through expanding set of automated tasks. | Zeira's machines are generic capital; we distinguish compute from the complements it cannot replace and price both. |
| HemousOlsen2022 | Endogenous automation vs. horizontal innovation; explains rising income inequality and falling labor share with directed technical change. | We hold the direction of technical change fixed (AGI) and focus on capital prices and household wealth. |
| AghionJonesJones2019 | AI and growth: Baumol-type bottlenecks, automation of production and of idea production, singularity conditions. | We adopt the "bottleneck" logic but apply it to physical complements (structures, land, energy) and derive asset-price consequences; no idea-production automation. |
| KorinekSuh2024 | Scenarios for the transition to AGI: task-based model with a compute constraint; wages can collapse or not depending on whether the automatable task set is bounded. | We make compute reproducible and add scarce complements, households and an SDF; their "wages → 0 vs positive" dichotomy is one of the cases we re-derive with capital prices. |
| TrammellKorinek2023 | Survey of growth models under transformative AI (growth theory with AI as capital, singularity, bottlenecks). | Survey; we contribute a specific GE-with-asset-pricing model. |
| KorinekStiglitz2019 | AI and income distribution/unemployment; Pareto-improving redistribution arguments, role of market power. | We quantify distributional outcomes through wealth accumulation and compare specific instruments (land vs compute tax vs citizen fund). |
| KorinekJuelfs2022 | Policy discussion of a future with little labor demand; income support and distribution. | We model the policy menu inside the GE model with asset-price feedback. |
| Acemoglu2025 | Hulten-style bounds on the TFP and GDP effects of AI over a decade; modest numbers. | We study long-run limits rather than decadal bounds; we use Hulten/Baqaee–Farhi tools for complements, not for near-term accounting. |
| Nordhaus2021 | Tests for an economic singularity using substitution elasticities and capital/information shares; finds no evidence yet. | Our model nests the elasticity conditions he tests; we add asset valuation. |
| MollRachelRestrepo2022 | Automation in a heterogeneous-agent (Aiyagari-type) economy raises returns, boosting wealth inequality and the wealth of capital owners. | Closest HA antecedent. We add reproducible compute with falling price, scarce complements and an explicit SDF pricing several asset classes, and we study valuation ratios (which can fall even as output rises). |
| LehrRestrepo2022 | Optimal gradualism: speed of adopting labor-displacing technology when workers face adjustment frictions. | We do not model the transition speed as a policy; our policy section is about ownership and taxation of scarce vs reproducible factors. |
| Eloundou2024 | Occupational exposure of US work to LLMs (task-level ratings). | Calibration input for the automatable task share; not a model. |
| FeltenRajSeamans2021 | AI occupational exposure index (AIOE) linking AI capabilities to occupational abilities. | Calibration input only. |
| Webb2019 | AI exposure from overlap of patents and job tasks. | Calibration input only. |
| BrynjolfssonLiRaymond2025 | Field evidence: generative-AI assistant raises customer-support productivity, most for less-experienced workers. | Micro evidence on complementarity in the short run; our analysis concerns the AGI limit. |
| Hanson1998 | Early growth model with machine intelligence: machines first complement then substitute for labor; wages may fall as fast as computer prices. | Antecedent for "wages tied to compute price"; we add scarce complements, heterogeneous households and asset pricing. |
2. Growth theory and structural change
| Key | What it does | How our paper differs |
|---|---|---|
| Jones1995 | Semi-endogenous growth: ideas production with diminishing returns; growth tied to population growth. | We take productivity in compute production as given and study factor pricing; no idea dynamics. |
| Kortum1997 | Search-theoretic model of research and patenting consistent with Jones-type semi-endogenous growth. | Background for semi-endogenous growth; not used in our mechanism. |
| Segerstrom1998 | Endogenous growth without scale effects (increasing difficulty of innovation). | Same. |
| Jones2022 | Review of semi-endogenous growth; discusses AI and the future of growth. | Background; we use the "ideas" view only to motivate the exogenous fall in the compute price. |
