The Rails Were Never the Railroad
9 min read
›Contents
Twice in a generation, the United States built far more railroad than it could pay for. The money came as mortgage bonds, sold in London and Amsterdam and New York on the promise of traffic that had not arrived yet, and when the promise ran ahead of the traffic the bonds defaulted, the builders went broke, and the courts took the lines. In 1893 alone, more than 27,000 miles of railroad passed into receivership, and between 1894 and 1898 some 40,503 miles changed hands at foreclosure sales, more than a fifth of the country's track.1
The people who got rich were the ones standing at those sales. Everyone knows that part. What is worth looking at closely is what they bought, because it was not what the sellers thought they were selling.
The foreclosure sale#
James J. Hill's fortune started with a wreck. The St. Paul and Pacific had gone into receivership in the panic of 1873 with its line to the Canadian border unfinished and its bonds in the hands of Dutch investors who had never seen Minnesota. Hill and three partners bought those bonds for a fraction of their face value, mostly with borrowed money, finished the road, and ran it as the St. Paul, Minneapolis and Manitoba. By 1889 it was the Great Northern.2 When the next panic came in 1893 and the Northern Pacific, the Union Pacific and the Santa Fe all went into receivership, the Great Northern did not, and Hill spent the rest of the decade taking control of the Northern Pacific out of its bankruptcy.
E. H. Harriman's fortune started with someone else's wreck too. The Union Pacific, the first transcontinental, failed in 1893 and came out of reorganization in 1897. Harriman arrived the next year, toured the line, and over the following decade put something like $160 million of capital into it: lower grades, straighter curves, heavier rail, bigger engines, longer trains.3 The road that had been a byword for shoddy construction became the standard of the industry, and the return on that money was one of the great returns of the era.
It is tempting to read these two stories as "buy the dip." That is the wrong lesson, and the next two stories show why.
What they actually bought#
Start with the iron. A ton of Bessemer steel rail cost about $120 in 1873. By 1885 it was under $30, and by 1898 under $18.4 Two forces did that, and they are the same two forces that act on any capital good at the end of a boom: demand collapsed, and the technology moved. Iron rail, the stuff most of the 1870s track was laid with, was being torn out for steel within a decade; by 1900 the trade association reported that iron rails had practically disappeared.5 Whoever had paid boom prices for iron owned, within ten years, scrap.
What held value was everything under the rails. The surveyed route. The grade, with its cuts and fills already dug. The tunnels and the bridges. The land, and above all the terminal land in the cities. The charter and the franchise. The engineers' knowledge of where the line could be bent and where it could not. Hill and Harriman were not buying track; they were buying routes, and then re-equipping them. Harriman's $160 million went into the roadbed and the rolling stock, not into the rails he inherited.
Two episodes from the same decade make the point in the extreme. In 1883 a syndicate opened the West Shore Railroad up the far bank of the Hudson, parallel to the New York Central the whole way to Buffalo. It was bankrupt within two years. J. P. Morgan, the Central's banker, brokered a settlement in 1885 under which the Central took the West Shore over at a heavy discount to what it had cost to build, and the Central and its successors have run freight on that route ever since.6 The other half of the same settlement concerned the South Pennsylvania Railroad, which William Vanderbilt had been boring across the Alleghenies to break the Pennsylvania Railroad's hold on Pittsburgh: nine tunnels through six ridges, half the grading done, and then abandoned. The grade sat for fifty years. In 1940 the Pennsylvania Turnpike opened on it, running through six of the tunnels the railroad had dug.7
A route, once cut, outlasted the company that cut it, the financing that paid for it, the technology that first ran on it, and in the South Penn's case the entire mode of transport it was built for. The rails did not outlast anything.
The rebate#
There was a second kind of winner in the bust, and it was not a railroad at all.
When too many lines were built between the same two cities, they fought rate wars, and the party who could offer a line steady volume could name his price. John D. Rockefeller understood this earlier than anyone. His first great rebate came from the Lake Shore at the end of the 1860s, in exchange for a guarantee of sixty carloads a day; by the end of the 1870s Standard Oil was shipping on terms no competitor could get, and that freight advantage was as much a part of the monopoly as the refineries were.8 A glut of capacity did not enrich the owners of the capacity. It enriched the shipper who could turn cheap capacity into something.
