In 1882, America laid more than 11,000 miles of railroad track in a single year. Eleven years later, half the industry was in receivership and unemployment ran between 12 and 18 percent. The track stayed.
As many know, I like trains, so the parallel was hard to miss. Today four companies, Microsoft, Alphabet, Meta, and Amazon, plan to spend about 725 billion dollars on AI infrastructure in 2026, up 77 percent in a single year. Amazon alone is guided to around 220 billion. Every second post calls it a bubble and predicts the pop. They are probably right about the pop. They are asking the wrong question.
The question is not whether AI is a bubble. It is what survives the bust, and who. History answers both. Three things worth holding onto.
The pattern is old, not new. In the 1880s, railroads were 15 to 20 percent of all national investment. Companies laid competing lines side by side into empty prairie, funded by debt, chasing territory nobody had asked them to win. By 1892, fewer than half of rail shares paid a dividend. Then the Philadelphia and Reading collapsed in February 1893 and pulled the rest down. Union Pacific, Northern Pacific, and the Santa Fe all went bankrupt. Swap miles of track for gigawatts of compute and you have read this morning’s headlines.
The infrastructure survived. The speculators did not. About 200 railroads went into receivership after 1893, roughly 41,000 miles of track, and by 1895 a third of the network was bankrupt. The track did not rust into nothing. J.P. Morgan and a handful of disciplined operators bought it cheap, put in fresh capital, cut the ruinous competition, and ran it at a profit for fifty years. The winners were not the ones who laid the most track fastest. They were the ones who bought proven track cheap and made it earn. Discipline beat speed.
This time the track spoils. Railway iron lasted decades, which is exactly why buying it cheap in 1894 was a good trade. AI compute does not. In January 2025, Amazon cut the assumed life of its servers from six years to five and took about 920 million dollars in accelerated depreciation. Its Trainium 2 chips look set for replacement at around 20 months, while the hardware needs roughly three years just to break even. Read that again. The replacement cycle is shorter than the payback. The track was iron. A lot of this buildout is fresh milk with a date stamped on the carton.
The analogy is imperfect, and pretending otherwise would be the nonsense I am in my posts point out. The 1893 crash became a banking panic because the railroads ran on risky debt. The hyperscalers fund most of this from operating cash, so a slowdown is less likely to take the banks down with it. And some of the demand is real today, not a bet on empty prairie. Fair. That said, Amazon still burned about 7.6 billion in free cash flow as spending outran income, and a chip that is scrap in two years does not care whether you bought it with cash or debt. The bill comes due either way.
So we should stop asking whether AI is a bubble. It has the shape of one and will probably deflate like one. The interesting question is which side of 1893 you want to be on. The durable layer survives every cycle: power, land, buildings, fiber, and the workflows on top that earn their keep before the silicon ages out. The rest is fresh milk. The operators who can tell the two apart inherit the network. The ones who cannot are laying redundant track into empty land and calling it strategy.
Here’s a real question before a franc or a dollar is committed to anything with AI stamped on it: Does this spend pay for itself before it depreciates, and how fast. If nobody can tell you, you are not buying infrastructure. You are buying milk.
