The Hidden Cost of AI Infrastructure: Hardware Depreciation Lags Actual Obsolescence

ai · x · 2026-08-01

As hyperscalers enter earnings season, the true cost of the AI infrastructure business is coming to light. On paper, the cost of purchasing GPUs is spread out over 5-7 years for accounting purposes, meaning a 5-year-old chip remains a productive asset on the balance sheet.

In reality, however, a new generation of AI chips drops roughly every 18 months, offering dramatic improvements in speed and energy efficiency. Running frontier AI workloads on older GPUs costs significantly more power per unit of output, and customers actively demand inference on the latest hardware. This puts hyperscalers on a never-ending treadmill where the actual replacement cycle is much shorter than financial depreciation suggests, meaning the real cost of AI infrastructure is far higher than reported.

Original post →

More from Venture

Venture channel →