Reading the AI capex cycle like an old industrial cycle

Strip away the vocabulary and the current AI build-out is a familiar story: a new general-purpose technology, a race to build capacity before competitors do, and capital spending growing far faster than the revenue it currently supports. Railroads, fiber in 1999, shale in 2014 — the technology differs, the financing rhythm does not.

Three numbers are worth following. The first is the gap between hyperscaler capex growth and cloud revenue growth; a widening gap is not automatically a bubble, but it is automatically a question. The second is depreciation policy: extending assumed server lifetimes flatters current earnings and defers the reckoning. The third is who funds the marginal data center — corporate balance sheets are one thing, off-balance-sheet vehicles and private credit are another.

The historical pattern is not that the technology fails. Railroads and fiber both worked; the overbuilding was real anyway, and returns accrued to whoever bought the capacity cheaply in the second act. For an investor the useful stance is neither cheerleading nor denial: own the cash flows that exist today, and let someone else pay full price for the ones projected for 2030.