The Maze: AI infrastructure spending has moved from boom to historical outlier in three years. Normalised to its 2023 trough, hyperscaler capex reached 4.46× by 2026. Canal mania took five years to peak at 4.09×; railway mania never passed 2.65×. That does not make a crash dateable. It does make the commercial hurdle brutal: revenue, utilisation and productivity now have to catch infrastructure that was financed before demand fully arrived.
AI compressed the biggest historical build-out into the shortest window. The BIS comparison starts every episode at 1.0. AI then rises to 1.50× after one year, 2.54× after two and 4.46× after three. Canal spending needed five years to reach 4.09× before collapsing below its starting level in year seven. Railway investment peaked at 2.65× in year four; the Roaring Twenties and dotcom cycles stayed below 2×. The point is pace, not equal dollars: AI has already cleared the historical amplitude benchmark while its revenue model is still being negotiated.
The denominator makes the result useful—and easy to misuse. These are different series normalised to different troughs: US canal construction from 1835, British real investment from 1843, US fixed-asset investment from 1921, information-processing investment from 1995, and hyperscaler capex from 2023. The AI line also includes 2026 expectations gathered from earnings calls and press releases. So this is not a like-for-like valuation test or a countdown to a bust. It is a capital-cycle alarm: an unusually large commitment has arrived unusually early.
Competition can make rational companies collectively overspend. Alphabet, Amazon, Meta, Microsoft and Oracle are set to deploy more than $1 trillion across 2025-26. Their average capex-to-revenue ratio rises from roughly 32% in 2025 to a forecast 48% in 2026. The investment-race model explains why: when only a few firms may dominate, waiting is strategically dangerous. Each player buys capacity to avoid being locked out. The conservative calibration still produces investment about 50% above the socially efficient level. Less elastic demand can push the gap toward three times.
The financing structure turns a capex miss into a network problem. Spending is moving beyond operating cash flow toward debt and private credit. Hyperscalers, chipmakers, AI labs, neoclouds and data-centre developers are also connected through equity stakes, purchase commitments and long leases. The financing shift means a revenue disappointment would not stop at lower equipment orders. It could weaken suppliers, trigger asset sales and tighten credit. Specialised hardware is productive when demand is strong, but poor collateral when everybody tries to exit together.
Why it matters: The technology can work and the investment cycle can still break. Canals, railways and the internet all created durable economic infrastructure while destroying capital for investors who funded the wrong capacity at the wrong moment. AI operators should stop treating capex as proof of demand. The defensible edge is utilisation: workloads that customers renew, pricing that covers compute, and assets that stay valuable if one model or tenant disappoints. The race is no longer to build fastest. It is to make committed capacity earn.
