Four Build-Outs: an exploratory data note on capital, revenue and survival in British railways, US railroads, telecom and AI infrastructure

This is an exploratory data note, not a working paper and not a finding. It compiles a sourced, row-level dataset of 259 observations across four infrastructure investment booms, the British railway mania of 1844 to 1850, the US railroad build-outs of 1865 to 1873 and 1880 to 1893, the US telecom boom of 1995 to 2002, and the AI infrastructure build-out of 2023 to September 2026, and derives every reported number from it with one standard-library script. Five measures and four falsification criteria were registered in advance of the data. The main result is about the instrument rather than about AI: not one of the four pre-registered criteria yields a verdict on the present build-out. Three cannot be adjudicated at all, because their benchmarks or populations are not the quantities the criteria were written about, and one of the five measures was never obtained on a common basis. The fourth criterion, that the four largest builders' capital spending stays inside their operating cash flow, is met for 2025 at a ratio of 0.648, which the pre-registration counts as a break in the analogy. For January to September 2026 it is not robust and no verdict is reported: the ratio is 0.997 on cash purchases of property and equipment and at least 1.0056 once principal payments on finance leases are counted, as one of the four companies' own definition of capital expenditure does, and the pre-registration never said which definition governs. That the one adjudicable criterion turns on a definitional choice left unfixed in advance is the clearest single result of the exercise. The scale comparison is reported as measured percentages with no verdict and no direction, because its AI numerator is four companies' global capital spending over US GDP rather than a domestic investment share. The three historical cases were selected because they ended badly, so nothing here establishes a base rate, and this is the first limitation stated. Of the 52 rows that carry a reported number, 49 have been read back at a cited source across two checking rounds and 3 have not, those three being figures for which the checks found and read no company or official source, which is weaker than none existing: one capital spending guidance figure corroborated only in a third-party transcript, and two revenue run-rates reported by the press and attributed to people familiar with the matter. Each is marked in the data and no reported result rests on one. The note was written and reviewed by AI systems and no human has verified the data rows; the AI use statement sets this out in full, including one material error the first source check made, the independent review round that caught it, and the second round whose finding went against the note's own adjudicated result. Against the absence of human verification the deposit carries a machine-auditable trail: the snapshot scripts record for each source the URL, the retrieval time, the HTTP status, the byte length and the SHA-256 of the exact bytes read, so a reader can re-fetch and compare rather than take the account on trust. The deposit includes the dataset, the scripts, the pre-registration, all three source checks and the two review rounds that objected to the work.