AI Chips vs Railroad Tracks: Depreciation Changes the Bubble Debate

The image of railway tracks surviving a financial bust makes a powerful argument: even if investors lose money, society may retain useful infrastructure. Applying that image to AI requires separating the building, power connection, network and computing equipment.
Do not treat a data center as one asset
Each component can have a different useful life, maintenance requirement and alternative use. The relevant question is not whether a facility physically exists in ten years. It is whether customers will pay enough for the services it can then provide.
Odlyzko's Railway Mania research examines historical accounting and investor understanding. The modern analytical lesson is to inspect asset economics rather than infer returns from the continued existence of infrastructure.
Depreciation is an estimate, not a demand guarantee
An accounting useful life spreads a cost across periods. Economic competitiveness depends on the work an asset can perform and the price of alternatives. Neither should be inferred solely from the other.
A hypothetical $120 million equipment purchase depreciated evenly over six years creates $20 million of annual depreciation, ignoring residual value. Over four years, that becomes $30 million. The cash purchase is unchanged; the reported annual expense differs. These are illustrative assumptions, not useful-life estimates for a specific chip.
Test replacement and resale assumptions
Ask what happens if a customer prefers newer equipment before the original asset is fully depreciated. Could the older equipment serve another workload? At what price, power cost and conversion expense?
Avoid assuming either that older equipment becomes worthless immediately or that it preserves its original economics indefinitely. Both are claims requiring evidence.
Who bears the risk?
The equipment buyer, lessor and chip supplier can have different exposure. A supplier's current sale does not create the same balance-sheet risk as a customer's financed equipment purchase.
The stock-analysis task is to connect replacement spending, lease obligations and residual values to the entity you actually own. That makes the railroad analogy more informative than a blanket prediction that all infrastructure survives or all chips become obsolete.
More in this analysis series
Start with AI stocks versus railroad mania. Continue with The Bull Case Against the AI Bubble Thesis: The Productivity J-Curve.
Analysis published October 1, 2026. Historical comparisons and hypothetical examples are not forecasts. Educational information, not investment advice.