The AI industry's spending narrative has a blind spot: capital costs get the headlines; ongoing maintenance costs get almost none.
The past two years of hyperscaler earnings calls have produced a precise vocabulary for AI build-out — GPU procurement, power purchase agreements, data center footprints. What that vocabulary omits, according to reporting on engineer Shashidhar Bhat's work, is the recurring cost of keeping those clusters healthy after they're built. The figure attached to that gap: $2 trillion. Bhat is working on addressing it.
Capital expenditure is easy to track because it shows up in discrete announcements, investor days, and earnings guidance. The ongoing cost of maintaining GPU clusters at scale is diffuse, harder to aggregate, and therefore systematically underdiscussed. As hyperscalers collectively commit hundreds of billions to new infrastructure each year, the maintenance obligation on existing installations compounds alongside it.
The industry that built careers on measuring compute efficiency has spent two years measuring only the part that generates press releases.