Which part is the rails#
Now the analogy, with one rule: nothing here is a forecast. Whether the price of compute goes up or down from here is a question reasonable people disagree about, and this piece does not need an answer to it. The argument is about what survives either way, and it is this: in any AI stack a company runs, the model is the rail.
It is the most capital-intensive part, the most visible part, and the part most certain to be torn out. Every model your team runs today will be replaced by a better one, the way iron came out for steel, and the replacement will be cheaper per unit of work than the thing it replaces. That has been the pattern so far, and it holds whatever happens to the chips. So the question an operator should ask of an AI investment is Harriman's question: if the rails were pulled up tomorrow, what would I still own?
The usual answer, at most companies, is "the prompts." A library of instructions written for one model's quirks, in one vendor's format, that will be rewritten when the model changes. That is iron. It was expensive to lay, it was skilled work, and its scrap value is nil.
The right-of-way is the other thing, the thing that is slow and expensive to build because it has to be cut through the actual ground of the business: what the company has decided, and the words those decisions rest on; how it actually operates, exceptions included; the record that says which version of a rule is current and who approved it; and the ground truth that any piece of work, by a person or an agent, can be checked against. None of it belongs to a model. All of it survives a model change intact. A business that has cut that route can re-equip it with every generation of model that comes along, and each one runs faster on it than the last.
The rebate is where cheap compute comes in, and it is the part people get backwards. If capacity does get cheap, the beneficiary is not whoever owns the most of it. It is whoever can convert it into something, and the conversion that matters in a world of abundant generation is correctness: running the work more than once, checking it against the record, replaying it, holding a change that passes every test but contradicts a decision someone made. Every one of those moves spends compute and every one of them is impossible without a record to check against. Cheap capacity plus no record is just more unverified output, faster. Cheap capacity plus a record is Rockefeller's position.
That record is what we build at Kelright. The Brain keeps a cited record of what a company decided, drawn from its own meetings and documents, approved by its own people, and holds the work agents do to it before the work lands. It is route, not rail: it does not care which model did the work.
The Operator's Move#
The free move this quarter is the swap test. Take one workflow your team runs with agents and list everything in it: the model, the prompts and instruction files, the tools, the checks that say the output is right, and the record (if there is one) of the decisions the work is supposed to honor. Then sort each item into two columns. Route: survives a change of model intact. Rail: has to be relaid. Count the route column. If it is short, you are a tenant on someone else's line, and the rent is whatever the next model costs. Where the decisions a workflow depends on exist only in someone's head, that is where the route has not been cut yet, and the cutting is the work. Nobody is going to do it but you, and once it is done every model you ever run will run on it.
References#
-
Stuart Daggett, Railroad Reorganization (Houghton Mifflin, 1908), preface. The share is against roughly 180,000 miles of road operated in the mid-1890s. ↩
-
Albro Martin, James J. Hill and the Opening of the Northwest (Oxford University Press, 1976; Minnesota Historical Society Press, 1991). ↩
-
Maury Klein, The Life and Legend of E. H. Harriman (University of North Carolina Press, 2000). ↩
-
Annual averages of the "Wholesale Price of Steel Rails for Pennsylvania," NBER Macrohistory Database series m04181, compiled from Iron Age and the American Iron and Steel Association's reports; dollars per long ton. ↩
-
Victor S. Clark, History of Manufactures in the United States, vol. 3 (McGraw-Hill for the Carnegie Institution of Washington, 1929), "Rails," p. 118: "Iron rails had practically disappeared, less than a thousand tons being made that year." ↩
-
Jean Strouse, Morgan: American Financier (Random House, 1999), on the 1885 settlement between the New York Central and the Pennsylvania. ↩
-
"Turnpike Tunnel History Spans Generations," Roads & Bridges, April 18, 2005. ↩
-
Ron Chernow, Titan: The Life of John D. Rockefeller, Sr. (Random House, 1998). ↩