The Semiconductor Capex Supercycle Has Only Just Begun

Every prior semiconductor cycle was demand-led and inventory-driven: customers over-ordered, lead times stretched, and then the whole complex round-tripped into a glut. This cycle is capacity-led, and that distinction changes almost everything about how it should be modeled and understood.

The consensus view still treats advanced-node capex as a coiled spring that will snap back toward historical intensity ratios once AI enthusiasm cools. We think that framing mistakes a structural shift in who is paying for capacity, and why, for a cyclical overshoot.

The buyer of leading-edge capacity has changed, and the market hasn't finished repricing what that means for the sellers.

Historically, foundry demand was aggregated across dozens of fabless customers with staggered product cycles, which naturally smoothed utilization. Today, a handful of hyperscale buyers are committing multi-year capacity in advance, often with take-or-pay terms that shift inventory risk away from the foundry.

That shift matters because it changes the volatility profile of foundry revenue without changing how most models treat it. Utilization can stay structurally higher through a downturn than prior cycles suggest, because the marginal customer isn't reacting to a sell-through signal — they're building against a training-compute roadmap set 18 months earlier.

Public capex guidance understates the picture because it excludes the packaging and interconnect build-out that increasingly gates how much compute can actually ship. Advanced packaging capacity, not wafer starts, is the binding constraint in several product lines we track, and it is far less visible in headline capital intensity figures.

We think this is the more durable structural edge for incumbents with packaging scale: it is harder to replicate quickly, less commoditized than leading-edge lithography access, and increasingly where design wins are actually decided.

The clearest risk is a sharp deceleration in frontier model training runs, which would remove the marginal buyer setting current capacity commitments. Contracted backlog cancellation clauses are worth watching closely, since these determine how much downside protection foundries actually have if hyperscaler capex plans are revised.

A second risk is geographic: incentive-driven capacity additions outside existing hubs could arrive faster than expected, adding supply into a market that is currently capacity-constrained rather than demand-constrained.

Equipment and packaging suppliers whose revenue scales directly with capacity additions face a structurally different exposure profile than foundries, whose valuations may already reflect much of this capex thesis. Distinguishing between these parts of the value chain — rather than treating "semiconductor exposure" as a single category — is central to analyzing how this cycle could play out differently across it.

The read-through to test-and-assembly names is a useful illustration: they sit downstream of packaging capacity decisions, which means analyzing their exposure requires a different set of demand drivers than those governing front-end equipment makers.

This section is offered as a framework for understanding differentiated exposure across a value chain, not as a recommendation to buy, sell, or hold any security. Any conclusions readers draw from it should reflect their own research, judgment, and consultation with a licensed financial professional.