AI infrastructure does not amortize like concrete

The case against an AI bubble often rests on an accounting assumption that the trillions committed to data centers represent a one-time buildout that will taper off once the concrete is poured In an essay arguing that the AI bubble does not exist, Matteo Raso calculates that paying off roughly $5 trillion in industry commitments over twenty years amounts to $250 billion annually, a figure he argues enterprise demand can eventually digest.

That math treats accelerated compute like municipal bridges rather than silicon. Frontier AI clusters do not run for twenty years, they run hot under continuous load and GPUs need to be replaced every 4 years (estimated average for an NVIDIA H200, many older and lower-tier GPUs need replacement in just 1-3 years with AI heavy workloads). When hardware depreciates on a four-year horizon, capital requirements do not plummet after an initial wave. Operators must fund the next silicon refresh before the debt on the previous generation matures, transforming what looked like a temporary construction phase into permanent capital expenditure.

Anthropic reporting $2.08 in revenue for every dollar spent on compute shows that operational unit economics can function when clusters are fully utilized. The hurdle is whether enterprise software budgets can expand fast enough to outpace recurring hardware replacement. In corporate IT, procurement schedules and legacy integration move at quarter-speed, regardless of how quickly the underlying hardware burns through its depreciable life.