The dominant question in technology investing through this cycle is not whether AI demand is real. It is what kind of cycle AI capex represents, and whether the rhythm of that cycle resembles the software cycles that shaped market intuition over the last fifteen years.
It does not. AI capex behaves much more like an industrial cycle than a software one. That has consequences for forecasting, valuation, and risk.
Software cycle vs. industrial cycle
A software cycle is characterised by negative working capital, low incremental cost of serving the next customer, and operating leverage that compounds quickly once distribution is established. Drawdowns are usually demand-driven and short. Capex is small relative to free cash flow.
An industrial cycle is characterised by long-dated, lumpy capacity decisions made under uncertainty, with depreciation that runs on a schedule the business cannot pause. Supply takes years to come online, but once it is online, it has to be filled at falling marginal cost to clear. Cycle troughs are deep because the fixed-cost base cannot be retracted.
AI infrastructure — chips, accelerators, power, cooling, hyperscale data centers, optical and networking — is in the second category. The unit economics resemble semiconductor capacity and utility expansion more than SaaS subscription growth.
Where the binding constraints sit
The bottleneck is migratory and worth tracking, not just at the chip layer:
- Advanced packaging (CoWoS-class capacity) often constrains throughput before wafer supply does.
- HBM memory has its own cycle distinct from logic, with a different vendor structure.
- Grid interconnects and substation capacity now drive data-center site selection more than land or fiber.
- Power purchase agreements are increasingly the long-pole item for new hyperscale builds, and they constrain how quickly capacity can be brought online even when capital is willing.
Whichever constraint binds determines who captures economic rent at the margin and where pricing power lives in the current vintage of demand.
Depreciation is the hidden cycle
The single most under-discussed feature of this build-out is the depreciation schedule applied to short-lived AI accelerators. A four-to-six year useful life — generous for current-generation training silicon — means that hyperscaler operating expense will climb on a known cadence regardless of revenue realisation. If demand fails to grow into the depreciation step-ups, operating leverage works in reverse, and the cycle troughs the way an industrial cycle troughs: through margin compression, not through demand collapse alone.
Implications for research framing
For investment work, three implications follow:
- Forecasting horizon. Quarterly framing under-models how long it takes for supply to respond. Multi-year capacity ramps require multi-year demand triangulation, not just bookings momentum.
- Cross-asset signals. Industrial cycles propagate through commodities, power markets, real assets, and credit, not just equity multiples. Watching only the equity tape misses early warning.
- Reflexivity risk. Valuation depends on capex assumptions, capex depends on cost of capital, cost of capital depends on confidence in valuation. This loop is durable on the way up and on the way down.
The signal worth tracking is not whether AI is “real.” It is whether the ratio of incremental useful compute to incremental dollar spent is improving, flat, or deteriorating. That is the variable an industrial cycle turns on.