AI's CapEx Boom Rhymes With the Railroad Bond Bust of 1873

The piece uses Jay Cooke and the Panic of 1873 as a lens on today’s AI infrastructure spending. Cooke financed the Northern Pacific Railway by pioneering mass retail bond sales — patriotic appeals, a controlled press, and an army of salespeople — but the railroad’s bottomless capital needs met a global credit crunch in 1873, his firm collapsed, and the resulting depression rippled for decades. Liaquat Ahamed’s new book ‘1873’ draws the parallel explicitly: adjusted for the size of economies, the railroad boom’s annual bond sales map to roughly $600 billion, about what the major tech companies are projected to invest in AI in 2026. Satya Nadella flagged the same book on Microsoft’s earnings call.

The real shift is how that buildout is now funded. A year ago nearly all of Big Tech paid for capex out of free cash flow; since late 2025, Oracle, Meta, Alphabet, and Amazon have turned to debt, raising $194 billion by mid-2026 after $108 billion in all of 2025. The strain is visible — spreads are widening, 86% of this year’s bonds already trade above their issuance yields, and demand coverage has fallen from 5x to under 2x. Microsoft is the outlier still funding capex from cash ($19.6 billion last quarter), while Google’s $85 billion equity raise — including a symbolic $10 billion slug bought by Berkshire Hathaway — reads as a bet that whoever can bring the most cash to bear will win the fight for scarce compute.

Against that backdrop, Google’s frontier-model position looks shaky. With Demis Hassabis kicked upstairs to chairman and Jeff Dean and other senior researchers leaving, SemiAnalysis argues DeepMind is no longer a frontier lab and its odds of retaking the state of the art have collapsed. The counterintuitive winner, per that analysis, is Google Cloud: with Gemini no longer competing for scarce chips, GCP’s compute allocation frees up and its revenue growth should accelerate.