1\section{Related literature and what is different here}2\label{sec:literature}34\paragraph{Task models of automation.} The work equation $X=L+Q/\alpha$ descends from the task framework of5\citet{AutorLevyMurnane2003}, \citet{AcemogluAutor2011} and \citet{Zeira1998}, in which technology replaces6labor task by task. \citet{AcemogluRestrepo2018} embed the task set in a growth model with endogenous creation of7new tasks, \citet{AcemogluRestrepo2019b} separate displacement from reinstatement, and8\citet{AcemogluRestrepo2022} estimate that task displacement accounts for most of the rise in US wage inequality9since 1980; \citet{AcemogluRestrepo2020} and \citet{HemousOlsen2022} provide the industrial-robot and10automation--inequality counterparts. \citet{Restrepo2025} takes the framework to its limit: when every task can11be replicated by compute, growth is the growth of compute, wages are the value of the compute they save, and12the labor share goes to zero. We keep his primitives and change one thing in the production block --- compute is a13reproducible, priced, depreciating capital good that needs energy and works alongside structures and land --- and14we add everything on the household side.1516\paragraph{The economics of transformative AI.} \citet{AghionJonesJones2019} show that Baumol-type bottlenecks17determine whether automation produces a growth explosion or a sequence of level effects; our regime map18(Table~\ref{tab:regimes}) is the asset-pricing counterpart of that observation. \citet{TrammellKorinek2023}19survey growth under transformative AI; \citet{KorinekSuh2024} characterize the transition to AGI in task models20with a distribution of task complexity and show that wages can collapse once compute crosses a threshold, which is21the mechanism of Propositions~\ref{prop:regimes}--\ref{prop:moore} in a different parametrization.22\citet{KorinekStiglitz2019} and \citet{KorinekJuelfs2022} discuss distribution and the ``future of work'';23\citet{Acemoglu2025} argues from Hulten's theorem that the near-term TFP gains from current AI are modest, a24discipline we respect by starting the 2026 base point at the observed compute share. \citet{Nordhaus2021} tests25and rejects the singularity hypothesis on historical data; our ``Moore's law never ends'' scenario is where the26model itself produces one, and our baseline does not. \citet{MollRachelRestrepo2022} is the closest antecedent on27distribution: automation raises returns to wealth and so wealth inequality in an Aiyagari economy; we obtain28their mechanism as one of three cells of Table~\ref{tab:regimes} and add the valuation side. \citet{Hanson1998}29and \citet{Bostrom2014} are the long-run speculative antecedents the source cites.3031\paragraph{Growth, bottlenecks and networks.} \citet{Jones1995}, \citet{Kortum1997} and \citet{Segerstrom1998}32provide the semi-endogenous growth structure of the source's science block; \citet{Bloom2020} and33\citet{Jones2022} the evidence that ideas are getting harder to find. \citet{Baumol1967} and34\citet{CominLashkariMestieri2021} are the references for unbalanced growth with gross complements, the force35behind our scarce-complement regimes. \citet{Hulten1978}, \citet{Acemoglu2012}, \citet{Jones2011} and36\citet{BaqaeeFarhi2019,BaqaeeFarhi2020} supply the production-network machinery; we use it only as the37accounting discipline that gross output of compute and energy is intermediate, not final, output.3839\paragraph{Heterogeneous households and incomplete markets.