Point of View
AI Infrastructure · Capital Markets · Economic History
The AI Infrastructure Buildout Is the Closest Thing to a 19th-Century Railroad Boom Modern Capital Markets Have Produced
By scale, this isn't a comparable technology cycle -- it's a comparable capital cycle. The financing structure is drifting toward the same fragility that broke the railroads. But unlike the railroads, the demand uncertainty may point toward under-building, not over-building.
Feb 2026
Points of View is Vaxa's opinion column. Each article states a position on a question where reasonable people disagree — and makes the argument for that position directly.
VAXA POINT OF VIEW
The AI infrastructure buildout is the closest thing to a nineteenth-century railroad boom that modern capital markets have produced — but the comparison cuts both ways. The financing structure is drifting toward the same fragility that broke the railroads. The demand uncertainty, unlike the railroads, may be pointing toward under-building rather than over-building. Both things can be true at once, and most of the current debate is only arguing one side of it.
Parallel
The four largest hyperscalers are on track to spend somewhere between $650 billion and $725 billion on AI infrastructure in 2026 alone — a 60-plus percent increase over 2025. Measured as a share of U.S. GDP, that spending sits in the same range as the railroad buildout of the 1850s and is approaching the peak of the 1870s railroad boom, when annual spending hit roughly 5% of GDP. Adjust for the fact that AI chips depreciate far faster than railroad track ever did, and by some estimates today's buildout already exceeds the railroad peak in real terms.
This is not a casual comparison. A historian who wrote the definitive account of the 1873 railroad bust — the crash that followed the last private capital boom of comparable scale — has said publicly that the AI buildout is the parallel that worries him most, precisely because the pattern of previous infrastructure manias is so consistent: fiber-optic build in the late 1990s, electrification in the 1920s, railroads in the 1870s. In every case, the underlying technology was real and the long-run economic value was real. The capital cycle funding it was still a bubble, and the timing mismatch between capital deployed and revenue realized still broke a large number of the companies that built it.
Power
The four largest hyperscalers are on track to spend somewhere between $650 billion and $725 billion on AI infrastructure in 2026 alone — a 60-plus percent increase over 2025. Measured as a share of U.S. GDP, that spending sits in the same range as the railroad buildout of the 1850s and is approaching the peak of the 1870s railroad boom, when annual spending hit roughly 5% of GDP. Adjust for the fact that AI chips depreciate far faster than railroad track ever did, and by some estimates today's buildout already exceeds the railroad peak in real terms.
This is not a casual comparison. A historian who wrote the definitive account of the 1873 railroad bust — the crash that followed the last private capital boom of comparable scale — has said publicly that the AI buildout is the parallel that worries him most, precisely because the pattern of previous infrastructure manias is so consistent: fiber-optic build in the late 1990s, electrification in the 1920s, railroads in the 1870s. In every case, the underlying technology was real and the long-run economic value was real. The capital cycle funding it was still a bubble, and the timing mismatch between capital deployed and revenue realized still broke a large number of the companies that built it.
The railroads had hundreds of competitors financed by debt they could not service if revenue was late. The AI buildout has four companies financed by their own profits — until very recently. That gap is closing faster than the comparison usually accounts for.
Demand
"Build it and they will come" assumes the "they" is knowable in advance — a fixed population of freight and passengers whose volume simply needed the track to exist. AI's demand curve is a different kind of unknown. Enterprise adoption is still early. The shift from conversational AI to agentic workflows has driven a sharp increase in the amount of computing each interaction actually consumes, and consumer applications are still expanding into categories that did not exist eighteen months ago. Several analysts now frame the binding constraint on this buildout as physical — power availability, grid interconnection, land — rather than demand. McKinsey's own estimate is that the industry needs 125 gigawatts of incremental AI capacity by 2030, a figure premised on adoption that has not yet been fully priced into current build plans.
That is the genuine asymmetry with the railroads. A 19th-century railroad company that overbuilt track in a low-traffic corridor had built something close to its full addressable value on day one — track doesn't get more useful as an technology matures. AI infrastructure is not fixed in the same way. If model capability and the depth of enterprise and consumer adoption keep expanding at anything close to the current rate, today's build plans may prove to be a floor, not a ceiling, on what is ultimately required — which is a very different risk profile than a rail line built through a town that never grew.
Implication
Treating the AI buildout as a single monolithic bet — bubble or not, overbuilt or not — misses the more useful question, which is that the capital structure risk and the demand risk are not the same risk, and they resolve on different timelines. The financing risk (whether hyperscalers can service the debt and lease obligations now layering onto what was once a cash-funded buildout) is a near-term balance sheet question that could surface in the next downturn regardless of how AI adoption ultimately plays out. The demand risk (whether the applications built on this infrastructure ever generate revenue commensurate with the capital) is a longer, structural question that will not be answered for years — and it may resolve differently by layer, with power, cooling, and memory providers proving out their economics well before the application layer does.
For organizations making decisions in and around this buildout — as suppliers, customers, or capital allocators — the railroad parallel is most useful as a reminder that the technology being real and the capital cycle being fragile are not contradictory positions. Every major infrastructure mania in the last 150 years has produced both outcomes at once: a genuinely transformative asset base, and a financially brutal shakeout for a meaningful share of the companies that built it. Betting on which side of that split any given position falls on is the actual strategic question — not whether the buildout as a whole is justified.
The Question to Ask
Is your organization's exposure to this buildout positioned to survive a financing shakeout even if the demand thesis ultimately proves correct — or does your plan quietly assume both have to go right at the same time?