1\section{Calibration and identification}2\label{sec:calibration}34The base point is the US economy in 2026, normalized so that GDP per efficiency unit is one and the two skills5together supply one unit of labor. It is constructed as a balanced growth path of the frozen-technology economy6and then re-derived for each scenario so that the compute sector's 2026 rental is consistent with the price path7the scenario anticipates. Table~\ref{tab:calibration} lists every parameter with its value, its status ---8\emph{data} (read from a retrieved series), \emph{calibrated} (implied by base-point shares and the model's9steady-state conditions), \emph{literature}, \emph{assumed} (a documented assumption with no series behind10it) or \emph{scenario} (a forward-looking input that no data identify) --- and its source. The replication package11records the provenance of each value in \texttt{results/data/parameters.json}; Appendix~\ref{app:data} documents12the series.1314\begin{table}[htbp]15\centering\scriptsize16\caption{Parameters of the baseline, with status and source.}17\label{tab:calibration}18\resizebox{\textwidth}{!}{\input{tables/tab02_calibration}}19\end{table}2021\subsection{What the data pin}2223Shares and stocks come from the national accounts: the labor share (PWT, 0.568 in 2023), residential structures24(1.17 times GDP) and their depreciation (2.3\%), nonresidential depreciation (8.9\%), and the investment share of25information-processing equipment (2.57\% of GDP), which capitalized at $\delta_Q+g_Z+g_{pQ}$ gives a value of26compute capital of 6.2\% of GDP. The data-center share of electricity (4.4\% in 2023) pins the energy intensity27of compute. Given these, the model's steady-state conditions imply the rest of the base point: the stock of28structures is the residual that makes factor payments add up to GDP (1.71 times GDP, against 1.22 for private29nonresidential assets in the data --- the difference is the model's broader definition, which includes public30and intangible capital), the housing land share (0.42) and the energy resource share (0.35) make the housing and31energy Euler equations hold, and the CES weights are the base-point cost shares. The price of compute services32and the 2026 compute-equivalent of a routine worker ($\alpha_{1,0}=\AlphaOneBase$) follow from the compute Euler33equation and the task-1 tie $A=w_L/\alpha_1$; they imply that compute supplies 22\% of the effective input of the34routine task in 2026, an untargeted magnitude that we find plausible for the current penetration of software and35AI in routine work.3637\subsection{What the data do not pin}3839Three groups of parameters are not identified by any historical observation. The elasticities between the40task composite and physical complements ($\sigma_N=0.7$, $\sigma_T=0.6$) are literature values estimated at41compute intensities far below those of the scenarios \citep{OberfieldRaval2021,Chirinko2008}; every asymptotic42statement of Section~\ref{sec:theory} changes sign at $\sigma_T=1$, and Section~\ref{sec:robustness} reports43$\sigma\in\{0.4,0.6,1,1.5\}$. The technology paths --- the rate of capability progress $g_\alpha$, the rate of44decline of the compute price $g_{pQ}$ and its floor, the energy-efficiency path --- are scenario inputs. The45Epoch AI hardware data imply a doubling of FLOP/\$ every two years ($g_{pQ}=0.34$); the baseline uses $0.15$46because the price of \emph{installed} capacity includes buildings, power and cooling that do not follow Moore's47law, and treats $0.34$ as the fast-Moore scenario. The floors matter most: Proposition~\ref{prop:bgp} says that the48wage in the AGI limit is proportional to the product of the compute-price floor and the terminal compute per49worker, so the baseline floor of $10^{-2}$ (compute capacity one hundred times cheaper than in 2026 once Moore's50law ends, around 2057) is the single assumption on which the level of wages after 2060 depends. Finally, the51terminal compute requirements $\alpha^\infty_i=(1,2,4,8)$ are a normalization plus an ordering: one unit of522026 capacity, which today does about one fifteenth of a routine worker's work, eventually does the whole of it,53and the expert task needs eight times as much. The reader who finds this optimistic about algorithms should read54the slow-capability scenario; the reader who finds it pessimistic, the fast one.5556\subsection{Heterogeneous households}5758The household block uses a productivity process with persistence 0.95 and innovation standard deviation 0.20,59two assumed discount factors (0.93, 0.97) and a third calibrated so that stationary savings at the 2026 prices60equal the two-class wealth ($\betaHigh$). The implied wealth Gini ($\GiniBase$), top-10\% share ($\TopTenBase$)61and bottom-50\% share ($\BottomFiftyBase$) compare with 0.85, 0.69 and 0.02 in the Distributional Financial62Accounts; the top-1\% share ($\TopOneBase$ against 0.325) is under-predicted for the usual reason that the model has63no return heterogeneity.64