AI data centers.

Models turn over in like 18 months.

GPUs depreciate over 3 years, but the hyperscalers are playing accounting games by stretching the depreciation out to something like five years.

The physical data center itself--the concrete shell that houses everything--depreciates over 15-30 years.

Three different clocks. Duration mismatch. Hard to reconcile all of it.

Consequences:

Valuation distortion: Cash flows are front-loaded (GPU spend, training revenue spikes), but depreciation is back-loaded.

Capital rigidity: You can’t repurpose a 2025 H100 cluster easily when 2027 NPUs need new network fabric and power envelopes.

Systemic reflexivity: As accounting smooths earnings, investors underprice the true volatility of the asset base, inviting over-investment: the essence of the bubble dynamic.