The pace of investment in AI, some of it circular, has raised real questions over the past year about how long the current supercycle can hold—the industry has been building at extraordinary intensity for several years now, and the doubts aren't new. What has shifted is the comparison people reach for to think them through.
The habit so far has been to size the buildout against a small set of famous American engineering programs: the Manhattan Project, the Apollo program. Both are attached to almost every headline about compute spend, and both fall short of the scale. Manhattan peaked around 0.4% of US GDP, the equivalent of about $120 billion a year today. Apollo peaked around 0.8%, roughly $240 billion. By some measures, AI infrastructure spending is already running at or above the GDP share either program reached at its peak, and Manhattan and Apollo were missions with a budget line to begin with, not the kind of open-ended, general-purpose buildout underway now.
There's a much less familiar episode in American economic history that fits the category and, at its peak, was considerably larger—and it happens to be the one a very visible person in this industry has just pointed everyone toward.
The Book to Be Read
On Microsoft's latest earnings call, Satya Nadella told analysts: "all of us are reading this, '1873' is the book to be read." He meant Liaquat Ahamed's new book, about 1873, the year the global bond market cracked under railroad overbuilding and opened a depression that ran for the rest of the decade. Ahamed himself has since suggested that Nadella may be drawing the wrong lesson: each hyperscaler can be careful and individually rational while the system collectively overbuilds.
1873 followed a railroad mania. Money went into new lines faster than the freight and passengers using them could justify, much of it borrowed against demand that hadn't arrived yet. When the bonds behind that debt started failing, the panic wiped out banks, brokerages, and a large share of the capital that had financed the boom. Railroads, at their financing peak in the 1880s, absorbed at least 6% of US GDP—several times what AI capex represents today. On scale alone, the reference makes sense: few infrastructure booms since have run at that scale, relative to the economy financing them. Railroads are closer to what's happening now than either Apollo or Manhattan: general-purpose infrastructure financed years ahead of the demand that would eventually justify it.
What the Panic Didn't Destroy
The analogy holds because of what the panic did and didn't destroy: the financing collapsed, not the railroads. Too many lines, laid too fast, on debt sized to demand that hadn't arrived, and the crash wiped out the capital and companies built on it. The tracks stood. The financial record didn't. Within a decade, those overbuilt networks had become some of the cheapest infrastructure the economy had ever had—feeding new cities, industrial belts and agricultural markets. The equivalent tracks this time may not be the GPUs themselves, which depreciate quickly, but the power, fiber, land, cooling and data-center capacity built around them, and the immaterial transformation of companies and mindsets.
The distinction that matters: overinvestment and overvaluation aren't the same thing. Investors can lose a fortune building the very infrastructure that lets someone else make one afterward: the same money can be badly overspent while the infrastructure it paid for turns out enormously valuable. Apollo and Manhattan never got to test that, because they were missions with an end date, not infrastructure built to keep running once the money was gone. Rails did, and history has since forgotten which fortunes were made or lost in building them—the tracks disappeared into the landscape, the way rail, electricity, and TCP/IP eventually do.
1883
Ten years after the panic, the same railroads changed something else. Their expansion had surfaced a problem: until 1883, every American town kept its own time, set by the sun, and each railroad ran on its own clock, until shared schedules across several towns made that dangerous. On November 18, 1883, the railroads imposed four standard time zones across the continent. Clocks were reset as standard noon reached each zone, and some towns saw noon twice, hence the name the day kept: the Day of Two Noons.
What forced the change was scale. A convention that had worked for centuries became a liability once a network large enough to need coordination sat on top of it. The railroads set the standard and the towns followed, because the network could not run on dozens of clocks.
Similarly, AI has already had its moment of convergence on the technical side: since 2024, open protocols and conventions have multiplied—MCP, A2A, AG-UI, AGENTS.md, Skills, UCP, ARD, OKF, among others—and gradually synchronized the clocks. Each player pushes its own and some overlap, but rivals adopt each other's, and several now sit under the Linux Foundation.
But the clocks are synchronized between systems, not inside companies, which still run on "local time". Much of what a company knows how to do is not written down: which version of a document is the reference, who has to approve what before it ships, which system people actually trust. Employees pick this up over time. An agent working inside the same company has no way to tell, and no colleague to correct it. If agents end up doing real work, companies will probably have to make explicit what they have so far left implicit, much as the railroads forced towns to give up local time.
If 1873 shows that infrastructure can be real and still overbuilt, 1883 shows what comes next: once the network is everywhere, everything connected to it has to conform. Beyond how much is being spent, the open question for this cycle is which of our unwritten local conventions AI will force us to make explicit.
