k and research, but whether profits will show up fast enough to justify today’s valuations and repay borrowed money used to build all this computing capacity. JPMorgan wrote in August that broad-based US productivity gains “remain elusive,” and warned that tech booms can cool when new infrastructure doesn’t deliver returns quickly. Bain & Company framed the hurdle: it estimates hyperscalers and other AI infrastructure builders like Google, Amazon, and Microsoft would need more than $4.2 trillion in new revenue over the next five years, implying entirely new markets have to form.
Why should I care?
Zooming out: Debt has a calendar but AI payoffs might not.
Columbia Business School economist Stijn Van Nieuwerburgh argues the buildout is unusually debt-funded, so even modest demand shortfalls, delays, or lower resale values can create outsized losses. The reason is simple: interest and principal payments are fixed, while the revenue from new data centers can take years to mature, leaving equity holders as a thin shock absorber.
That’s why past manias often had a split ending. The rail network kept expanding after the Panic of 1873, and the internet kept growing after the dotcom bust, but the financing layer repriced first. If AI timelines slip, funding conditions and valuation assumptions could reset for the biggest infrastructure spenders – and for “backbone” beneficiaries like Nvidia that are priced for a long, smooth boom.