| Bloom2020 | Evidence that research productivity is falling (ideas harder to find), including Moore's law. | Motivates our treatment of the compute-price decline as an exogenous but decelerating trend. |
| JonesTonetti2020 | Nonrivalry of data and its ownership (consumers vs firms). | We do not model data; ownership of compute and scarce factors is our object. |
| Baumol1967 | Unbalanced growth: sectors with stagnant productivity absorb spending and labor ("cost disease"). | Our scarce complements play the role of stagnant inputs; we derive asset prices for them. |
| CominLashkariMestieri2021 | Structural change with nonhomothetic CES preferences; long-run income and price effects. | We use nested CES on the production side; nonhomothetic demand appears only as a robustness extension. |
3. Production networks and Hulten-type aggregation
| Key | What it does | How our paper differs |
|---|---|---|
| Hulten1978 | Sales shares as sufficient statistics for the first-order aggregate effect of sectoral productivity shocks. | We use Hulten's theorem for the first-order bound and go to the nonlinear limit where compute's share rises. |
| Acemoglu2012 | Input–output network structure propagates micro shocks to aggregate fluctuations. | We use a two-to-three input nested CES rather than a full network; no fluctuation analysis. |
| BaqaeeFarhi2019 | Second-order (nonlinear) effects beyond Hulten: elasticities of substitution and factor shares reallocate. | Directly relevant: our labor-share and complement-share limits are nonlinear Baqaee–Farhi effects; we add asset prices. |
| BaqaeeFarhi2020 | Misallocation and productivity in general equilibrium with distortions and networks. | We do not model markups/distortions in the baseline. |
| Jones2011 | Complementarities ("weak links") among intermediate goods amplify productivity differences. | Weak-link logic underlies our scarce complements; we apply it to AGI and asset valuation. |
4. Heterogeneous agents and incomplete markets
| Key | What it does | How our paper differs |
|---|---|---|
| Bewley1986 | Stationary equilibrium with a continuum of consumers facing idiosyncratic shocks and a borrowing constraint (monetary setting). | We build on Bewley–Huggett–Aiyagari but with several priced assets and a drifting technology (falling compute price). |
| Huggett1993 | Risk-free rate in an endowment economy with idiosyncratic risk and borrowing limits. | Same lineage; we have production with compute and complements. |
| Aiyagari1994 | Precautionary saving with uninsurable risk in a production economy; aggregate capital exceeds the complete-markets level. | We add multiple capital goods with different scarcity and price dynamics; the wealth distribution is tied to the AGI transition. |
| KrusellSmith1998 | Aggregate shocks with heterogeneous agents; approximate aggregation. | Our aggregate shock is the AGI transition (deterministic or stochastic arrival); we use sequence-space methods. |
| Carroll2006 | Endogenous gridpoint method for solving consumption-saving problems. | Numerical method we use (or adapt); no overlap in substance. |
| KaplanMollViolante2018 | HANK: heterogeneous agents with liquid/illiquid assets and monetary policy. | We are real, not monetary; our assets differ in reproducibility, not liquidity. |
| Auclert2021 | Sequence-space Jacobian method for solving and estimating HA models. | Numerical method we use for transition paths. |
| KrusellSmith1997 | Portfolio choice and equilibrium asset returns with heterogeneous agents; equity premium with idiosyncratic risk. | We also price risky claims with heterogeneous agents, but on compute, structures, land and housing. |
| ConstantinidesDuffie1996 | Uninsurable persistent idiosyncratic risk can generate any equity premium (SDF with cross-sectional variance of consumption growth). | We use a similar SDF construction but study the level of asset prices across asset classes under AGI. |
| HeatonLucas1996 | Quantitative effect of incomplete markets on risk sharing and asset pricing; modest effects with trading. | Background on the magnitude of incomplete-market premia. |
| Carroll2017 | Wealth distribution and heterogeneous MPCs from preference heterogeneity. | Calibration reference for the wealth distribution. |
| BenhabibBisinZhu2011 | Pareto tail of wealth from capital-income risk with finitely lived agents. | Return heterogeneity in our model comes from asset-class exposure (compute equity vs land vs housing), not from exogenous return draws. |
