\section{Data sources}
\label{app:data}

All series were retrieved on 2026-10-27 from key-less public endpoints by \texttt{scripts/01\_download\_data.py};
every file is listed with its URL, vintage, size and SHA-256 hash in \texttt{data/raw/manifest.json}, and every
calibration target with its series identifier, period, units and construction in
\texttt{data/processed/calibration\_targets.csv} (162 rows). No number in the paper is typed from memory. The
FRED copies were obtained through a browser-side fetch because the FRED edge servers timed out for scripted
requests on the retrieval day; the files are byte-identical copies and the script re-downloads them when the
server answers. Table~\ref{tab:data} lists the sources used in the calibration and validation.

\begin{table}[htbp]
\centering\scriptsize
\caption{Sources of the empirical inputs.}
\label{tab:data}
\begin{tabularx}{\textwidth}{@{}lXll@{}}
\toprule
Object & Series / document (provider) & Vintage used & Value \\
\midrule
Labor share & PWT 11.0 \texttt{LABSHPUSA156NRUG} via FRED (Groningen; \citealp{Feenstra2015}); cross-checks: BLS \texttt{PRS85006173} index, BEA compensation/GDI and /net domestic income & 2023 & 0.568 \\
Fixed assets & BEA Fixed Assets via FRED: private nonresidential net stock \texttt{K1NTOTL1ES000}, total \texttt{K1PTOTL1ES000}, depreciation \texttt{M1NTOTL1ES000}, \texttt{M1PTOTL1ES000}; GDP \texttt{GDPA} \citep{BEA2026} & 2024 & 1.22, 2.39; $\delta$ 0.089, 0.057 \\
Compute investment & BEA NIPA via FRED: information-processing equipment \texttt{Y034RC1Q027SBEA}, software \texttt{B985RC1Q027SBEA}, IPP \texttt{Y001RC1Q027SBEA} / GDP & 2026Q2 & 2.57\%, 2.49\%, 5.77\% \\
Household balance sheet & Federal Reserve Z.1 via FRED: net worth \texttt{TNWBSHNO}, real estate \texttt{HNOREMQ027S}, owners' equity \texttt{OEHRENWBSHNO} \citep{FedZ1_2026} & 2026Q2 & 6.02, 1.66, 1.10 (× GDP) \\
Wealth concentration & Distributional Financial Accounts, net worth shares \citep{FedDFA2026} & 2026Q2 & top 1\% 32.5, top 10\% 68.9, bottom 50\% 2.3 \\
Real rates & FRED \texttt{DFII10} (10-year TIPS), \texttt{REAINTRATREARAT10Y} (Cleveland Fed), \texttt{GS10}; Treasury real yield curve & 2025 avg; 2026-10 & 1.96\%; 2.47\%; 4.99\% nominal \\
Equity valuation & \citet{Shiller2026} \texttt{ie\_data.xls}: CAPE, D/P, E/P & 2023-09 (end of file) & 30.8; 1.58\%; 4.17\% \\
Housing & \citet{Shiller2026} real home price index; FRED \texttt{CSUSHPISA}, \texttt{CUSR0000SEHA} (rent CPI) & 2023; 2026-07 & real growth 1.5\%/yr (1990--) \\
Data-center electricity & \citet{Shehabi2024}, LBNL-2001637 (eScholarship copy), pp.~5--6, 52 & 2023 & 176 TWh, 4.4\% of US electricity; 6.7--12\% in 2028 \\
Electricity & EIA Electric Power Monthly Table 1.1 \citep{EIA2026}; BLS \texttt{APU000072610} & 2025 & 4\,425 TWh generated; 0.185 \$/kWh \\
Compute price--performance & Epoch AI \texttt{ml\_hardware.csv} \citep{HobbhahnBesiroglu2022}; own OLS of $\log_2$(FLOP/s per \$) on release year, 60 accelerators & 2008--2026 & doubling 2.02 years ($g_{pQ}=0.34$) \\
Input--output & BEA summary use table 2023 (\texttt{AllTablesIO.zip}): compensation / value added by industry & 2023 & 0.513 (all industries) \\
Canada (comparison) & Statistics Canada 36-10-0103-01, 36-10-0580-01 \citep{StatCan2026}; PWT \texttt{LABSHPCAA156NRUG} & 2026Q2; 2023 & labor share 0.49 (COE/GDP), 0.66 (PWT); net worth 5.65 × GDP \\
\bottomrule
\end{tabularx}
\end{table}

\paragraph{Definitions and caveats.} The PWT labor share adjusts for self-employment income and is the
concept closest to the model's compensation of all workers; the BEA compensation/GDI ratio (0.508 in 2025) is
lower because it excludes proprietors' income and includes taxes on production. Depreciation rates divide
current-cost depreciation during the year by the end-of-year net stock. The information-processing-equipment
investment share is capitalized at $\delta_Q+g_Z+g_{pQ}$ to obtain a stock value; it includes computers and
communications equipment beyond AI accelerators, and excludes software, so it is a proxy for the value of
``compute capital'' in the model's sense. The Shiller spreadsheet ends in September 2023; the Case--Shiller index
runs to July 2026. The Epoch price--performance column mixes numerical precisions across generations, which
overstates the trend; the paper uses 0.34 only as the ``fast Moore'' scenario and 0.15 in the baseline. The
LBNL data-center figures refer to electricity consumption; the model's energy block is interpreted as
electricity.