} The household block is a40\citet{Bewley1986}--\citet{Huggett1993}--\citet{Aiyagari1994} economy with discount-factor heterogeneity as in41\citet{KrusellSmith1998} and \citet{Carroll2017}, solved by the endogenous grid method of \citet{Carroll2006}.42\citet{KaplanMollViolante2018} and \citet{Auclert2021} are the references for transition dynamics with43heterogeneous agents; our relaxation on the interest-rate path is the simplest member of that family.44\citet{KrusellSmith1997}, \citet{ConstantinidesDuffie1996} and \citet{HeatonLucas1996} study asset prices with45heterogeneous agents; because the two-class model nests the heterogeneous economy, the marginal investor in our46pricing equations is always an unconstrained saver. On the dynamics of top wealth, \citet{BenhabibBisinZhu2011},47\citet{Gabaix2016}, \citet{HubmerKrusellSmith2021} and \citet{Fagereng2020} document the role of return48heterogeneity, which we do not model and which is why the model under-predicts the top-1\% share.49\citet{MianStraubSufi2021} is the reference for the demand-side consequences of wealth concentration.5051\paragraph{Asset pricing and valuation.} The stochastic discount factor is the one of \citet{Lucas1978}; we work52under perfect foresight for the transition paths and discuss \citet{MehraPrescott1985}, \citet{EpsteinZin1989}53and \citet{BansalYaron2004} when we interpret the discount-rate channel. The decomposition of valuation changes54into cash-flow and discount-rate news follows \citet{CampbellShiller1988a,CampbellShiller1988b} and55\citet{Campbell1991}, computed exactly along the deterministic path. Investment valuation uses56\citet{Tobin1969}, \citet{Hayashi1982}, \citet{AbelBlanchard1983} and \citet{Cochrane1991}.57\citet{GreenwaldLettauLudvigson2025} attribute most of the post-1989 rise in US equity values to a reallocation of58rewards from labor to shareholders; \citet{FarhiGourio2018} decompose macro-finance trends into market power,59intangibles and risk premia; \citet{RachelSummers2019} and \citet{EggertssonMehrotraRobbins2019} on the secular60decline in real rates. Our Proposition~\ref{prop:wealth} is a warning against reading the first of these results61forward: the income that labor loses to compute is capitalized at the duration of chips, not of equities.6263\paragraph{Factor shares, markups and intangibles.} \citet{KarabarbounisNeiman2014}, \citet{ElsbyHobijnSahin2013},64\citet{Rognlie2015}, \citet{Barkai2020}, \citet{Autor2020}, \citet{KehrigVincent2021} and65\citet{BergholtFurlanettoMaffeiFaccioli2022} document and dispute the decline of the labor share; the66compute-equivalent wage gives a structural reading of a further decline. \citet{DeLoeckerEeckhoutUnger2020},67\citet{CrouzetEberly2019} and \citet{CorradoHultenSichel2009} motivate the markup and intangible scenarios.68\citet{OberfieldRaval2021} and \citet{Chirinko2008} inform the elasticities $\sigma_N$ and $\sigma_T$;69\citet{CooperHaltiwanger2006} the adjustment costs; \citet{GommeRavikumarRupert2011} the return on capital.7071\paragraph{Land, housing and wealth.} \citet{KnollSchularickSteger2017} and \citet{DavisHeathcote2007} show that72land, not structures, drives long-run house prices; our housing block separates the two. \citet{Piketty2014},73\citet{PikettyZucman2014}, \citet{SaezZucman2016}, \citet{SmithZidarZwick2023}, \citet{KuhnSchularickSteins2020}74and \citet{Jorda2019} provide the wealth and return facts. We do not identify $r>g$ with rising concentration:75Proposition~\ref{prop:wealth} shows that $r$ and the composition of claims matter separately.7677\paragraph{Policy.} \citet{GuerreiroRebeloTeles2022}, \citet{Thuemmel2023} and \citet{CostinotWerning2023}78analyze robot taxes; our Proposition~\ref{prop:tax} gives the limit-economy version (a compute tax raises wages79one-for-one with the tax-inclusive user cost). The land-tax and citizen-fund scenarios go back to80\citet{George1879}, \citet{Meade1964} and \citet{Atkinson2015}, and to the ``universal dividend'' the source81mentions in its discussion.8283\paragraph{Compute and energy.} \citet{Sevilla2022} and \citet{HobbhahnBesiroglu2022} document compute trends and84the price--performance of hardware (doubling about every two years in the data we use); \citet{Koomey2011} the85historical efficiency of computing; \citet{Masanet2020}, \citet{Shehabi2024} and \citet{IEA2025} the energy use86of data centers (4.4\% of US electricity in 2023). These are the sources behind the compute-price,87energy-efficiency and energy-share inputs of Section~\ref{sec:calibration}.88