| Gabaix2016 | Dynamics of inequality: speed of transitions to Pareto tails; need for type dependence/scale dependence. | We use their results to interpret transition speeds in our wealth simulations. |
| HubmerKrusellSmith2021 | Quantitative decomposition of US wealth-inequality drivers (returns, taxes, earnings). | We ask the forward-looking question (AGI) with the same machinery. |
| Fagereng2020 | Persistent heterogeneity in returns to wealth (Norwegian registry data). | Empirical motivation for allowing asset-class and return heterogeneity. |
| MianStraubSufi2021 | Indebted demand: rising inequality lowers natural rates; debt dynamics. | Our falling-rate channel comes from compute abundance and saving glut by capital owners; we discuss but do not model debt feedback. |
5. Asset pricing, q-theory and macro-finance
| Key | What it does | How our paper differs |
|---|---|---|
| Lucas1978 | Consumption-based asset pricing in an exchange economy. | Our SDF is consumption-based but aggregated over heterogeneous households in a production economy. |
| MehraPrescott1985 | Equity premium puzzle. | We calibrate premia rather than resolve the puzzle. |
| EpsteinZin1989 | Recursive preferences separating risk aversion from intertemporal substitution. | Used as an extension for pricing long-run AGI risk. |
| BansalYaron2004 | Long-run risk model of asset prices. | AGI arrival is a long-run growth shock; we borrow the framing for the stochastic-arrival case. |
| CampbellShiller1988a / CampbellShiller1988b / Campbell1991 | Log-linear present-value identities; decomposition of price-dividend ratios and returns into cash-flow and discount-rate news. | We use the identity to show when valuation ratios fall despite higher output (discount-rate and share effects dominate cash-flow news). |
| Tobin1969 / Hayashi1982 / AbelBlanchard1983 | q-theory: investment depends on marginal q; equals average q under homogeneity; intertemporal investment with adjustment costs. | Compute investment in our model follows q-theory with a falling investment-good price; q for structures and land behaves differently (scarce). |
| Cochrane1991 | Production-based asset pricing: investment returns equal asset returns. | We use the investment-return link to price compute equity from the compute-production technology. |
| GreenwaldLettauLudvigson2025 | Factor-share shifts (to capital) explain much of US equity gains since 1989. | Directly related: we derive when a labor-share fall raises or lowers equity valuations under AGI. |
| FarhiGourio2018 | Accounting for macro-finance trends with market power, intangibles and risk premia. | Our accounting adds compute abundance and scarce complements as drivers. |
| RachelSummers2019 | Secular stagnation: falling neutral rates in advanced economies. | Our model generates endogenous rate paths under AGI; we relate to their r* decline. |
| EggertssonMehrotraRobbins2019 | Quantitative model of secular stagnation with OLG and ZLB. | We abstract from nominal rigidities. |
| Cochrane2005 | Textbook treatment of SDF asset pricing. | Reference for methods. |
| GommeRavikumarRupert2011 | Measured return to capital in the US and its cyclicality. | Calibration target for returns. |
6. Factor shares, markups, intangibles, elasticities
| Key | What it does | How our paper differs |
|---|---|---|
| KarabarbounisNeiman2014 | Global labor-share decline linked to falling relative price of investment goods (σ > 1). | Our falling compute price is the AGI analogue; we characterise the σ conditions in a nested structure with scarce complements. |
| ElsbyHobijnSahin2013 | US labor-share decline: measurement and offshoring. | Measurement reference. |
| Barkai2020 | Both labor and capital shares fell; profit share rose. | We distinguish the compute capital share from the scarce-factor (rent) share, which maps onto their profit share. |
| DeLoeckerEeckhoutUnger2020 | Rising markups and their macro implications. | Markups are not in our baseline; a robustness extension. |
| Autor2020 | Superstar firms and the labor-share fall. | Firm heterogeneity is absent from our baseline. |
| KehrigVincent2021 | Micro anatomy of the labor-share decline (reallocation to low-share plants). | Same. |
| CrouzetEberly2019 | Weak investment with concentration and intangibles. | We treat compute as tangible, reproducible capital; intangibles are a possible extension. |
| CorradoHultenSichel2009 | Intangible capital measurement and growth accounting. | Same. |
| OberfieldRaval2021 | Aggregate elasticity of substitution between capital and labor built from micro data (below 1 at the aggregate). | We use their estimates as a benchmark for the non-compute part of the nested CES. |
