Foundry capex cycle confirms itself with pulled-in EUV deliveries
Revenue per wafer has a structural slope that no single customer's budget cycle can flatten. The near-term watch items are construction schedules — cleanrooms running behind tool deliveries — because those convert committed orders into charged interest rather than early revenue.
The capex cycle is confirming across customer types at the same time. Logic's roadmap adds EUV layers per wafer with each node, and one leading foundry is re-timing deliveries rather than waiting; a second foundry restarted previously deferred orders. Memory's share of bookings nearly doubled year over year as HBM4 capacity gets built on EUV, and cleanroom construction lags tool delivery at the flagship fab. The common thread is intensity: every node and every memory generation needs more EUV per wafer than the last.
- TSMC plans roughly $48 billion of 2027 capex, up from $42 billion, aimed at A16 and High-NA readiness
- A portion of TSMC's 2027 EUV orders was pulled in by one quarter
- SK hynix ordered six add-on EUV systems for HBM4 and 1c DRAM capacity
- Memory reached 39 percent of system bookings, up from 26 percent a year earlier
- Samsung restarted deferred EUV orders for its 1.4-nanometer pilot line
All Signals (2)
Related Themes
Cluster Debate
The capex cycle is confirming across customer types: TSMC raising 2027 spending to roughly $48 billion with A16's higher EUV layer count, deliveries pulled in by a quarter, Samsung restarting deferred orders, and SK hynix adding six EUV systems for HBM4. EUV intensity per wafer rises with every node, which gives tool demand a structural slope.
- TSMC 2027 capex planned at roughly $48 billion, up from $42 billion, with a portion of EUV orders pulled in one quarter
- A16 requires more EUV layers per wafer than the current N3 family
- SK hynix ordered six add-on EUV systems for HBM4 and 1c DRAM capacity
- Memory reached 39 percent of system bookings, up from 26 percent a year earlier
Two customers made news this cycle, one pulled orders forward and one cancelled six units, and both stories were told as timing. The pattern is that capacity arrives when the fab shell does, and the shells are running late — the M15X cleanroom is behind its tool deliveries, and the cancelled order blamed a delayed shell.
- Cleanroom construction at the M15X fab runs roughly one quarter behind tool delivery
- The cancelled order was tied to a delayed fab shell, not a demand reassessment
- Slot reallocation pushes six Low-NA systems a year out
- Analysts cut 2026 EUV shipment forecasts by four to six units
- Leading-edge demand is broadening across logic and memory
- EUV layers per wafer increase with each node
- Construction and qualification schedules, not orders, gate near-term revenue
Timing changes around a rising multi-year trend are ordinary in capital equipment. The direction is set by technology transitions, and both A16 and HBM4 increase EUV intensity.
When one customer pulls in and another cancels in the same quarter, the book's aggregate hides the volatility. The revenue is booked; the quarter it lands in is now a construction conversation.
- TSMC has pulled in delivery dates for a portion of its 2027 EUV orders by one quarter
- a large memory customer cancelled six Low-NA EUV systems previously slotted for 2026 delivery
HBM4 and 1c DRAM capacity require EUV, and the memory upcycle has multi-year spending behind it. A 39 percent share reflects a technology transition, not a one-quarter spike.
Memory buyers are the least committed customers in this industry and the most exposed to a single product cycle. Six-system add-ons are what a customer looks like before it decides whether the HBM price holds.
- an add-on order for six EUV systems to support HBM4 and 1c DRAM capacity
- 'incremental EUV tool commitments' for its 1-gamma node without specifying quantities
The agents agree on the structural mechanics: more layers per node, two customer classes spending, and construction schedules as the binding constraint. They diverge on what the lumpiness signals. Pull-ins prove conviction, cancellations prove optionality, and both can be true because the customers differ. The next two quarters of shipment data resolve it: if pulled-in logic deliveries land on schedule while memory conversions hold, the cluster is a cycle; if the pulled-in units slide to match their cleanrooms, the analyst's slope is right but the timing is not.