| Chirinko2008 | Survey of σ estimates (σ ≈ 0.4–0.6). | Same. |
| BergholtFurlanettoMaffeiFaccioli2022 | SVAR evidence on drivers of the labor-share decline (automation, markups, bargaining). | Empirical background. |
| Rognlie2015 | Net capital-share rise is mostly housing; accumulation vs scarcity. | Central to our housing/land channel: we model the scarce-factor share explicitly. |
| CooperHaltiwanger2006 | Nature of capital adjustment costs (nonconvexities). | Adjustment costs for structures/compute are parameters in our model. |
7. Land, housing and wealth
| Key | What it does | How our paper differs |
|---|---|---|
| KnollSchularickSteger2017 | Long-run house prices driven mainly by land prices. | Our land asset is priced by the SDF under AGI; we derive conditions for land to absorb AGI rents. |
| DavisHeathcote2007 | Price and quantity of US residential land. | Calibration reference. |
| Piketty2014 | Capital/income ratios and r > g as a force for concentration. | We do not treat r > g as automatically concentrating; distribution follows from who owns compute vs scarce factors and from the SDF. |
| PikettyZucman2014 | Wealth–income ratios across rich countries 1700–2010. | Calibration target. |
| SaezZucman2016 | Top wealth shares from capitalized incomes. | Calibration target. |
| SmithZidarZwick2023 | Top wealth with heterogeneous returns. | Calibration target; motivates return heterogeneity by asset class. |
| KuhnSchularickSteins2020 | Income and wealth inequality 1949–2016 with portfolio composition (housing vs equity). | Portfolio composition across the distribution drives who gains from compute equity vs housing in our model. |
| Jorda2019 | Rates of return on all major assets 1870–2015. | Calibration targets by asset class. |
| Stiglitz2015 | Theoretical perspectives on wealth distribution (Part I: the wealth residual). | Conceptual background on rents vs produced capital, which we formalise. |
8. Policy
| Key | What it does | How our paper differs |
|---|---|---|
| GuerreiroRebeloTeles2022 | Optimal robot taxation: positive when workers cannot retrain, zero in the long run. | We compare a compute tax to a land tax and a citizen capital fund inside a GE with asset prices and wealth heterogeneity. |
| Thuemmel2023 | Optimal taxation of robots in a Mirrleesian setting. | Same contrast. |
| CostinotWerning2023 | Sufficient-statistics approach to optimal technology regulation (robots, trade). | We do not provide optimal-tax formulas; we compare incidence and distributional outcomes of three instruments. |
| George1879 | Land-value taxation. | Our land-tax experiment is the AGI version of George's argument. |
| Atkinson2015 | Proposals including capital endowments and sovereign wealth funds. | Our citizen capital fund experiment is a modelled version. |
| Meade1964 | Property ownership and equality; proposals for widespread capital ownership. | Same. |
9. Compute, energy and AI trajectories
| Key | What it does | How our paper differs |
|---|---|---|
| Sevilla2022 | Training compute trends across three eras of ML. | Calibration of the compute growth trend. |
| HobbhahnBesiroglu2022 | FLOP/s per dollar doubles roughly every 2.5 years (GPU data 2006–2021). | Calibration of the falling compute price. |
| Koomey2011 | Long-run trends in electrical efficiency of computing. | Calibration of the energy complement. |
| Masanet2020 | Global data-center energy use estimates. | Same. |
| Shehabi2024 | US data-center electricity use and projections to 2028. | Same. |
| IEA2025 | Energy and AI report. | Same. |
| HilbertLopez2011 | World capacity to store, communicate and compute information. | Background on compute stock measurement. |
| Amodei2024 / Kokotajlo2025 / DragoLaine2025 / Bostrom2014 | Non-academic (or philosophical) scenarios for transformative AI, including "the intelligence curse" (owners of AI need not share rents). | We provide a formal GE model of the scenario; cited for the narrative only. |
10. Data sources
| Key | Content |
|---|---|
| BEA2026 | NIPA tables (output, income shares). |
| BLS2026 | Productivity and Costs (labor share). |
| FedZ1_2026 | Z.1 Financial Accounts (balance sheets, asset values). |
| FedDFA2026 | Distributional Financial Accounts (wealth by percentile). |
| Shiller2026 | CAPE and US home-price series. |
| Feenstra2015 / PWT1001 | Penn World Table 10.01 (capital, labor shares, prices). |
| StatCan2026 | National Balance Sheet Accounts (Canada). |
| EIA2026 | Electric Power Monthly (prices and generation). |